id string | text string |
|---|---|
1501.04049 | Title: The Geometry and Moduli of K3 Surfaces
Abstract: These notes will give an introduction to the theory of K3 surfaces. We begin
with some general results on K3 surfaces, including the construction of their
moduli space and some of its properties. We then move on to focus on the theory
of polarized K3 surfaces, studying their moduli, degenerations and the
compactification problem. This theory is then further enhanced to a discussion
of lattice polarized K3 surfaces, which provide a rich source of explicit
examples, including a large class of lattice polarizations coming from elliptic
fibrations. Finally, we conclude by discussing the ample and Kahler cones of K3
surfaces, and give some of their applications.
Body: \title{The Geometry and Moduli of {K3} Surfaces} \author{Andrew Harder} \address{Department of Mathematical and Statistical Sciences, 632 CAB, University of Alberta, Edmonton, Alberta, T6G 2G1, Canada} \email{aharder@ualberta.ca} \thanks{A. Harder was supported by an NSERC PGS D scholarship and a University of Alberta Doctoral Recruitment Scholarship.} \author{Alan Thompson} \address{Department of Mathematical and Statistical Sciences, 632 CAB, University of Alberta, Edmonton, Alberta, T6G 2G1, Canada} \email{amthomps@ualberta.ca} \thanks{A. Thompson was supported by a Fields-Ontario-PIMS postdoctoral fellowship with funding provided by NSERC, the Ontario Ministry of Training, Colleges and Universities, and an Alberta Advanced Education and Technology Grant.} \begin{abstract} These notes will give an introduction to the theory of K3 surfaces. We begin with some general results on K3 surfaces, including the construction of their moduli space and some of its properties. We then move on to focus on the theory of polarized K3 surfaces, studying their moduli, degenerations and the compactification problem. This theory is then further enhanced to a discussion of lattice polarized K3 surfaces, which provide a rich source of explicit examples, including a large class of lattice polarizations coming from elliptic fibrations. Finally, we conclude by discussing the ample and K\"ahler cones of K3 surfaces, and give some of their applications.\end{abstract} \maketitle \section{General Results on K3 Surfaces} We begin by recalling the definition of a K3 surface. \begin{definition} A \emph{K3 surface} $S$ is a smooth compact complex surface with trivial canonical bundle $\omega_S \cong \calO_S$ and $h^1(S,\calO_S) = 0$. \end{definition} \begin{remark} Note that an arbitrary K3 surface $S$ is not necessarily projective, but every K3 surface is K\"{a}hler. This was first proved by Siu who, by treating the K3 case, completed the proof of a conjecture of Kodaira \cite[Sect. XII.1]{cm} stating that every smooth compact complex surface with even first Betti number is K\"{a}hler. A direct proof of this conjecture may be found in \cite[Thm. IV.3.1]{bpv}. \end{remark} Unless otherwise stated, throughout these notes $S$ will denote an arbitrary K3 surface. In the remainder of this section we will study the geometry of $S$, then use this to initiate our study of the moduli space of K3 surfaces. Our main reference for this section will be \cite[Chap. VIII]{bpv}. \subsection{Hodge Theory} We begin by studying the Hodge theory of a K3 surface $S$. The Hodge diamond of $S$ has the form \[\begin{array}{ccccc} && h^{0,0} &&\\ &h^{1,0} & & h^{0,1} & \\ h^{2,0} & & h^{1,1} & & h^{0,2} \\ & h^{2,1} & & h^{1,2} & \\ && h^{2,2} && \end{array}\quad =\quad \begin{array}{ccccc} && 1 &&\\ &0 & & 0 & \\ 1 & & 20 & & 1 \\ & 0 & & 0 & \\ && 1 && \end{array}.\] We note that this is largely trivial: the only interesting behaviour happens in the second cohomology group. As we shall see, the structure of this cohomology group determines the isomorphism class of a K3 surface, so can be used to construct a moduli space for K3 surfaces. The second cohomology group $H^2(S,\Z)$ with the cup-product pairing $\langle \cdot , \cdot\rangle$ forms a lattice isometric to the \emph{K3 lattice} \[ \Lambda_{\mathrm{K}3} := H \oplus H \oplus H \oplus (-E_8) \oplus (-E_8),\] where $H$ is the hyperbolic plane (an even, unimodular, indefinite lattice of rank $2$) and $E_8$ is the even, unimodular, positive definite lattice of rank $8$ corresponding to the Dynkin diagram $E_8$. The lattice $\Lambda_{\mathrm{K}3}$ is a non-degenerate even lattice of rank $22$ and signature $(3,19)$ (for the reader unfamiliar with lattice theory, we have included a short appendix containing results and definitions relevant to these notes). There are two important sublattices of $H^2(S,\Z)$ that appear frequently in the study of K3 surfaces. The first is the \emph{N\'{e}ron-Severi lattice} $\NS(S)$, given by \[\NS(S) := H^{1,1}(S) \cap H^2(S,\Z) \] (here we identify $H^2(S,\Z)$ with its image under the natural embedding $H^2(S,\Z) \hookrightarrow H^2(S,\C)$). By the Lefschetz theorem on $(1,1)$-classes \cite[Thm. IV.2.13]{bpv}, $\NS(S)$ is isomorphic to the Picard lattice $\mathrm{Pic}(S)$, with isomorphism induced by the first Chern class map. The second important sublattice of $H^2(S,\Z)$ is the \emph{transcendental lattice} $\T(S)$. It is defined to be the smallest sublattice of $H^2(S,\Z)$ whose complexification contains a generator $\sigma$ of $H^{2,0}(S)$. In the case where $\NS(S)$ is nondegenerate (which happens, for instance, when $S$ is projective), then the transcendental lattice is equal to the orthogonal complement of $\NS(S)$ in $H^2(S,\Z)$. The structure of the second cohomology of $S$ is an important object to study, as it determines the isomorphism class of $S$. \begin{theorem}[Weak Torelli] \textup{\cite[Cor. VIII.11.2]{bpv}} Two K3 surfaces $S$ and $S'$ are isomorphic if and only if there is a lattice isometry $H^2(S,\Z) \to H^2(S',\Z)$, whose $\C$-linear extension $H^2(S,\C) \to H^2(S',\C)$ preserves the Hodge decomposition \textup{(}such an isometry is called a \emph{Hodge isometry}\textup{)}. \end{theorem} \subsection{The Period Mapping} We can use the weak Torelli theorem to begin constructing a moduli space for K3 surfaces. We start by defining a \emph{marking} on the K3 surface $S$. \begin{definition} A \emph{marking} on $S$ is a choice of isometry $\phi\colon H^2(S,\Z) \to \Lambda_{\mathrm{K}3}$. We say that $(S,\phi)$ is a \emph{marked K3 surface}. \end{definition} Since the canonical bundle of $S$ is trivial, we have $H^{2,0}(S) := H^0(S,\Omega^2_S) = H^0(S,\calO_S)$. Let $\sigma \in H^{2,0}(S)$ be any nonzero element. Then $\sigma$ is a nowhere vanishing $2$-form on $S$. Using the Hodge decomposition, we may treat $\sigma$ as an element of $H^2(S,\C)$. This cohomology group carries a bilinear form $\langle \cdot,\cdot\rangle$, given by the $\C$-linear extension of the cup-product pairing, with respect to which we have $\langle\sigma,\sigma\rangle = 0$ and $\langle\sigma,\overline{\sigma}\rangle >0$. If $\phi$ is a marking for $S$ and $\phi_{\C}\colon H^2(S,\C) \to \Lambda_{\mathrm{K}3} \otimes \C$ is its $\C$-linear extension, then $\phi_{\C}(H^{2,0}(S))$ is a line through the origin in $\Lambda_{\mathrm{K}3} \otimes \C$ spanned by $\phi_{\C}(\sigma)$. Projectivising, we see that $\phi_{\C}(H^{2,0}(S))$ defines a point in \[\Omega_{\mathrm{K}3} := \{[\sigma]\in \Proj(\Lambda_{\mathrm{K}3} \otimes \C) \mid \langle\sigma,\sigma\rangle = 0, \langle\sigma,\ \overline{\sigma}\rangle > 0\}.\] $\Omega_{\mathrm{K}3}$ is a $20$-dimensional complex manifold called the \emph{period space of K3 surfaces}. The point defined by $\phi_{\C}(H^{2,0}(S))$ is the \emph{period point of the marked K3 surface $(S,\phi)$}. The Weak Torelli theorem (Thm. ) gives that two K3 surfaces are isomorphic if and only if there are markings for them such that the corresponding period points are the same. Now we extend this idea to families. Let $\pi \colon \calS \to U$ be a flat family of K3 surfaces over a small contractible open set $U$ and let $S$ be a fibre of $\pi$. A choice of marking $\phi\colon H^2(S,\Z) \to \Lambda_{\mathrm{K3}}$ for $S$ can be extended uniquely to a marking $\phi_U\colon R^2\pi_*\Z \to (\Lambda_{\mathrm{K3}})_U$ for the family $\calS$, where $(\Lambda_{\mathrm{K3}})_U$ denotes the constant sheaf with fibre $\Lambda_{\mathrm{K3}}$ on $U$. Applying the above construction to the marked K3 surfaces in the family $\calS$, we obtain a holomorphic map $U \to \Omega_{\mathrm{K3}}$, called the \emph{period mapping} associated to the family $\pi\colon \calS \to U$. Applying this to the case where $\pi\colon \calS \to U$ is a representative of the versal deformation of $S$, one finds: \begin{theorem}[Local Torelli] \textup{\cite[Thm. VIII.7.3]{bpv}} For any marked K3 surface $S$, the period mapping from the versal deformation space of $S$ to $\Omega_{\mathrm{K}3}$ is a local isomorphism.\end{theorem} This shows that the period mapping is well-behaved under small deformations of a marked K3 surface. Moreover, we have: \begin{theorem}[Surjectivity of the Period Map] \textup{\cite[Cor. VIII.14.2]{bpv}} Every point of $\Omega_{\mathrm{K}3}$ occurs as the period point of some marked K3 surface.\end{theorem} Putting these elements together, we seem to be close to constructing a (coarse) moduli space for K3 surfaces: we have a space $\Omega_{\mathrm{K3}}$ whose points correspond to marked K3 surfaces, and any family of marked K3 surfaces $\pi\colon \calS \to U$ gives rise to a map $U \to \Omega_{\mathrm{K3}}$. All that remains is to quotient $\Omega_{\mathrm{K}3}$ by the action of the group $\Gamma$ of isometries of $\Lambda_{\mathrm{K}3}$ to identify period points corresponding to different markings on the same K3 surface. However, on closer inspection one finds that this group action is not properly discontinuous, so the quotient will have undesirable properties: in particular, it won't be Hausdorff. More details may be found in \cite[Sect. VIII.12]{bpv}. \section{Polarized K3 Surfaces} One way to solve this problem is to restrict our attention to a subclass of K3 surfaces that have better properties: the \emph{pseudo-polarized} K3 surfaces. \begin{definition} A (\emph{pseudo}-)\emph{polarized K3 surface of degree $2k$} (for $k>0$) is a pair $(S,h)$ consisting of a K3 surface $S$ and a primitive (pseudo-)ample class $h \in \NS(S)$ with $\langle h,h\rangle = 2k$. Two (pseudo-)polarized K3 surfaces $(S,h)$ and $(S',h')$ of degree $2k$ are \emph{equivalent} if there exists an isomorphism $f\colon S \to S'$ of K3 surfaces such that $f^*(h') = h$. \end{definition} \begin{remark} If $\NS(S)$ contains a pseudo-ample class, then $S$ is a Moishezon manifold by \cite[Thm. 2.2.15]{hmibk}. As $S$ is also K\"{a}hler, \cite[Thm. 2.2.26]{hmibk} implies that $S$ is projective. Thus every pseudo-polarized K3 surface is projective. \end{remark} The geometry of pseudo-polarized K3 surfaces was studied by Mayer . The following easy consequence of Props. 1 and 2 from his paper is particularly useful for studying them explicitly. \begin{proposition} Let $(S,h)$ be a pseudo-polarized K3 surface of degree $2k$ and let $D$ be an effective divisor on $S$ with $[D] = h$ in $\NS(S)$. Then the map $f\colon S \to \Proj(H^0(S,\calO_S(D)))$ defined by the linear system $|D|$ is \begin{itemize} \item \textup{(}generic case\textup{)} a birational morphism onto a normal surface of degree $2k$ in $\Proj^{k+1}$ if the general member of $|D|$ is a smooth non-hyperelliptic curve; or \item \textup{(}hyperelliptic case\textup{)} a morphism of degree $2$ onto a normal surface of degree $k$ in $\Proj^{k+1}$ if the general member of $|D|$ is a smooth hyperelliptic curve; or \item \textup{(}unigonal case\textup{)} a regular map $S \to \Proj^{k+1}$ whose image is a rational curve of degree $k+1$ if the general member of $|D|$ is reducible. \end{itemize} \end{proposition} Using this, we can introduce two of the most widely studied classes of K3 surfaces. \begin{example}[Sextic double planes] Suppose first that $(S,h)$ is a pseudo-polarized K3 surface of degree $2$ and let $D$ be an effective divisor on $S$ with $[D] = h$. Then the general member of $|D|$ is either a smooth hyperelliptic curve or is reducible. In the hyperelliptic case, which for degree $2$ is generic, the linear system $|D|$ defines a generically $2:1$ map from $S$ onto $\Proj^2$. We thus see that $S$ is birational to a double cover of $\Proj^2$ ramified over a sextic curve. Such surfaces may be realized as sextic hypersurfaces in the weighted projective space $\mathbb{W}\Proj(1,1,1,3)$. In the unigonal case, which for degree $2$ can only occur when $D$ is pseudo-ample but not ample, the linear system $|D|$ defines a regular map from $S$ onto a smooth conic in $\Proj^2$. The general fibre of this map is a smooth elliptic curve. \end{example} \begin{example}[Quartic hypersurfaces] For our second example, suppose that $(S,h)$ is a pseudo-polarized K3 surface of degree $4$ and let $D$ be an effective divisor on $S$ with $[D] = h$. Then all three cases from Prop. can occur. In the generic case the linear system $|D|$ defines a birational morphism onto a quartic hypersurface in $\Proj^3$. In the hyperelliptic case, the linear system $|D|$ defines a generically $2:1$ map from $S$ onto a quadric hypersurface in $\Proj^3$. This hypersurface is isomorphic to $\Proj^1 \times \Proj^1$, so $S$ is birational to a double cover of $\Proj^1 \times \Proj^1$ ramified over a curve of bidegree $(4,4)$. Such surfaces may be realized as complete intersections of degree $(2,4)$ in the weighted projective space $\mathbb{W}\mathbb{P}(1,1,1,1,2)$, where the degree two relation does not involve the degree two variable (since if it did then we could eliminate it, putting us back in the generic case of a quartic hypersurface in $\Proj^3$). Finally, in the unigonal case the linear system $|D|$ defines a regular map from $S$ onto a twisted cubic in $\Proj^3$. The general fibre of this map is again a smooth elliptic curve. \end{example} \subsection{Moduli of Polarized K3 Surfaces} For polarized K3 surfaces we have an upgraded version of the Weak Torelli Theorem (Thm. ), which will enable us to build a moduli space for them. \begin{theorem}[Strong Torelli] \textup{\cite[Cor. VIII.3.12 and Thm. VIII.11.1]{bpv}} Let $(S,h)$ and $(S',h')$ be polarized K3 surfaces of the same degree $2k$. Assume that there is a Hodge isometry $\varphi\colon H^2(S',\Z) \to H^2(S,\Z)$ with $\varphi(h')=h$. Then there is a \emph{unique} isomorphism $f\colon S \to S'$ with $\varphi = f^*$ \textup{(}i.e. $S$ and $S'$ are equivalent\textup{)}. \end{theorem} Following our previous discussion, we next construct a period space for pseudo-polarized K3 surfaces. Fix once and for all a primitive class $h \in \Lambda_{\mathrm{K}3}$ with $\langle h,h\rangle = 2k >0$. Then a \emph{marked} \mbox{(\emph{pseudo}-)}\emph{polarized K3 surface of degree $2k$} is a marked K3 surface $(S,\phi)$ such that $\phi^{-1}(h)$ is a (pseudo-)ample class in $\NS(S)$. If $(S,\phi)$ is a marked pseudo-polarized K3 surface of degree $2k$ and if $\sigma \in H^{2,0}(S)$ is any nonzero element, then we have $\langle\sigma,\sigma\rangle = 0$, $\langle\sigma,\overline{\sigma}\rangle > 0$ and $\langle\sigma,\phi^{-1}(h)\rangle = 0$. So the period point of $(S,\phi)$ lies in \[\Omega_{2k} := \{[\sigma]\in \Proj(\Lambda_{\mathrm{K}3} \otimes \C) \mid \langle\sigma,\sigma\rangle = 0,\ \langle\sigma,\overline{\sigma}\rangle > 0,\ \langle\sigma,h\rangle = 0\} \subset \Omega_{\mathrm{K}3}.\] $\Omega_{2k}$ is called the \emph{period space of pseudo-polarized K3 surfaces of degree $2k$}. It is a $19$-dimensional complex manifold with two connected components, each of which is a bounded symmetric domain of type IV \cite[Rmk. VIII.22.2]{bpv}. By the Surjectivity of the Period Map (Thm. ), every point of $\Omega_{2k}$ corresponds to a marked K3 surface $(S,\phi)$. Furthermore, for any generator $\sigma \in H^{2,0}(S)$ we have $\langle \sigma, \phi^{-1}(h) \rangle = 0$, so $\phi^{-1}(h) \in \NS(S)$. Thus if $\phi^{-1}(h)$ is an ample class, then $(S,\phi)$ will be a marked polarized K3 surface of degree $2k$. Using this and the Torelli Theorems (Thms. and ) we can construct a coarse moduli space for polarized K3 surfaces of degree $2k$. First, however, we have to remove the points in $\Omega_{2k}$ corresponding to the K3 surfaces that are pseudo-polarized but not polarized. If a marked pseudo-polarized K3 surface $(S,\phi)$ is not polarized, then the Hodge Index Theorem \cite[Cor. IV.2.16]{bpv} and the genus formula imply that it must contain a rational curve $C$, such that the class $\delta$ of $C$ in $\NS(S)$ satisfies $\langle \delta, \delta \rangle = -2$ and $\langle \delta, \phi^{-1}(h) \rangle = 0$. The converse is also true: if $\NS(S)$ contains such a $\delta$, then by \cite[Prop. VIII.3.7]{bpv} there exists a rational curve $C$ on $S$ with $\langle[C],\phi^{-1}(h)\rangle = 0$, so $\phi^{-1}(h)$ is not ample and $(S,\phi)$ is not polarized. Using this, we see that a marked pseudo-polarized K3 surface $(S,\phi)$ is not polarized if and only if its period point $[\sigma]$ is orthogonal to a point in the set \[ \Delta_{2k} := \{\delta \in \Lambda_{\mathrm{K}3} \mid \langle \delta, \delta \rangle = -2, \ \langle \delta, h \rangle = 0 \}.\] For each $\delta \in \Delta_{2k}$, define a hyperplane \begin{equation} H_{\delta} := \{[\sigma]\in \Proj(\Lambda_{\mathrm{K}3} \otimes \C) \mid \langle \sigma, \delta \rangle = 0\}.\end{equation} Then define \[ \Omega_{2k}^0 := \Omega_{2k} - \bigcup_{\delta \in \Delta_{2k}} (H_{\delta} \cap \Omega_{2k}).\] $ \Omega_{2k}^0$ is called the \emph{period space of polarized K3 surfaces}. The Torelli Theorems (Thms. and ) and the Surjectivity of the Period Map (Thm. ) show that its points are in bijection with marked polarized K3 surfaces of degree $2k$. It just remains to quotient by an appropriate group to identify period points corresponding to different markings on the same K3 surface. Let $\Gamma(h)$ denote the group of isometries of $\Lambda_{\mathrm{K}3}$ that fix the class $h$. Then $\Gamma(h)$ acts properly discontinuously on $\Omega_{2k}^0$, so the quotient $\Gamma(h) \setminus \Omega_{2k}^0$ will not have the same problems that we experienced before. Thus we have: \begin{theorem} \textup{\cite[Thm. VIII.22.4]{bpv}} The quotient \[\calF_{2k}^0 := \Gamma(h) \setminus \Omega_{2k}^0\] is a coarse moduli space for polarized K3 surfaces of degree $2k$. \end{theorem} \begin{remark} At first glance this definition appears to depend upon the choice of the class $h \in \Lambda_{\mathrm{K}3}$. However, it can be shown that all choices of $h$ yield the same moduli space and that, in fact, it is possible to construct $\calF_{2k}$ without making reference to a specific choice of $h$. However, this construction requires somewhat more theoretical background than the one presented above, so we will not detail it here: the interested reader may refer to \cite[Sect. 1.1]{cmsak3s} for details. \end{remark} $\calF_{2k}^0$ is a $19$-dimensional complex space with only finite quotient singularities. By exhibiting a projective compactification, Baily and Borel showed that it is even quasi-projective (their compactification will be discussed further in Sect. ). It may be obtained from $\calF_{2k} := \Gamma(h) \setminus \Omega_{2k}$ by removing finitely many divisors. $\calF_{2k}$ is also a $19$-dimensional quasi-projective variety with only finite quotient singularities and, as $\Gamma(h)$ contains an isometry that interchanges the two connected components of $\Omega_{2k}$, it is even connected \cite[Thm. VIII.22.3]{bpv}. One can think of the points in the complement $\calF_{2k} - \calF_{2k}^0$ as corresponding to K3 surfaces that are pseudo-polarized but not polarized, but the definitions required to make this rigourous are somewhat subtle; we refer the interested reader to . $\calF_{2k}$ may therefore be thought of as a coarse moduli space for pseudo-polarized K3 surfaces of degree $2k$. It can be realized as a quotient of a bounded symmetric domain of type IV (given by one of the connected components of $\Omega_{2k}$), by an arithmetically defined discrete group of automorphisms, a fact that makes it very amenable to explicit study: details may be found in \cite[Sect. 1]{cmsak3s}. \subsection{Degenerations} In the remainder of this section, we will discuss what happens when we proceed to the boundary of this moduli space. In order to do this we study \emph{degenerations}. Our main reference for this study will be . \begin{definition} A \emph{degeneration of K3 surfaces} is a proper, flat, surjective morphism $\pi\colon \calS \to \Delta$ from a smooth threefold $\calS$ to the unit disc $\Delta \subset \C$, whose general fibre $S_t = \pi^{-1}(t)$ for $t\neq 0$ is a smooth K3 surface. Note that we do not assume that $\calS$ is algebraic, but we will make the assumption that the components of the central fibre $S_0 = \pi^{-1}(0)$ are K\"{a}hler.\end{definition} Let $\pi\colon \calS \to \Delta$ be any degeneration of K3 surfaces. We begin our analysis by converting this degeneration into a form that has certain desirable properties. The first step is to arrange for \emph{semistability}, using the Semistable Reduction Theorem of Knudsen, Mumford and Waterman: \begin{theorem}[Semistable Reduction] \textup{} Let $\pi \colon \calS \to \Delta$ be a degeneration of surfaces. Then there exists an $m$ such that, if $\pi'\colon \calS' \to \Delta$ is the base change by the map $\varrho\colon \Delta \to \Delta$ given by $\varrho(t) = t^m$, there is a birational morphism $\hat{\calS} \to \calS'$ so that $\psi\colon \hat{\calS} \to \Delta$ is semistable, i.e. $\hat{\calS}$ is nonsingular and $\hat{S}_0 :=\psi^{-1}(0)$ is a reduced divisor with normal crossings. \[\xymatrix{ \hat{\calS} \ar[r] \ar[d]_{\psi}& \calS' \ar[r] \ar[d]_{\pi'} & \calS \ar[d]_{\pi} \\ \Delta \ar@{=}[r] & \Delta \ar[r]^{\varrho} & \Delta \\ }\] \end{theorem} \begin{remark} Note that this theorem holds for degenerations of surfaces in general, not just for degenerations of K3 surfaces, although we will only use the K3 version here.\end{remark} To illustrate the computation of semistable reduction in an example, we will simplify matters by considering a degeneration of elliptic curves. The basic theory is largely unchanged from the K3 surface case, but the equations are substantially simpler. \begin{example}[The cuspidal elliptic curve] Consider the family $\pi\colon \calE \to \Delta := \{t \in \C \mid |t|<\varepsilon\}$ of elliptic curves in $\mathbb{A}^2 \times \Delta$ given by the equation \[y^2 = x^3 + tf_3(x),\] where $(x,y)$ are coordinates on $\mathbb{A}^2$ and $f_3(x)$ is a smooth cubic polynomial in $x$ with $f_3(0) \neq 0$. $\calE$ is smooth, but the central fibre $E_0 = \pi^{-1}(0)$ of $\calE$ is a cuspidal elliptic curve, which does not have normal crossings. $\calE \to \Delta$ is thus not a semistable degeneration of elliptic curves. To make it semistable, we use Thm. . However, first we need to determine the order $m$ of the cover $\varrho$ that we need to take. To do this, first blow up $\calE$ until the central fibre has only normal crossings. Let $m_1,\ldots,m_n$ denote the multiplicities of the irreducible components of the new central fibre. Then $m = \mathrm{lcm}(m_1,\ldots,m_n)$. In our case, to obtain a fibre with normal crossings we need to blow up the point $(x,y;t) = (0,0;0)$ three times. The strict transform of $E_0$ under this blow up has multiplicity $1$, and the three exceptional curves have multiplicities $2$, $3$ and $6$. We thus have $m = 6$. Let $\calE'$ denote the pull-back of $\calE$ by the map $\varrho\colon \Delta \to \Delta$ given by $\varrho(t) = t^6$. Then $\calE'$ is given in $\mathbb{A}^2 \times \Delta$ by \[y^2 = x^3 + t^6f_3(x).\] $\calE'$ is singular at $(x,y;t) = (0,0;0)$. The singularity is locally analytically isomorphic to $\{y^2 = x^3 + t^6\} \subset \C^3$. This is an example of a \emph{minimally elliptic singularity}. Such singularities have been studied by Laufer . The resolution of this singularity is given in \cite[Table 5.1]{omes}. To resolve it, we blow up the point $(0,0;0) \in \calE'$ once. The resulting exceptional curve is an elliptic curve with self-intersection $(-1)$. The resolved family $\hat{\calE} \to \Delta$ is semistable, with central fibre consisting of a rational $(-1)$-curve meeting an elliptic $(-1)$-curve at a single node, both with multiplicity $1$. In fact, in this case we can go one step further, by contracting the rational $(-1)$-curve in the central fibre of $\hat{\calE}$. This does not introduce any new singularities into $\hat{\calE}$, so the resulting family is semistable and all of its fibres are smooth elliptic curves. \end{example} Once our denegeration of K3 surfaces is semistable, we may additionally arrange for the canonical bundle of the total space to be trivial, using the following theorem of Kulikov, Persson and Pinkham: \begin{theorem} \textup{} If $\psi\colon \hat{\calS} \to \Delta$ is a semistable degeneration of K3 surfaces, and if all components of $\hat{S}_0 = \psi^{-1}(0)$ are K\"{a}hler, then there exists a birational modification $\hat{\calS}'$ of $\hat{\calS}$ such that $\psi'\colon \hat{\calS}' \to \Delta$ is semistable, isomorphic to $\hat{\calS}$ over $\Delta - \{0\}$, and has $\omega_{\hat{\calS}'} \cong \calO_{\hat{\calS}'}$. \end{theorem} Motivated by this theorem, a \emph{Kulikov model} is defined to be a semistable degeneration of K3 surfaces $\pi\colon \calS \to \Delta$ with $\omega_{\calS} \cong \calO_{\calS}$; the discussion above shows that any degeneration of K3 surfaces may be converted into a Kulikov model by a base change and a birational modification. \begin{remark} It is important to note that the construction of the Kulikov model is very non-algebraic in nature, so even if $\hat{\calS}$ is algebraic, its Kulikov model $\hat{\calS}'$ may not be. We do, however, know that the Kulikov model $\hat{\calS}'$ is complex analytic and that all components of its central fibre are K\"{a}hler. \end{remark} Kulikov models are useful because there exists a rough classification of their central fibres, first proven by Kulikov, Persson, Friedman and Morrison. However, in order to state it we first need to introduce the \emph{dual graph} of the central fibre of a degeneration. \begin{definition} Let $S_0 = \bigcup V_i$ be the central fibre in a semistable degeneration. Define the dual graph $\Gamma$ of $S_0$ as follows: $\Gamma$ is a simplicial complex whose vertices $P_1, \ldots , P_r$ correspond to the components $V_1,\ldots,V_r$ of $S_0$; the $k$-simplex $\langle P_{i_0},\ldots,P_{i_k} \rangle$ belongs to $\Gamma$ if and only if $V_{i_0} \cap \cdots \cap V_{i_k} \neq \emptyset$. \end{definition} This enables us to state: \begin{theorem}[Classification of Kulikov Models] \textup{} Let $\pi\colon \calS \to \Delta$ be a semistable degeneration of K3 surfaces with $\omega_{\calS} \cong \calO_{\calS}$, such that all components of $S_0 = \pi^{-1}(0)$ are K\"{a}hler. Then either \begin{enumerate}[\textup{(}Type I\textup{)}] \item $S_0$ is a smooth K3 surface; \item $S_0$ is a chain of elliptic ruled components with rational surfaces at each end, and all double curves are smooth elliptic curves; \item $S_0$ consists of rational surfaces meeting along rational curves which form cycles in each component. If $\Gamma$ is the dual graph of $S_0$, then $\Gamma$ is a triangulation of the $2$-sphere. \end{enumerate} \end{theorem} These cases can also be distinguished by the action of monodromy on the second cohomology $H^2(S_t,\Z)$ of a general fibre. Let $T$ denote the Picard-Lefschetz transformation on $H^2(S_t,\Z)$ obtained by the action of monodromy around $0$ and let $N = \log T$. Then $N$ is nilpotent and has $N = 0$ if $S_0$ is Type I, $N^2 = 0$ and $N\neq 0$ if $S_0$ is Type II, and $N^3 = 0$ and $N^2 \neq 0$ if $S_0$ is Type III. We conclude this section by giving two examples of degenerations of K3 surfaces, one of Type II and one of Type III. \begin{example}[Type II degeneration] We begin with the Type II example. Consider the family $\pi\colon \calS \to \Delta := \{t \in \C \mid |t| < \varepsilon\}$ given in $\mathbb{W}\Proj(1,1,1,3) \times \Delta$ by the formula \[y^2 = (f_3(x_1,x_2,x_3))^2 + tg_6(x_1,x_2,x_3),\] where $(x_1,x_2,x_3,y)$ are coordinates on $\mathbb{W}\Proj(1,1,1,3)$ of weights $(1,1,1,3)$ respectively and $f_3$, $g_6$ are generic homogeneous polynomials in the $x_i$ of degrees $3$ and $6$ respectively. The general fibre of $\calS$ is a sextic hypersurface in $\mathbb{W}\Proj(1,1,1,3)$ which, by Example , is a generic polarized K3 surface of degree two. The central fibre $S_0 = \pi^{-1}(0)$ is isomorphic to two copies of $\Proj^2$ glued along an elliptic curve $\{f_3(x_1,x_2,x_3) = 0\}\subset \Proj^2$. Note that $\omega_{\calS} \cong \calO_{\calS}$, but $\calS$ is \emph{not} a Kulikov model as $\calS$ is not smooth, so the family $\calS \to \Delta$ is not semistable. In fact, $\calS$ has eighteen singularities at the points $\{f_3(x_1,x_2,x_3) = g_6(x_1,x_2,x_3) = y = t = 0\}$. Each of these singularities is locally analytically isomorphic to $\{y^2 = x^2 + tz\} \subset \C^4$, which is a threefold node. To solve this, one's first instinct would be to blow up each of the eighteen singularities individually. This introduces eighteen exceptional divisors $E_1,\ldots,E_{18}$, each of which is isomorphic to $\Proj^1 \times \Proj^1$. Indeed, the resulting family is semistable, but it is still not a Kulikov model: in this case, the canonical bundle is isomorphic to $\calO(E_1 + \cdots + E_{18})$, which is non-trivial. Instead, we notice that $\calS$ contains two Weil divisors that are not Cartier, given by $\{y \pm f_3(x_1,x_2,x_3) = t = 0\}$. Choosing one of these divisors to blow up, we find that the resolved family $\calS' \to \Delta$ is semistable and the exceptional locus is eighteen copies of $\Proj^1$. As this resolution has not introduced any new divisors (it is an example of a \emph{small resolution}: a resolution with exceptional locus of codimension $\geq 2$), we must have $\omega_{\calS'} \cong \calO_{\calS'}$, so $\calS'$ is a Kulikov model. Its central fibre is a copy of $\Proj^2$ glued to a rational surface (obtained by blowing up $\Proj^2$ at eighteen points) along a smooth elliptic curve. This is an example of a Type II degeneration from Thm. . At this point we make a crucial note: when we performed the resolution to go from $\calS$ to $\calS'$, we had a choice of which divisor to blow up. This illustrates an important point, which is that \emph{Kulikov models are not unique}. In fact, if we allow ourselves to perform analytic blow ups (and, as noted in Rmk. , in certain cases we have to, as an algebraic resolution with trivial canonical bundle will not always exist) then the situation gets much worse, as we have to make a choice of which divisor to blow up \emph{locally} in a neighbourhood of each node. In our example above this gives $2^{18}$ possible analytic Kulikov models! \end{example} \begin{example}[Type III degeneration] Next we look at a Type III example. Consider the family $\pi\colon \calS \to \Delta := \{t \in \C \mid |t| < \varepsilon\}$ given in $\Proj^3 \times \Delta$ by the formula \[wxyz + tf_4(w,x,y,z) = 0,\] where $(w,x,y,z)$ are coordinates on $\Proj^3$ and $f_4$ is a generic homogeneous polynomial of degree $4$ in $(w,x,y,z)$. The general fibre of $\calS$ is a quartic hypersurface in $\Proj^3$ which, by Ex. , is a generic polarized K3 surface of degree four. The central fibre $S_0 = \pi^{-1}(0)$ is isomorphic to four copies of $\Proj^2$ given by the coordinate hyperplanes in $\Proj^3$. As in the previous example, we have $\omega_{\calS} \cong \calO_{\calS}$, but $\calS$ is singular and so not a Kulikov model. There are $24$ singularities, occurring at the intersections of $\{f_4(w,x,y,z) = 0\}$ with the six lines $\{w = x = 0\}$, $\{w = y = 0\}$, etc. As before, each of these singularities is locally analytically isomorphic to a threefold node $\{wx + tz = 0\} \subset \C^4$. We may resolve these singularities in the same way as the previous example to get a Kulikov model. The central fibre consists of four rational surfaces meeting along six rational curves, with dual graph homeomorphic to a tetrahedron. This is an example of a Type III degeneration from Thm. . \end{example} \subsection{Compactifications} Given that we have such a good description of the moduli space for pseudo-polarized K3 surfaces, it is natural to ask whether there is a nice way to compactify this moduli space, i.e. find a compact variety $\overline{\calF}_{2k}$ that contains $\calF_{2k}$ as an open subset. Preferably, one would like to do this in such a way that the boundary $\overline{\calF}_{2k} - \calF_{2k}$ encodes some geometric data about the corresponding degenerate K3 surfaces (in fact, ideally, we would like the boudary to provide moduli for degenerate K3's). Probably the best known compactification of $\calF_{2k}$ is the \emph{Baily-Borel compactification}, first constructed in . This is a method to compactify any arithmetic quotient of a bounded symmetric domain, of which $\calF_{2k}$ is an example. In the case of $\calF_{2k}$, this compactification was studied in detail by Scattone . Its boundary is a union of $0$- and $1$-dimensional strata, which have some geometric meaning: the $0$-dimensional strata correspond to degenerate K3's of Type III, and the $1$-dimensional strata to degenerate K3's of Type II. Furthermore, the $1$-dimensional strata are all rational curves, each of which is parametrised by the $j$-invariant of the elliptic double curves appearing in the corresponding Type II degenerate K3. \begin{example}[K3 surfaces of degree $2$] In the case of $\calF_2$, Friedman \cite[Sect. 5]{npgttk3s} showed that the boundary of the Baily-Borel compactification has four components of dimension $1$, which meet in a unique $0$-dimensional component. Furthermore, he gave a coarse geometric description of the degenerate fibres corresponding to each Type II boundary component; one of the four possibilities is given by the central fibre in Ex. . \end{example} \begin{example}[K3 surfaces of degree $4$] In the case of $\calF_4$, Scattone \cite[Sect. 6]{cmsak3s} has shown that the boundary of the Baily-Borel compactification has nine $1$-dimensional components, which meet in a unique $0$-dimensional component. This serves to illustrate that the number of boundary components in the Baily-Borel compactification of $\calF_{2k}$ grows very quickly with $k$; in fact, Scattone has shown that it grows like $k^8$. \end{example} Unfortunately this is about all one can say about the Baily-Borel compactification: the boundary is simply too small to encode more detailed geometric data about degenerate K3's, let alone provide moduli for them. Several other compactifications also exist. \emph{Toroidal compactifications}, originally constructed by Mumford , can be described as toroidal blow-ups of the boundary in the Baily-Borel compactification; a brief overview of this construction may be found in \cite[Sect. 2.2]{cmsak3s}. They have the advantage of being fairly easy to construct and their boundary components admit a clean explicit description. Furthermore, they can encode more detailed geometric data about degenerate K3's than is found in the Baily-Borel compactification (for instance, see ). However, there is a large collection of such compactifications, corresponding to possible choices of blow-up, with no clear canonical choice amongst them. One may also compactify using tools from the log minimal model program developed by Koll\'{a}r, Shepherd-Barron and Alexeev . Such so-called \emph{KSBA compactifications} avoid many of the problems encountered by toroidal compactifications, but come with some of their own. They can be defined in a way that avoids choices, making them in some sense canonical, and their construction ensures that their boundary provides moduli for degenerate K3 surfaces. However, it is difficult to describe the boundaries of KSBA compactifications explicitly, making them hard to study. Probably the best compactifications currently known are those that use techniques from Geometric Invariant Theory. Such \emph{GIT compactifications} were originally studied by Shah and, much more recently, Laza . These solve both problems: they admit clean explicit descriptions and their boundaries provide moduli for degenerate K3 surfaces. However, they are difficult to construct in general, so thus far have only been studied for small values of $k$: and treat only the case $k = 1$, and treats the case $k = 2$. \section{Lattice Polarized K3 Surfaces} This general theory is all well and good, but so far we have relatively few explicit examples of K3 surfaces to work with (the polarized K3 surfaces of Exs. and are effectively the only ones we have constructed). Many more interesting families of K3 surfaces may be studied using the machinery of lattice polarizations. The concept of lattice polarization extends the idea of polarization discussed in Sect. . It was first introduced by Dolgachev , whose results we will summarize here. We begin with a central definition. \begin{definition} Let $S$ be an projective K3 surface and let $L$ be a non-degenerate lattice. Assume that there is a lattice embedding $\iota \colon L \hookrightarrow \NS(S)$. Then we say that the pair $(S,\iota)$ is a (\emph{pseudo-})\emph{ample $L$-polarized K3 surface} if \begin{enumerate} \item The embedding $\iota$ is primitive, and \item The image of $\iota$ contains an (pseudo-)ample class. \end{enumerate} Two (pseudo-)ample $L$-polarized K3 surfaces $(S,\iota)$ and $(S',\iota')$ are \emph{equivalent} if there is an isomorphism $f\colon S \rightarrow S'$ of K3 surfaces such that $f^* \circ \iota' = \iota$ and $f^*$ takes some (pseudo-)ample class on $S'$ to a (pseudo-)ample class on $S$. \end{definition} We will often informally use the phrase ``$L$-polarization'' to indicate a pseudo-ample $L$-polarization. The argument from Rmk. shows that every pseudo-ample $L$-polarized K3 surface is projective. This definition is a natural generalization of the concept of polarization from Sect. . Indeed, if $L$ is a non-degenerate lattice of rank $1$ and if $(S,\iota)$ is a pseudo-ample $L$-polarized K3 surface, then the image $\iota(L)$ contains a pseudo-ample class $h$ with $\langle h, h\rangle > 0$. Furthermore, since $\Lambda_{\mathrm{K}3}$ is even, we must have $\langle h, h\rangle = 2k > 0$, so $(S,h)$ is a pseudo-polarized K3 surface of degree $2k$. More examples of lattice polarized K3 surfaces will be given in Sects. and . In general, since $\iota(L)$ contains a pseudo-ample class, $L$ must contain some $h$ with $\langle h,h \rangle > 0$. This, along with the condition that $\iota(L)$ must be contained in $\NS(S)$ and the Signature Theorem \cite[Thm. IV.2.14]{bpv}, shows that $L$ must be a lattice of signature $(1,n-1)$ where $n$ is the rank of $L$. \subsection{Moduli of Lattice Polarized K3 Surfaces} We next discuss the moduli space of K3 surfaces with lattice polarization. The construction works in largely the same way as the construction of the moduli space for polarized K3 surfaces discussed in Sect. . Fix once and for all a nondegenerate lattice $L$ of rank $n$ and a primitive embedding of $L$ into $\Lambda_{\mathrm{K}3}$. We consider $L$ as a sublattice of $\Lambda_{\mathrm{K}3}$. Then a marked K3 surface $(S,\phi)$ is a \emph{marked \textup{(}pseudo-\textup{)}ample $L$-polarized K3 surface} if the restriction $\phi^{-1}|_L \colon L \rightarrow H^2(S,\mathbb{Z})$ is a (pseudo-)ample $L$-polarization on $S$. Note that for an generator $\sigma$ of $H^{2,0}(S)$, we have $\langle \phi_\mathbb{C}(\sigma), u \rangle = 0$ for any $u \in L$, since $\phi^{-1}(u)$ is contained in $\NS(S)$. Thus the period points of marked $L$-polarized K3 surfaces are contained in \[\Omega_{L^\perp} := \{[\sigma] \in \mathbb{P}(\Lambda_{\mathrm{K}3} \otimes \C) \mid \langle \sigma,\sigma\rangle = 0,\ \langle \sigma, \overline{\sigma} \rangle > 0,\ \langle \sigma, L \rangle = 0 \} \subset \Omega_{\mathrm{K}3}.\] $\Omega_{L^\perp}$ called the \emph{period space of pseudo-ample $L$-polarized K3 surfaces}. It is a complex manifold of dimension $(20 - n)$ with two connected components, each of which is a bounded symmetric domain of type IV \cite[Sect. 3]{mslpk3s}. \begin{remark} By \cite[Sect. X.6.3]{dglgss}, a bounded symmetric domain of type IV (also called ``type BD I ($q = 2$)'' in ) and complex dimension $n$ is isomorphic to $\mathrm{SO}_0(2,n)/\mathrm{SO}(2) \times \mathrm{SO}(n)$. Furthermore, by the isomorphisms in \cite[Sect. X.6.4]{dglgss}, in dimensions $1$, $2$ and $3$ these domains coincide with the classical modular domains $\mathbb{H}$, $\mathbb{H} \times \mathbb{H}$ and $\mathbb{H}_2$ respectively, where $\mathbb{H}$ denotes the upper half space in $\mathbb{C}$ and $\mathbb{H}_2$ denotes the Siegel upper half space of genus $2$. \end{remark} By the Surjectivity of the Period Map (Thm. ), every point in $\Omega_{L^\perp}$ is the period point of some marked K3 surface $(S,\phi)$. Furthermore, for any such marked K3 surface we still have $\langle \sigma, L \rangle = 0$, so $\phi^{-1}(L)$ is contained in $\NS(S)$. In fact, outside of certain codimension $1$ loci, $\phi^{-1}(L)$ actually contains an ample class on $S$. Define, as before, \[\Delta_{L^\perp}:= \{ \delta \in \Lambda_{\mathrm{K}3} \mid \langle \delta,\delta\rangle = -2,\ \langle \delta,L \rangle = 0\}.\] To each $\delta$ in $\Delta_{L^\perp}$, we may assign a hyperplane $H_\delta$ as in Eq. \eqref{eq:H}. Then define \[\Omega^0_{L^\perp} := \Omega_{L^\perp} - \bigcup_{\delta \in \Delta_{L^\perp}} (H_\delta \cap \Omega_{L^\perp}).\] $\Omega^0_{L^\perp}$ called the \emph{period space of ample $L$-polarized K3 surfaces}. Dolgachev proved that: \begin{theorem}\textup{\cite[Thm. 3.1 and Cor. 3.2]{mslpk3s}} The points of the space $\Omega^0_{L^\perp}$ \textup{(}resp. $\Omega_{L^\perp}$\textup{)} are in bijection with ample \textup{(}resp. pseudo-ample\textup{)} marked $L$-polarized K3 surfaces. \end{theorem} Using the Torelli Theorems (Thms. and ) and the Surjectivity of the Period Map (Thm. ), this theorem is not very difficult to prove; the proof essentially amounts to notation and keeping track of K\"ahler data. We refer the interested reader to \cite[Sect. 2]{mslpk3s} for details. To construct a coarse moduli space for (pseudo-)ample $L$-polarized K3 surfaces, we once again have to perform a quotient to get rid of the choice of marking. Let $\Gamma_{L^\perp}$ be the subgroup of elements of $\mathrm{O}(\Lambda_{\mathrm{K}3})$ satisfying \[\Gamma_{L^\perp} := \{ \gamma \in \mathrm{O}(\Lambda_{\mathrm{K}3}) \mid \gamma |_L = \Id\},\] i.e. the subgroup of elements of $\mathrm{O}(\Lambda_{\mathrm{K}3})$ which fix $L \subseteq \Lambda_{\mathrm{K}3}$ pointwise. Then we find: \begin{theorem}\textup{\cite[Rmk. 3.4]{mslpk3s}} The quotients \[\mathcal{F}_{L^\perp}^0 := \Gamma_{L^\perp} \setminus \Omega_{L^\perp}^0 \] and \[\mathcal{F}_{L^\perp}:=\Gamma_{L^\perp} \setminus \Omega_{L^\perp}\] are coarse moduli spaces for ample and pseudo-ample $L$-polarized K3 surfaces respectively. \end{theorem} As before, $\mathcal{F}_{L^\perp}$ may be seen as an arithmetic quotient of a bounded symmetric domain of type IV (given by one of the connected components of $\Omega_{L^{\perp}}$), so $\mathcal{F}_{L^\perp}$ is connected and one may use the work of Baily-Borel to show that both $\mathcal{F}^0_{L^\perp}$ and $\mathcal{F}_{L^\perp}$ are quasi-projective varieties. Note that $\mathcal{F}_{L^\perp}$ has dimension $(20 - n)$. \begin{example}[Lattices of rank 19] If $L$ has rank $19$ then we see that $\Omega_{L^\perp}$ is a $1$-dimensional space so, by Rmk. , each connected component of $\Omega_{L^\perp}$ is analytically isomorphic to $\mathbb{H}$, the upper half plane in $\mathbb{C}$. We thus see that $\calF_{L^\perp}$ is isomorphic to the quotient of $\mathbb{H}$ by the action of the group $\Gamma_{L^\perp}$. It can be shown that this group acts on $\mathbb{H}$ as a discrete subgroup of $\mathrm{SL}_2(\mathbb{R})$. For instance, if $L \cong H \oplus (-E_8) \oplus (-E_8) \oplus \left<-2n\right>$, then \cite[Thm. 7.1]{mslpk3s} shows that the group $\Gamma_{L^\perp}$ acts on $\mathbb{H}$ in the same way as $\Gamma_0(n)^+$. Thus \[\mathcal{F}_{L^\perp} = \Gamma_0(n)^+ \setminus \mathbb{H},\] which is a classical modular curve. A more complicated example is provided by the lattice $L \cong H \oplus (-E_8) \oplus (-D_7) \oplus (-A_2)$. In this case, Elkies \cite[Sect. 3]{scck3snsr19} has proved that $\mathcal{F}_{L^\perp}$ is isomorphic to a quotient of the \emph{Shimura curve} $X(6)$. We refer the interested reader to for full details of this and several other Shimura curve examples. \end{example} \begin{example}[Lattices of rank 18] If $L$ is a lattice of rank 18, then we find a similar structure. In this case, by Rmk. , each connected component of $\Omega_{L^\perp}$ is analytically isomorphic to the product $\mathbb{H} \times \mathbb{H}$, and $\calF_{L^\perp}$ is isomorphic to the quotient of $\mathbb{H} \times \mathbb{H}$ by the action of $\Gamma_{L^\perp}$. This group acts on $\mathbb{H} \times \mathbb{H}$ as a subgroup of $\left((\mathrm{SL}_2(\mathbb{R} ) \times \mathrm{SL}_2(\mathbb{R})\right) \rtimes \Z / 2\Z$, where the $\Z / 2\Z$ acts to exchange the factors of $\mathbb{H} \times \mathbb{H}$. In particular, if \[L \cong \left( \begin{matrix} 2 & D \\ D & (D^2 - D)/2 \end{matrix} \right) \oplus (-E_8) \oplus (-E_8)\] for $D$ square-free, then $\mathcal{F}_{L^\perp}$ is birational to the classically known \emph{Humbert surface of discriminant $D$} by \cite[Thm. 9]{k3sehms}. If $L \cong H \oplus (-E_8) \oplus (-E_8)$, then $\Gamma_{L^\perp}$ is isomorphic to $\left(\mathrm{SL}_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})\right) \rtimes \Z / 2\Z$. In this case $\mathcal{F}_{L^\perp}$ is isomorphic to the symmetric product of two copies of the classical modular curve; see or for more details. \end{example} One may notice that the moduli spaces of $L$-polarized K3 surfaces with $\rank L \geq 17$ are often related to the moduli spaces of abelian surfaces. This is no accident. According to , K3 surfaces with lattice polarization by $L = (-E_8) \oplus (-E_8) \oplus \langle 2n \rangle$ can be constructed geometrically from abelian surfaces with $(1,n)$ polarization. More generally, Galluzzi and Lombardo show that a large class of rank $17$ polarized K3 surfaces admit algebraic correspondences with abelian surfaces. This is part of a more general relation between periods of K3 surfaces and abelian varieties called the Kuga-Satake construction, which was first described in . A modern introduction to this construction may be found in . \subsection{Examples of Lattice Polarized K3 Surfaces} In this subsection we will provide some simple examples of lattice polarized K3 surfaces. The aim is to show that the lattice polarization construction given above arises naturally and geometrically. \subsubsection{Examples Arising from Embeddings} If $S$ is a K3 surface embedded as a subvariety of a smooth projective variety $X$, then there is a natural restriction map $r\colon \NS(X)\otimes \mathbb{Q} \rightarrow \NS(S)\otimes \mathbb{Q}$. Let $L$ be the lattice \[L:= \Image r \cap \NS(S) \subseteq \NS(S) \otimes \mathbb{Q}.\] We claim that $S$ is $L$-polarized. The image of this restriction morphism is non-trivial, first of all, since any ample class on $X$ restricts to an ample class on $S$; this also shows that condition 2 of Def. is satisfied. By construction, the lattice $L$ is a primitive sublattice of $\NS(S)$, so condition 1 of Def. is satisfied. Therefore $S$ is $L$-polarized, as claimed. As an interesting example of this, let $X$ be a smooth Fano threefold (i.e., a threefold whose anticanonical bundle is ample) and let $S$ be a smooth member of the anticanonical linear system $|-K_X|$; by adjunction $S$ is a K3 surface. The Lefschetz Hyperplane Theorem shows that the map $r\colon \NS(X) \rightarrow \NS(S)$ is a a primitive embedding. Furthermore, by , the map $r$ can be seen as a lattice embedding if we equip $\NS(X)$ with the bilinear form \begin{equation} \langle u,v \rangle = \langle [-K_X],u,v \rangle_X, \end{equation} where $\left<\cdot,\cdot,\cdot\right>_X$ is the trilinear intersection form on $\NS(X)$. \begin{example}[Anticanonical hypersurfaces in $\mathbb{P}^1 \times \mathbb{P}^2$] For instance, if we take $X = \mathbb{P}^1 \times \mathbb{P}^2$, then $\NS(X) \cong \mathbb{Z}^2$ is spanned by the divisor classes \[D_1 := [p \times \mathbb{P}^2], \quad D_2 := [\mathbb{P}^1 \times H]\] where $p$ is a point in $\mathbb{P}^1$ and $H$ is a line in $\mathbb{P}^2$. One may check that $-K_X \sim i_1^* (-K_{\mathbb{P}^1}) + i_2^*(-K_{\mathbb{P}^2})$ where $i_1$ is the projection map onto $\mathbb{P}^1$ and $i_2$ is the projection map onto $\mathbb{P}^1$. Thus $[-K_X] = 2D_1 + 3D_2$. Now it is easy to compute that \[\langle D_1, D_1, D \rangle_X = 0\] for any divisor $D$, and that \[\langle D_1, D_2, D_2\rangle_X = 1,\quad \langle D_2,D_2,D_2 \rangle_X = 0.\] Using Eq. \eqref{equation:intersection} we deduce that any $S$ embedded as an anticanonical hypersurface in $X$ is lattice polarized by a lattice of rank $2$ with Gram matrix \[\left( \begin{matrix} 0 & 3 \\ 3 & 2 \end{matrix} \right).\] \end{example} If $X$ is a smooth toric weak Fano threefold (i.e. a threefold whose anticanonical bundle is pseudo-ample) then one may effectively compute the induced lattice polarization induced on the anticanonical K3 surface $S$ using toric geometry. A large number of K3 surfaces of this form were found by Reid , as toric resolutions of hypersurfaces in weighted projective spaces (the full list is given in \cite[Sect. 13.3]{wwwci}); the induced lattice polarizations on them were computed by Belcastro . Further examples of this type have been computed by Rohsiepe . \subsubsection{Examples Arising from Singularities} Another useful way of producing lattice polarizations on K3 surfaces is by introducing controlled singularities. In particular, if $\overline{S}$ is a compact algebraic surface with singularities of ADE type (see, for instance \cite[Sect. 4.2]{chapters}) and trivial dualizing sheaf $\omega_{\overline{S}} \cong \calO_{\overline{S}}$, then the minimal resolution $S$ of $\overline{S}$ is a K3 surface and the exceptional divisor associated to each singularity is a configuration of rational curves whose dual intersection graph is a Dynkin diagram of ADE type. The associated (negative definite) root system is then contained inside of the N\'eron-Severi lattice of $S$. \begin{example}[Nodal quartics] Let $\overline{S}$ be a quartic hypersurface in $\mathbb{P}^3$ with a single $A_1$ singularity and minimal resolution $S \to \overline{S}$. The surface $S$ is a pseudo-polarized K3 surface of degree $4$ but is \emph{not} polarized, since the embedding into $\mathbb{P}^3$ induced by the hyperplane section is not a smooth embedding and so the class in $\NS(S)$ coming from the restriction of the hyperplane class in $\mathbb{P}^3$ is only pseudo-ample. After resolving the singularity, we obtain an exceptional rational curve $C$ in $S$. The class of $C$ has self-intersection $(-2)$ by the Riemann-Roch Theorem for surfaces. We also see that $\langle [H|_S], [C] \rangle = 0$ for a generic hyperplane section $H$ of $\mathbb{P}^3$. Therefore there is an embedding of the rank two lattice $L$ with Gram matrix \[\left( \begin{matrix} -2 & 0 \\ 0 & 4 \end{matrix} \right)\] into $\NS(S)$. As we have already observed, the class $[H|_S]$ is pseudo-ample, thus condition 2 of Def. holds. Furthermore, if $L$ were not embedded primitively into $\NS(S)$, then there would be some even integral sublattice of $L\otimes \mathbb{Q}$ containing $L$. One may argue using Thm. from the appendix that no such sublattice exists. Thus $L$ is a primitive sublattice of $\NS(S)$, so condition 1 of Def. is also satisfied. Therefore, we see that $S$ is $L$-polarized. We should note that, in terms of the moduli space of pseudo-polarized K3 surfaces of degree $4$ constructed in Sect. , such K3 surfaces lie in one of the hyperplanes $H_\delta$. \end{example} In general, one may try to produce lattice polarizations by imposing specific configurations of singularites. However, some care is required. Suppose that $\overline{S}$ is a singular surface with trivial dualizing sheaf $\omega_{\overline{S}} \cong \calO_{\overline{S}}$, a primitive pseudo-ample class $h \in \NS(\overline{S})$ with $\langle h,h \rangle = 2k$, and ADE singularities of types $R_1, \ldots, R_n$. Then the K3 surface $S$ obtained by resolving the singularities of $\overline{S}$ will not necessarily be lattice polarized by $\langle 2k \rangle \oplus (-R_1) \oplus \dots \oplus (-R_n)$, since this lattice may not be primitively embedded in $\NS(S)$. This phenomenon is illustrated in the following example. \begin{example}[Kummer surfaces] Suppose that $A$ is an abelian surface. The involution $\iota\colon A \to A$ given by $\iota(x) = -x$ (defined using the group law on $A$) has sixteen fixed points. If we quotient $A$ by this involution, we obtain a projective surface $\overline{S}$ with trivial dualizing sheaf and sixteen singularities of type $A_1$. Each of these singularities may be resolved by blowing up once, giving sixteen disjoint exceptional $(-2)$-curves $E_1,\ldots,E_{16}$. The resolution is a projective K3 surface $S$, called the \emph{Kummer surface associated to $A$}. Now let $h \in \NS(\overline{S})$ be a primitive ample class and suppose that $\langle h,h \rangle = 2k$. Then from the discussion above, there is an embedding of the lattice $L := \langle 2k \rangle \oplus (-A_1)^{\oplus 16}$ into $\NS(S)$. However, this embedding is not primitive, as we shall now demonstrate. Begin by noting that there exists a double cover of $S$ branched along the divisor $\sum_{i=1}^{16} E_i$: this cover is precisely the (non-minimal) surface obtained by blowing up the sixteen fixed points of $\iota$ in $A$. However, the existence of this cover is equivalent to the existence of a divisor $D \in \Pic(S)$ with $2D \cong \sum_{i=1}^{16} E_i$. The class $[D]$ of $D$ in $\NS(S)$ is thus equal to $\frac{1}{2}\sum_{i=1}^{16} [E_i]$, which lies in $\frac{1}{2}L$ but not in $L$. So we have found a class $[D] \in \NS(S)$ that lies in $L \otimes \mathbb{Q}$ but not in $L$ and, therefore, the embedding of $L$ into $\NS(S)$ cannot be primitive. \end{example} In the next section, we will see that the problem of primitive embeddings has a nice solution in the case where $\overline{S}$ is a singular elliptically fibred K3 surface. \section{Elliptically Fibred K3 Surfaces} In this section, we will give a detailed discussion of the construction of K3 surfaces through elliptic fibrations. As we shall see, this is an excellent source of K3 surfaces with lattice polarizations, but in order to get there we will have to develop some of the theory of elliptically fibred surfaces first. Most of the theory presented here was originally developed by Kodaira and Tate . A self-contained reference for the reader interested in algebraic elliptic surfaces over $\C$ is Miranda's book , which we will use as our main reference. However, readers who are interested in elliptic surfaces over arbitrary characteristic might find it useful to consult the more algebraic , whereas those interested in arithmetic results on elliptic surfaces may find or more to their tastes. In addition to this, both and contain discussions of elliptic fibrations on complex manifolds. \subsection{Elliptic Fibrations and $H$-Polarizations} We begin by studying the construction of elliptic surfaces, with an emphasis on K3 surfaces. Start by letting $S$ be an arbitrary algebraic surface. A \emph{genus one fibration} on $S$ is a pair $(S,\pi)$ where $\pi$ is a projective morphism $\pi \colon S \rightarrow B$ to some smooth curve $B$, with generic fibre a smooth curve of genus one. We say that $(S,\pi)$ is an \emph{elliptic fibration} if $\pi$ admits a section $\mathbb{O} \colon B \rightarrow S$ such that $\pi \circ \mathbb{O} = \Id$. By the Leray spectral sequence we have $h^1(S,\mathcal{O}_S) \geq h^0(B,\mathcal{O}_B)$, with equality if and only if $S$ is not a product $B \times E$ of $B$ with an elliptic curve $E$ \cite[III.4.1]{btes}. Thus if we want $S$ to be a K3 surface, then $B$ must be $\mathbb{P}^1$. If $(S,\pi)$ is an elliptic fibration on a K3 surface $S$, then we have two obvious curve classes in $S$. Firstly we have the class of the image of $\mathbb{O}$, which we call $[O]$, and secondly we have the class of a fibre $E$. It is easy to see that \[\langle [O], [O] \rangle = -2,\quad \langle[E],[E] \rangle = 0, \quad \langle [E], [O] \rangle = 1.\] The sublattice of $\NS(S)$ given by $L:= \mathbb{Z}([O] - [E]) \oplus \mathbb{Z}[E]$ has Gram matrix \[\left( \begin{matrix} 0 & 1 \\ 1 & 0 \end{matrix} \right),\] so it is isomorphic to the hyperbolic plane lattice $H$. Now, the class $(2[E]+[O])$ in $L$ is pseudo-ample, thus condition 2 of Def. holds. Furthermore, if $L$ were not embedded primitively into $\NS(S)$, then there would be some even integral sublattice of $L\otimes \mathbb{Q}$ containing $L$. Since $L$ is unimodular, Thm. from the appendix shows that no such sublattice exists. Thus $L$ is a primitive sublattice of $\NS(S)$, so condition 1 of Def. is also satisfied. We therefore see that any elliptic fibration on a K3 surface $S$ defines an $H$-polarization on $S$. Conversely, it is easy to see that any $H$-polarization on $S$ determines an elliptic fibration. In particular, this shows that any elliptically fibred K3 surface is projective. \begin{remark} We remind the reader here that our definition of an elliptic fibration comes with a section: the statements above are not necessarily true for more general genus one fibrations. In particular, a K3 surface with a genus one fibration may not admit an $H$-polarization (see Ex. ) and does not have to be projective. \end{remark} \subsection{Singular Fibres} To find a way to construct elliptic surfaces, we must dig a little deeper into the geometry of an elliptic fibration $\pi\colon S \rightarrow B$. We begin with a definition. \begin{definition} An elliptic fibration $\pi\colon S \rightarrow B$ is called \emph{relatively minimal} if it is smooth and there is no rational curve $C$ in $S$ with self-intersection $(-1)$ and $\pi(C)$ a point. \end{definition} Essentially, relative minimality means that all curves inside of fibres of $\pi$ that can be contracted smoothly have been contracted. Note that if $S$ is a minimal surface to begin with, it will certainly be a relatively minimal elliptic surface. In particular, this means that any elliptic fibration on a K3 surface is a relatively minimal elliptic fibration. On the other hand, consider a pair of cubics $C_1$ and $C_2$ in general position in $\mathbb{P}^2$. The intersection $C_1 \cap C_2$ consists of nine points, through which a pencil of cubics (given by taking projective linear combinations of the defining equations of $C_1$ and $C_2$) passes. If we let $S$ be the blow up of $\mathbb{P}^2$ at these nine points, then $S$ admits an elliptic fibration induced by the pencil of elliptic curves passing through the nine points. This fibration is relatively minimal, but the surface $S$ is not minimal (the nine $(-1)$-curves on $S$ are all sections of the fibration). We now turn our attention to the fibres of a relatively minimal elliptic fibration $\pi\colon S \to B$. Generically the fibres of $\pi$ are smooth elliptic curves. However at certain points, the fibres of $\pi$ may degenerate to singular curves. The number and type of these singular fibres is somewhat controlled by the following theorem: \begin{theorem} \textup{\cite[Lemma IV.3.3]{btes}} Suppose that $\pi\colon S \to B$ is a relatively minimal elliptic fibration. Let $e(S)$ be the Euler characteristic of the surface $S$ and let $e(p)$ be the Euler characteristic of a fibre $\pi^{-1}(p)$ of $\pi$. Then \[e(S) = \sum_{p\in B} e(p).\] \end{theorem} Note that this is actually a finite sum since the Euler characteristic of a smooth fibre is $0$. Since a K3 surface has Euler characteristic $24$, an elliptic fibration on a K3 surface must have singular fibres. A theorem, originally due to Kodaira , classifies the singular fibres of smooth minimal elliptic fibrations. This theorem will be presented in full generality in Sect. , but for now the following will suffice. \begin{theorem}\textup{\cite[Sect. I.4]{btes}} Let $\pi\colon S \rightarrow B$ be a smooth relatively minimal elliptic fibration. Then any fibre $E$ of $\pi$ is either: \begin{enumerate} \item Irreducible, in which case $E$ is either a smooth elliptic curve, or a nodal or cuspidal rational curve; or \item Reducible, in which case $E$ is a configuration of smooth rational curves $C_i$ with $\langle [C_i],[C_i]\rangle = -2$. In this case $E$ is either a pair of rational curves that are tangent at a point, three rational curves meeting at a single point, or a configuration of rational curves meeting transversely with dual intersection graph of extended ADE type. Furthermore, there are positive integers $a_i$ such that \begin{equation} [E] = \sum_{i = 0}^n a_i [C_i]. \end{equation} \end{enumerate} \end{theorem} \subsection{Weierstrass Fibrations} We have seen that if a singular fibre is not a singular elliptic curve, it is composed of smooth rational curves with self-intersection $(-2)$. If we relabel the curves in each singular fibre so that $C_0$ is the unique rational curve intersecting the section $O$, one sees that the configuration \[\sum_{i = 1}^na_i [C_i]\] has self-intersection $(-2)$ and can thus be contracted to an ADE singularity. Once this contraction has been performed, the component $C_0$ becomes a rational curve with a single node or cusp. If $(S,\pi)$ is a smooth relatively minimal elliptically fibred surface, then let $(\overline{S},\pi)$ be the singular elliptically fibred surface obtained by contracting components of all singular fibres as above. In this way, we may obtain an elliptically fibred surface whose fibres are all irreducible and have arithmetic genus one. Following \cite[Def. II.3.2]{btes}, we call such a surface a \emph{Weierstrass fibration} and we say that $\overline{S}$ is a \emph{Weierstrass model} for $S$. The rationale for these names will become clear in a moment. One may show without much difficulty (see \cite[II.3.5]{btes}) that if $\pi\colon \overline{S} \rightarrow B$ is a Weierstrass fibration, then the sheaf $R^1\pi_* \mathcal{O}_{\overline{S}}$ is a line bundle. \begin{definition} Let $\pi\colon \overline{S} \rightarrow B$ be a Weierstrass fibration. The \emph{fundamental line bundle} of $(\overline{S},\pi)$ is defined as \[\mathbb{L} := (R^1\pi_*\mathcal{O}_{\overline{S}})^{-1}.\] \end{definition} Using this theory, we are able to give a method by which elliptic fibrations can be explicitly constructed. \begin{theorem}\textup{\cite[Sect. III.1]{btes}} Let $(\overline{S},\pi)$ be a Weierstrass fibration over a smooth curve $B$. Then \begin{enumerate} \item There is an embedding, \[f\colon \overline{S} \hookrightarrow \mathbb{P}(\mathcal{O}_B \oplus \mathbb{L}^{-2} \oplus \mathbb{L}^{-3}).\] \item If $p\colon \mathbb{P}(\mathcal{O}_B \oplus \mathbb{L}^{-2} \oplus \mathbb{L}^{-3}) \rightarrow B$ is the natural projection map, then $p \circ f = \pi$. \item The hypersurface $\overline{S}$ is given by the vanishing of a section of $\mathcal{O}_\mathbb{P}(1)^3 \otimes p^*\mathbb{L}^6$ where $\mathcal{O}_\mathbb{P}(1)$ is the inverse of the tautological bundle on the projective bundle $\mathbb{P}(\mathcal{O}_B \oplus \mathbb{L}^{-2} \oplus \mathbb{L}^{-3})$. \item $\overline{S}$ can be written as the vanishing locus of \[ZY^2 = X^3 + \alpha XZ^2 + \beta Z^3\] where $\alpha$ and $\beta$ are global sections of $p^*\mathbb{L}^4$ and $p^*\mathbb{L}^6$, and $X,Y,Z$ are global sections of $\mathcal{O}_\mathbb{P}(1) \otimes p^*\mathbb{L}^2$, $\mathcal{O}_\mathbb{P}(1) \otimes p^*\mathbb{L}^3$ and $\mathcal{O}_\mathbb{P}(1)$ respectively. \end{enumerate} \end{theorem} \begin{remark} We will often refer to $\alpha$ and $\beta$ as sections of $\mathbb{L}^4$ and $\mathbb{L}^6$ respectively, using the fact that the spaces of sections of $\mathbb{L}^4$ and $\mathbb{L}^6$ are isomorphic to the spaces of sections of their pull-backs under $p$. \end{remark} This explains the meaning of the name ``Weierstrass fibration'': such surfaces admit expressions which are completely analogous to the Weierstrass form of an elliptic curve over a number field. This may be viewed as a refined version of the fact that the generic fibre of $(\overline{S},\pi)$ is an elliptic curve over the function field $\mathbb{C}(B)$, which may itself be expressed in Weierstrass form (see, for instance, \cite[Chap. II]{btes}). Our next aim is to find conditions under which this construction gives a K3 surface. We already know that $h^1(S,\mathcal{O}_S) = 0$ if and only if the base curve $B$ is $\Proj^1$, so it just remains to compute the canonical bundle. It is easy to compute the dualizing sheaf of a Weierstrass fibration. If $\omega_B$ is the canonical bundle on the curve $B$, then \[\omega_{\mathbb{P}(\mathcal{O}_B \oplus \mathbb{L}^{-2} \oplus \mathbb{L}^{-3})} \cong p^*(\omega_B \otimes \mathbb{L}^{-5}) \otimes \calO_{\Proj}(1)^{-3}. \] Using this, adjunction gives: \begin{proposition}\textup{\cite[Prop. III.1.1]{btes}} The dualizing sheaf of $\overline{S}$ is given by \[\omega_{\overline{S}} = \pi^*(\omega_B \otimes \mathbb{L}).\] \end{proposition} If $(S,\pi)$ is an elliptically fibred K3 surface and $(\overline{S},\pi)$ is the Weierstrass model associated to $(S,\pi)$, then $\omega_{\overline{S}} \cong \calO_{\overline{S}}$. Therefore, if $S$ is a K3 surface, then $\mathbb{L} \cong \omega_{\Proj^1}^{-1} \cong \mathcal{O}_{\mathbb{P}^1}(2)$. The converse is also true: if $\mathbb{L} \cong \mathcal{O}_{\mathbb{P}^1}(2)$ then the minimal resolution $S$ of the Weierstrass model $\overline{S}$ is a K3 surface (the proof is easy, but relies upon the fact that the exceptional curves in the resolution of an ADE singularity do not contribute to the canonical bundle; in technical language we say that ADE singularities admit \emph{crepant} resolutions). Thus we find: \begin{proposition} Let $\pi\colon S \to B$ be an elliptic fibration, $(\overline{S},\pi)$ be its Weierstrass model, and $\mathbb{L} := (R^1\pi_*\mathcal{O}_{\overline{S}})^{-1}$ be its fundamental line bundle. Then $S$ is a K3 surface if and only if $B = \Proj^1$ and $\mathbb{L} \cong \calO_{\Proj^1}(2)$. \end{proposition} Using this, we may express any elliptically fibred K3 surface as a hypersurface in $\mathbb{P}(\mathcal{O}_{\mathbb{P}^1} \oplus\mathcal{O}_{\mathbb{P}^1}(-4) \oplus\mathcal{O}_{\mathbb{P}^1}(-6))$ given by an equation \[ZY^2 = X^3 + \alpha(s,t)XZ^2 + \beta(s,t)Z^3\] where $(s,t)$ are coordinates on $\Proj^1$ and $\alpha(s,t)$ and $\beta(s,t)$ are homogeneous polynomials in $s$ and $t$ of degrees $8$ and $12$ respectively. \begin{remark} The projective bundle $\mathbb{P}(\mathcal{O}_{\mathbb{P}^1} \oplus\mathcal{O}_{\mathbb{P}^1}(-4) \oplus\mathcal{O}_{\mathbb{P}^1}(-6))$ admits a birational contraction map to the weighted projective space $\mathbb{WP}(1,1,4,6)$. Under this contraction, elliptic K3 surfaces are expressed as weighted projective hypersurfaces of the form \[y^2 = x^3 + \alpha(s,t)x + \beta(s,t)\] with $\alpha(s,t)$ and $\beta(s,t)$ exactly as above. Here $x$ has weight $4$, $y$ has weight $6$, and $s$, and $t$ both have weight $1$. \end{remark} \begin{remark} A general K3 surface with elliptic fibration is a smooth hypersurface in $\mathbb{P}(\mathcal{O}_{\mathbb{P}^1} \oplus\mathcal{O}_{\mathbb{P}^1}(-4) \oplus\mathcal{O}_{\mathbb{P}^1}(-6))$, which has Picard rank $2$. By the technique of Sect. , one may show that this embedding induces a pseudo-ample $H$-polarization on these K3 surfaces. This fits with our observation from Sect. , that elliptically fibred K3 surfaces all admit $H$-polarizations. \end{remark} \subsection{Singular Fibres Revisited} Now that we know how to construct elliptically fibred K3 surfaces, our next task is to find lattice polarizations on them. In order to do this, we will need to perform a closer study of the N\'{e}ron-Severi lattice of an elliptically fibred surface. We begin with a closer examination of the singular fibres. Assume that we begin with a (possibly singular) elliptically fibred surface $(\overline{S},\pi)$ in Weierstrass form. We would like to use the local behaviour of the Weierstrass equation of $(\overline{S},\pi)$ to describe the configurations of divisors arising from resolution of the singularities of $\overline{S}$. First of all, it is clear that a fibre of $(\overline{S},\pi)$ is a singular elliptic curve if and only if the discriminant of the cubic \[ZY^2 = X^3 + \alpha XZ^2 + \beta Z^3\] vanishes. As usual, we may express this discriminant as a polynomial in $\alpha$ and $\beta$, giving \[\Delta = 4\alpha^3 + 27\beta^2 \in H^0(B,\mathbb{L}^{12}).\] The points at which $\Delta$ vanishes correspond to the discriminant locus of the fibration $(\overline{S},\pi)$. Kodaira and Tate showed how to use the local behaviour of $\alpha,\beta$ and $\Delta$ to detect singularities in the surface $\overline{S}$ and computed the minimal resolutions of these singularities. \begin{theorem}\textup{\cite[Sect. IV.3]{btes}} Let $(S,\pi)$ be a smooth relatively minimal elliptically fibred surface and let $(\overline{S},\pi)$ be its Weierstrass model. Let $\mathbb{L}$ be the fundamental line bundle of $(\overline{S},\pi)$ and let $\alpha \in H^0(B,\mathbb{L}^4)$ and $\beta \in H^0(B,\mathbb{L}^6)$ be the sections defining $\overline{S}$. Let $\Delta \in H^0(B,\mathbb{L}^{12})$ be the discrimant. Denote by $\nu_p(\alpha)$ the order of vanishing of $\alpha$ at the point $p$, by $\nu_p(\beta)$ the order of vanishing of $\beta$ at $p$ and by $\nu_p(\Delta)$ the order of vanishing of $\Delta$ at $p$. Then the fibre $\pi^{-1}(p)$ is singular if and only if $\Delta(p) = 0$. The singularity of $\overline{S}$ lying over $p$, along with a description of the singular fibre $\pi^{-1}(p)$ in $S$ and its Euler characteristic $e$ are given by Table . \end{theorem} \begin{table} \caption{Singular fibres of a smooth relatively minimal elliptically fibred surface} \small{ \begin{tabular}{ccccccp{4cm}} \hline\noalign{\smallskip} Name & $\nu_p(\alpha)$ & $\nu_p(\beta)$ & $\nu_p(\Delta)$ & Singularity in $\overline{S}$ & $e$ & Description \\ \noalign{\smallskip}\hline\hline\noalign{\smallskip} $I_1$ & $0$ & $0$ & $1$ & Smooth & $1$ & Nodal rational curve\\ \hline\noalign{\smallskip} $I_n$ & $0$ & $0$ & $n$ & $A_{n-1}$ & $n$ & Cycle of $n$ smooth rational curves with dual graph $\widetilde{A}_{n-1}$ \\ \hline\noalign{\smallskip} $I_0^*$ & $2$ & $3$ & $6$ & $D_4$ & $6$ & Configuration of $5$ smooth\\ & $\geq 3$ & $3$ & $6$ & & & rational curves with dual \\ & $2$ & $\geq 4$ & $6$ & & & graph $\widetilde{D}_4$ \\ \hline\noalign{\smallskip} $I_n^*$ & $2$ & $3$ & $n+6$ & $D_{4+n}$ & $n+6$ & Configuration of $n+5$ smooth rational curves with dual graph $\widetilde{D}_{n+4}$ \\ \hline\noalign{\smallskip} $II$ &$\geq 1$ & $1$ & $2$ & Smooth & $2$ & Cuspidal rational curve \\ \hline\noalign{\smallskip} $III$ & $1$ & $\geq 2$ & $3$ & $A_1$ & $3$ &Two rational curves tangent at a point \\ \hline\noalign{\smallskip} $IV$ & $\geq 2$ & $2$ & $4$ & $A_2$ & $4$ & Three smooth rational curves meeting at a single point \\ \hline\noalign{\smallskip} $IV^*$ &$\geq 3$ & $4$ & $8$ & $E_6$ & $8$ & Configuration of $7$ smooth rational curves with dual graph $\widetilde{E}_6$ \\ \hline\noalign{\smallskip} $III^*$ &$3$ & $\geq 5$ & $9$ & $E_7$ & $9$ & Configuration of $8$ smooth rational curves with dual graph $\widetilde{E}_7$ \\ \hline\noalign{\smallskip} $II^*$ &$\geq 4$ & $5$ & $10$ & $E_8$ & $10$ & Configuration of $9$ smooth rational curves with dual graph $\widetilde{E}_8$ \\ \hline\noalign{\smallskip} \end{tabular} } \end{table} \begin{remark}If the values of $\nu_p(\alpha),\nu_p(\beta)$ and $\nu_p(\Delta)$ do not fall into one of the classes described in Table , then the singularities of $\overline{S}$ are worse than ADE singularities and $\overline{S}$ is not the Weierstrass model of a smooth relatively minimal elliptically fibred surface.\end{remark} \subsection{Mordell-Weil Group} Theorem enables us to compute the classes in the N\'{e}ron-Severi lattice that arise from components of fibres of the elliptic fibration. However, to obtain the full N\'{e}ron-Severi lattice, we also have to know about the classes coming from sections. This data is encoded by a second object, the \emph{Mordell-Weil group}. As usual, we let $(S,\pi)$ an elliptically fibred surface. Choose an arbitrary section $\mathbb{O}$ and, as before, let its image be denoted $O$. Let $\eta$ be the generic point on the base $B$. Any section $Q$ determines a $\mathbb{C}(B)$-rational point $\widetilde{Q}$ on the generic fibre $S_\eta$ and, in fact, there is a bijective correspondence between sections of $(S,\pi)$ and $\mathbb{C}(B)$-rational points of $S_\eta$. To see this, note that every $\mathbb{C}(B)$-rational point $\widetilde{Q}$ is actually a $\mathbb{C}[B_0]$ point of $S|_{B_0}$ over some open set $B_0 \subseteq B$; this bijection associates to $\tilde{Q}$ the closure in $S$ of the corresponding point in $\mathbb{C}[B_0]$. The group structure on $S_\eta$ allows us to add two sections ${Q}_1$ and ${Q}_2$, by letting $Q_1 + Q_2$ be the closure of the $\mathbb{C}(B)$-rational point $\widetilde{Q}_1 \oplus \widetilde{Q}_2$ (where $\oplus$ is used to indicate addition in the group structure on $S_\eta$, defined with respect to the zero section $\widetilde{O}$). \begin{definition} Let $(S,\pi)$ be an elliptically fibred surface with a chosen zero section $O$. Then the set of sections of $(S,\pi)$, equipped with the group structure defined above, is called the \emph{Mordell-Weil group} of $(S,\pi)$ and denoted $\MW(S,\pi)$. \end{definition} By this prescription, the group structure on $\MW(S,\pi)$ corresponds directly to the pointwise addition in each smooth fibre. Note that, as usual, the structure of $\MW(S,\pi)$ does not depend upon the choice of section $O$, hence our notation for the Mordell-Weil group does not make reference to it. \subsection{The N\'{e}ron-Severi Lattice} Now we determine the relationship between the N\'eron-Severi lattice of a smooth relatively minimal elliptically fibred surface $(S,\pi)$, its Mordell-Weil group, and its singular fibres. Let $O$ be a designated section in $\MW(S,\pi)$, which we will use as the identity element. To each fibre $\pi^{-1}(p)$ of $(S,\pi)$ we may associate a root lattice $R_p$ as follows: write $\pi^{-1}(p)$ as a sum of irreducible components $C_0,\ldots,C_n$ as before, labelled so that $C_0$ is the component which intersects the zero section $O$, then let $R_p$ be the sublattice of $\NS(S)$ generated by the classes of $C_1, \dots , C_n$. Note that if $\pi^{-1}(p)$ is irreducible, then $R_p$ is trivial. Next define a sublattice $L_\pi$ of $\NS(S)$ to be the sublattice spanned by the classes $[E], [O]$ and the lattices $R_p$ for all points $p \in B$. \begin{theorem}\textup{\cite[Thm. VII.2.1]{btes}} There is an exact sequence of abelian groups, \[0 \longrightarrow L_\pi \stackrel{a}{\longrightarrow} \NS(S) \stackrel{b}{\longrightarrow} \MW(S,\pi) \longrightarrow 0,\] where $a$ is the obvious embedding and $b$ is the composition of the restriction to the generic fibre $\NS(S) \to \Pic(S_{\eta})$ with the homomorphism $\Pic(S_{\eta}) \to \MW(S,\pi)$. In particular, $b$ assigns to a section $Q$ of $(S,\pi)$ the associated class in $\MW(S,\pi)$. \end{theorem} Note that the homomorphisms $a$ and $b$ depend upon the choice of section $O$, since the definition of $L_\pi$ depends upon the choice of $O$. Therefore, to determine the N\'eron-Severi lattice of any elliptic fibration, it is enough to know both the structure of the singular fibres of $(S,\pi)$ and the Mordell-Weil group. Note that Thm. implies that $\MW(S,\pi)$ is necessarily finitely generated. As we have seen, the singular fibres of an elliptic surface are quite easy to determine, given an explicit Weierstrass equation for $(S,\pi)$. In general, however, it is quite difficult to compute the non-torsion part of the Mordell-Weil group (see \cite[Chap. VII]{btes}), but we can often obtain bounds on the size of $\MW(S,\pi)_{\mathrm{tors}}$, the torsion part of $\MW(S,\pi)$. \begin{corollary} The group $\MW(S,\pi)_{\mathrm{tors}}$ is isomorphic to $ (L_\pi\otimes \mathbb{Q} \cap \NS(S)) / L_\pi$ and hence is a subgroup of the discriminant group of $L_\pi$. \end{corollary} \begin{proof} The first statement follows directly from Thm. . Note that $ (L_\pi \otimes \mathbb{Q} \cap \NS(S)) \subseteq L_\pi^*$, hence the second claim follows. \end{proof} \begin{corollary} Let $S$ be a K3 surface and let $\pi$ be an elliptic fibration on $S$. Let $N_\pi$ be the sublattice of $\NS(S)$ generated by $L_\pi$ and $\MW(S,\pi)_{\text{tors}}$. Then $S$ is $N_\pi$-polarized. \end{corollary} \begin{proof} This follows directly from Thm. . Since $\mathbb{Z}[O] \oplus \mathbb{Z}[E] \cong H$ is contained in $L_\pi$, the discussion in Sect. shows that $N_\pi$ contains a pseudo-ample class, so condition 2 of Def. is satisfied. To see that condition 1 is also satisfied, we need to show that $N_\pi$ is a primitive sublattice of $\NS(S)$. But $\NS(S)/L_\pi \cong \MW(S,\pi)$ and hence, by construction, $\NS(S)/N_\pi \cong \MW(S,\pi)_{\mathrm{free}}$, where $\MW(S,\pi)_{\mathrm{free}}$ denotes the torsion-free part of $\MW(S,\pi)$. Thus $N_\pi$ is primitively embedded in $\NS(S)$. \end{proof} \subsection{Examples} Now we will bring all of this theory together to examine some examples of K3 surfaces which are elliptically fibred, explaining how to use the techniques in the previous sections to compute lattice polarizations on them. \begin{example}[Polarization by a lattice of rank 18] Let us take the elliptically fibred K3 surface $(S,\pi)$ obtained as the minimal resolution of the Weierstrass fibration $(\overline{S},\pi)$ given by \[Y^2Z = X^3 + s^4t^4 XZ^2 + s^5t^5(as^2 + bst + ct^2)Z^3,\] where $a,b,c$ are parameters chosen generically in $\mathbb{C}$. According to Thm. , $(S,\pi)$ has two singular fibres of type $II^*$ located at $[s:t] = [0:1]$ and $[1:0]$. One may check that for a general choice of $a,b,c$, the discriminant $\Delta$ vanishes simply at four other points in $\mathbb{P}^1$, giving four further singular fibres of type $I_1$. Using Thm. and Table , we see that the Euler characteristic of this surface is $2(10) + 4(1) = 24$, as expected. For a singular fibre of type $II^*$, the root lattice $R_p$ is isomorphic to $(-E_8)$. Therefore, the lattice $L_\pi$ for this K3 surface is isomorphic to $H \oplus (-E_8) \oplus (-E_8)$ (recall that singular fibres of type $I_1$ are irreducible, so do not contribute to $L_{\pi}$). Since $L_\pi$ is unimodular, Cor. shows that $\MW(S,\pi)_{\mathrm{tors}}$ is trivial, so Cor. shows that $S$ is in fact $H \oplus (-E_8) \oplus (-E_8)$ polarized. This example is explored in great detail in the papers and . \end{example} \begin{example}[Polarization by a lattice of rank 19] Now take the elliptically fibred K3 surface $(S,\pi)$ obtained as the minimal resolution of the Weierstrass fibration $(\overline{S},\pi)$ given by \begin{eqnarray*}Y^2 & = &\ X^3 + \frac{1}{3}t^{3}s^3 (48 lt^{2} + 96 l ts - t + 48 l s^2 )XZ^2- \\ & & - \frac{2}{27}t^{5} s^5 (72 t^{2} l + 144 lts - ts + 72ls^2 )Z^3,\end{eqnarray*} where $l$ is a generic parameter in $\mathbb{C}$. Its discriminant is given by \[\Delta(s,t) = 256 l^{2} (t + 1)^{4} s^9 t^{9} (64 l t^{2} + 128 lts - ts + 64 ls^2).\] By Thm. , for generic $l$, this K3 surface has two singular fibres of type $III^*$ occurring at $[s:t] = [0:1]$ and $[1:0]$, one singular fibre of type $I_4$ at $[1:-1]$, and two $I_1$'s occurring at the zeros of $64 l t^{2} + 128 lts - ts + 64 ls^2$. We thus find that $L_\pi = H \oplus (-E_7) \oplus (-E_7) \oplus (-A_3)$. By Ex. in the appendix, the lattice $L_\pi$ has a unique overlattice of index $2$ isomorphic to $H \oplus (-E_8) \oplus (-E_8) \oplus \langle -4\rangle$. Therefore, by Cor. , the only possible torsion in $\MW(S,\pi)$ is of order two and, if $\MW(S,\pi)_{\text{tors}}$ has order two, then Cor. shows that $S$ is $H \oplus (-E_8) \oplus (-E_8) \oplus \langle -4\rangle$-polarized. In fact, one can check that $X = t^2s^2/3$ is a solution to the right hand side of the Weierstrass equation for $(\overline{S},\pi)$. Thus \[[s:t] \mapsto [X:Y:Z] = \left[\frac{t^2s^2}{3} : 0 : 1 \right]\] is a section of $\pi$. Since $Y=0$, we see easily that this is actually an order $2$ torsion section of $\pi$ on each fibre, hence it has order two in $\MW(S,\pi)$. This shows that $S$ is $H \oplus (-E_8) \oplus (-E_8) \oplus \langle -4\rangle$-polarized. \end{example} \section{Ample and K\"ahler Cones} In the final section of these notes we will discuss the ample and K\"ahler cones of a K3 surface. These are important objects: among other things, their geometry controls fibration structures and automorphisms on the K3 surface. Furthermore, we will find that the description of these cones essentially reduces to lattice theory. This should not come as a surprise: after all, we have already seen that the Torelli Theorems reduce the theory of moduli of lattice polarized K3 surfaces to essentially lattice theoretic considerations, so it does not seem unreasonable to expect that the birational geometry of K3 surfaces might also be lattice theoretic in nature. \subsection{The Ample Cone} Begin by letting $S$ be any smooth projective complex surface. Then we have: \begin{definition} The \emph{ample cone} of $S$ is the set $\mathrm{Amp}(S) \subset \NS(S) \otimes \R$ consisting of finite sums $\sum a_iu_i$, with $u_i \in \NS(S)$ ample and $a_i \in \R_{>0}$. \end{definition} By the Nakai-Moishezon ampleness criterion, a class $u$ in $\NS(S) \otimes \mathbb{R}$ is in $\Amp(S)$ if and only if it satisfies $\langle u,u\rangle > 0$ and $\langle u,[C] \rangle >0$ for every irreducible curve $C$ on $S$. In the case where $S$ is a K3 surface, we will see that it suffices to check that $\langle u, [C] \rangle > 0 $ for every smooth rational curve $C$ on $S$. To state this formally, we introduce some notation. The set \[\NS(S)^+ = \{ u \in \NS(S) \otimes \mathbb{R} \mid \langle u,u\rangle > 0 \}\] consists of two disjoint connected cones. All of the ample classes in $\NS(S)$ belong to one of them, the \emph{positive cone}, which we denote by $\calC_S$. The ample cone may then be described by the following theorem, which is an easy consequence of the Nakai-Moishezon criterion. \begin{theorem} Let $S$ be a projective K3 surface and let $\Delta^+(S)$ be the set of classes in $\NS(S)$ which are represented by smooth rational curves on $S$. Then the ample cone of $S$ is given by the intersection between the positive cone $\calC_S$ and the set \[\{u \in \NS(S) \otimes \mathbb{R} \mid \langle u,\delta \rangle > 0\ \mathrm{for}\ \mathrm{all}\ \delta \in \Delta^+(S) \}.\] \end{theorem} \begin{proof} By the Nakai-Moishezon criterion, $u \in \Amp(S)$ if and only if $\langle u, u \rangle > 0$ and $\langle u,[C] \rangle > 0$ for every irreducible curve $C$ on $S$. Suppose first that $C$ satisfies $\langle [C],[C] \rangle \geq 0$. Then $\langle h, [C] \rangle > 0$ for any ample class $h$ in $\NS(S)$ so, by \cite[Cor. IV.7.2]{bpv}, we must have $[C] \in \overline{\calC}_S$ (the closure of $\calC_S$). Applying \cite[Cor. IV.7.2]{bpv} again, we find that $\langle u, u \rangle > 0$ and $\langle u,[C] \rangle > 0$ if and only if $u \in \calC_S$. Thus $u \in \Amp(S)$ if and only if $u \in \calC_S$ and $\langle u,[C] \rangle > 0$ for every irreducible curve $C$ on $S$ with $\langle [C],[C] \rangle <0$. But, by the genus formula \cite[Ex. V.1.3]{hart}, the irreducible curves $C$ on $S$ with $\langle [C],[C] \rangle <0$ are precisely the smooth rational curves on $S$. \end{proof} Next, we define three special subgroups of the orthogonal group $\mathrm{O}(\NS(S))$. To define the first, note that any isometry in $\mathrm{O}(\NS(S))$ must either preserve or exchange the two components of $\NS(S)^+$. Let $\mathrm{O}^+(\NS(S))$ denote the subgroup of isometries that preserve them. To define the second, suppose that $\delta$ is any element of $\NS(S)$ with $\langle \delta,\delta \rangle = -2$. We can define an isometry of $\NS(S)$ by $u \mapsto u + \langle u, \delta\rangle \delta$. Such an isometry is called a \emph{Picard-Lefschetz reflection}. The subgroup of $\mathrm{O}(\NS(S))$ generated by all Picard-Lefschetz reflections is called the \emph{Weyl group} of the lattice $\NS(S)$ and is denoted $W_S$. It is easy to see that Picard-Lefschetz reflections preserve the positive cone $\calC_S$, so $W_S$ is a subgroup of $\mathrm{O}^+(\NS(S))$. In fact more is true. One may show (see \cite[Sect. 4.2]{iqfaag}) that $W_S$ is a normal subgroup of $\mathrm{O}^+(\NS(S))$ and that there is a third group $G_S$ giving a semidirect product decomposition \[\mathrm{O}^+(\NS(S)) \cong W_S \rtimes G_S.\] Furthermore, the discussion in \cite[Sect. 4.2]{iqfaag} shows that the closure of the cone $\Amp(S)$ is a fundamental domain for the action of $W_S$ on the positive cone $\calC_S$, so $G_S$ should be thought of as the group of symmetries of the ample cone of $S$. \begin{example}[Elliptically fibred K3 surface] Let $S$ be a K3 surface with $\NS(S) \cong H$ (recall that, by Sect. , this implies that $S$ is projective and admits an elliptic fibration). We will take a basis of $\NS(S)$ given by $[E]$ and $[O]$, the classes of a fibre and the section of the elliptic fibration $S \rightarrow \mathbb{P}^1$ respectively. We may calculate easily that there are only two classes $\delta$ in $H$ with $\langle \delta, \delta \rangle =-2$, given by $\delta = \pm [O]$. Clearly $-[O]$ is not the class of an effective divisor, so the only smooth rational curve on $S$ is $O$ itself. The set $\NS(S)^+$ is given by \[\langle a[E] + b[O],a[E]+b[O]\rangle = 2ab - 2b^2 > 0\] and the class $3[E] + [O]$ is ample (by Nakai-Moishezon), so the positive cone $\calC_S$ is determined by the conditions $b > 0$ and $a > b$. To find $\Amp(S)$, we also require the condition that \[ \langle a[E]+b[O], [O] \rangle = a-2b > 0.\] Therefore, we conclude that $\Amp(S)$ is given by the open cone in $\mathbb{R}^2$ defined by $b > 0$ and $a > 2b$. The ample and positive cones for this example are displayed in Fig. ; in this figure the positive cone $\calC_S$ is the entire shaded area, whilst the darker shaded subset of $\calC_S$ is the ample cone $\Amp(S)$. Note that the closure of the cone $\Amp(S)$ is rational polyhedral, hence $G_S$ must be finite. We will see in Sect. that this means $S$ has a finite group of automorphisms. \end{example} \begin{example}[Nodal quartics] Let $S$ be the minimal resolution of a nodal quartic in $\mathbb{P}^3$ as discussed in Ex. . In that example we found that $S$ is lattice polarized by the lattice $L$ with Gram matrix \[\left(\begin{matrix} -2 & 0 \\ 0 & 4 \end{matrix} \right);\] suppose now that $\NS(S) \cong L$. As in Ex. , let $H$ denote a divisor on $S$ with $H^2 = 4$ induced by a general hyperplane section in $\Proj^3$ and let $C$ denote the exceptional $(-2)$-curve arising from the blow-up of the node. Then $\NS(S)$ is generated by the classes $[H]$ and $[C]$ by assumption, so the set $\NS(S)^+$ is given by the condition \[ \langle a[C] + b[H], a[C] + b[H] \rangle = -2a^2 + 4b^2 > 0. \] Since the class $[H] - [C]$ is ample, the positive cone $\calC_S$ is given by the inequalities $b > 0$ and $-\sqrt{2}b < a < \sqrt{2}b$. In order to compute the ample cone, we have to find the smooth rational curves on $S$. The divisor $C$ is one such curve. Another may be computed as follows. Note that projection away from the node in $\Proj^3$ induces a double covering map $S \to \Proj^2$, which maps $C$ isomorphically to a conic in $\Proj^2$. The involution exchanging the sheets of this cover maps $C$ to another smooth rational curve $C'$. As the pull-back of a line in $\Proj^2$ under this double cover is a divisor in the linear system $|H-C|$ on $S$, it is not difficult to show that $C'$ lies in the linear system $|2H-3C|$. We claim that $C$ and $C'$ determine the ample cone of $S$. Consider the class \begin{align*}u_{\epsilon} &:= (1 + \epsilon)\left([H]-[C]\right) + \frac{1}{2}\left\langle [H]-[C],[C]\right\rangle [C] \\ &= (1 + \epsilon) [H] -\epsilon [C],\end{align*} where $\epsilon \geq 0$ is a real number. Now $u_{\epsilon}$ satisfies $\langle u_{\epsilon} ,[C] \rangle = 2\epsilon$ and, for any other class $\delta \in \Delta^+(S)$, we have $\langle u_{\epsilon} ,\delta \rangle \geq (1+\epsilon) \langle [H]-[C],\delta \rangle > 0$, since $[H]-[C]$ is ample and $\langle[C],\delta\rangle \geq 0$. So, by Thm. , $u_{\epsilon} \in \Amp(S)$ for all $\epsilon > 0$, but $u_{0} = [H] \notin \Amp(S)$ as $\langle [H] ,[C] \rangle = 0$. Thus the class $[H]$ lies in the boundary of the ample cone. By a similar argument, one can show that the class \begin{align*}u_0' &:= \left([H]-[C]\right) + \frac{1}{2}\left\langle [H]-[C],[C']\right\rangle [C']\\ & = 3[H] - 4[C] \end{align*} also lies in the boundary of $\Amp(S)$. But $\rank \NS(S) = 2$ in this example, so the ample cone is $2$-dimensional and thus has only two boundary rays. The ample and positive cones for this example are shown in Fig. ; as before, in this figure the positive cone $\calC_S$ is the entire shaded area, whilst the darker shaded subset of $\calC_S$ is the ample cone $\Amp(S)$. As in the previous example, the closure of the cone $\Amp(S)$ is rational polyhedral, so $G_S$ is finite and $S$ has a finite group of automorphisms. However, the structure of the Weyl group $W_S$ in this case is more interesting. $W_S$ is an infinite group, generated by the Picard-Lefschetz reflections corresponding to the classes $[C]$ and $[C']$. Repeated iteration of these reflections takes the boundary rays $[H]$ and $-4[C]+3[H]$ of the ample cone to the rays $2a[C] + b[H]$, where $a,b \in \mathbb{Z}$, $b > 0$ are solutions to the Pell equation $b^2 - 2a^2 = 1$. As the number of iterations tends to infinity, $\frac{b}{a}$ tends to $\pm\sqrt{2}$, so the limiting rays are $\pm\sqrt{2}[C] + [H]$ as expected. \end{example} \begin{remark} It can happen that the cone $\Amp(S)$ is not rational polyhedral. However, it can also be shown (see \cite[Lemma 2.4]{frak3s}) that $\Amp(S)$ always has a rational polyhedral fundamental domain under the action of $G_S$. Therefore if $\Amp(S)$ is not rational polyhedral, then $G_S$ must be infinite and results in Sect. can be used to show that the automorphism group of $S$ must also be infinite. This is discussed further in the section on the cone conjecture in Huybrechts' lecture notes \cite[Sect. 8.4]{lok3s}. \end{remark} \subsection{Genus One Fibrations and the Ample Cone} One application of this material is to detect genus one fibrations on a projective K3 surface $S$. Note here that we do not assume the existence of a section, so we cannot use the results from Sect. about the existence of an $H$-polarization. We begin by noting that, if $S$ has a genus one fibration $\pi\colon S \rightarrow B$, then the class $[E]$ of a fibre corresponds to a class in $\NS(S)$ which is in the boundary of the closure $\overline{\Amp(S)}$. To see this, first, note that $\langle [E],[E]\rangle = 0$. Furthermore, if $[C]$ is the class of an irreducible curve on $S$, then either $C$ is a curve in a fibre of $\pi$, in which case $\langle [E],[C]\rangle =0$, or $\pi|_C$ is a surjective map of curves, in which case $\langle [E], [C]\rangle = \deg \pi|_C > 0$. Thus by the Nakai-Moishezon criterion, $[E]$ lies in $\overline{\Amp(S)}$ but not in $\Amp(S)$. In fact, the converse of this statement is also true: \begin{theorem}\textup{\cite[Thm. 3.1]{ttastk3}} Let $S$ be a projective K3 surface and let $D$ be a class in the closure of $\Amp(S)$ that has $\langle D,D\rangle = 0$. Then there exists an $n$ such that $nD$ is the class of a fibre in a genus one fibration on $S$. \end{theorem} Now let $D$ be \emph{any} nonzero class in $\NS(S)$ with $\left<D,D \right> = 0$. Then, possibly after negating, we may assume that $D$ is in the closure of the positive cone $\calC_S$. Since $\Amp(S)$ is a fundamental domain for the action of $W_S$ on $\calC_S$, there must be some $\gamma \in W_S$ such that $\gamma(D)$ is in $\overline{\Amp(S)}$. Thus $\gamma(D)$ is the class of a fibre in a genus one fibration. Thus we have: \begin{corollary}\textup{\cite[Cor. 3.3]{ttastk3}} A projective K3 surface $S$ admits a genus one fibration if and only if the lattice $\NS(S)$ admits a nonzero element $u$ with $\langle u, u \rangle = 0$.\end{corollary} \begin{example}[Anticanonical hypersurfaces in $\mathbb{P}^1 \times \mathbb{P}^2$] Look at the anticanonical K3 surfaces in $\mathbb{P}^1 \times \mathbb{P}^2$ given in Ex. . Generically, such K3 surfaces have N\'eron-Severi lattice \[\left( \begin{matrix} 0 & 3 \\ 3 & 2 \end{matrix} \right).\] Hence $\NS(S)$ admits an element of square 0 and thus $S$ admits a genus one fibration. One may exhibit this fibration by restricting the natural projection $\mathbb{P}^1 \times \mathbb{P}^2 \to \mathbb{P}^1$ to the surface $S$. Note that $H$ is not a sublattice of $\NS(S)$, so this \emph{cannot} be an elliptic fibration. \end{example} Since every indefinite lattice of rank $n \geq 5$ contains an element of square $0$ \cite[Cor. IV.3.2]{aca} we have: \begin{corollary} Let $S$ be a projective K3 surface with $\rank \NS(S) \geq 5$. Then $S$ admits a genus one fibration. \end{corollary} \subsection{Automorphisms of K3 surfaces} A second application of this material is to the study of the automorphisms of a projective K3 surface $S$. We can produce such automorphisms using the Strong Torelli Theorem (Thm. ) along with some lattice theory. Begin by letting $f_0$ be an automorphism of $\NS(S)$. Then, by \cite[Cor. 1.5.2]{isbfa}, the automorphism \[f_0 \oplus \Id\colon \NS(S) \oplus \T(S) \rightarrow \NS(S) \oplus \T(S)\] extends uniquely to an automorphism $f$ of the lattice $\Lambda_{\mathrm{K3}}$ if and only if $f_0$ acts trivially on the discriminant lattice $A_{\NS(S)}$ of $\NS(S)$ (recall here that $\T(S)$ denotes the transcendental lattice of $S$, see Sect. ). Assume that this is the case. By definition, the holomorphic $2$-form $\sigma \in H^{2,0}(S)$ sits inside $\T(S)$, so $f$ fixes the period point of $S$. Therefore $f$ is induced by a non-trivial automorphism of $H^2(X,\mathbb{Z})$ which preserves the period $\sigma$. If we further assume that $f_0$ is contained in the group $G_S$ of automorphisms of $\NS(S)$ which preserve the ample cone $\Amp(S)$, then $f$ sends some ample class of $S$ to another ample class on $S$. Thus the Strong Torelli Theorem (Thm. ) tells us that $f$ induces a unique isomorphism on $S$. We have: \begin{proposition}\textup{\cite[Sect. 7]{ttastk3}} Let $S$ be a projective K3 surface. The subgroup of $\Aut(S)$ which fixes $\T(S)$ is isomorphic to the finite index subgroup of $G_S$ that acts trivially on the discriminant lattice $A_{\NS(S)}$. \end{proposition} Such automorphisms are called \emph{symplectic automorphisms}, since they preserve the holomorphic symplectic form $\sigma$ of $S$. We will denote the group of such automorphisms by $\Aut(S)^s$. There is an embedding $i\colon \Aut(S)^{s} \hookrightarrow \Aut(S)$. Another important theorem which follows with minimal effort from lattice theory and the Strong Torelli Theorem is: \begin{theorem}\textup{\cite[Sect. 7]{ttastk3}\cite[Cor. 4.2.4]{iqfaag}} The cokernel of the embedding $i$ is finite for all projective K3 surfaces $S$. Thus the group $\Aut(S)$ is finite if and only if $G_S$ is finite. \end{theorem} The subgroup of $\Aut(S)$ which does not fix $\T(S)$ tends to be quite small, and is called the \emph{group of non-symplectic automorphisms of $S$}. There has been much work done towards classification of finite groups of symplectic and non-symplectic automorphisms on K3 surfaces. Nikulin provided a classification of cyclic symplectic automorphisms of K3 surfaces in and Mukai completed the classification of symplectic automorphism groups of K3 surfaces in (see also ). Non-symplectic automorphism groups of K3 surfaces were also classified by Nikulin in . Recently, more work has been done towards explicitly exhibiting and classifying K3 surfaces admitting non-symplectic automorphisms; see, for example, . \begin{remark} Finiteness of the group $G_S$ is equivalent to the index of $W_S$ being finite in $\mathrm{O}(\NS(S))$ (since $\mathrm{O}^+(\NS(S))$ has finite index in $\mathrm{O}(\NS(S))$). Lattices whose Weyl groups are of finite index in their orthogonal groups are called \emph{reflexive}. Nikulin has produced a classification of reflexive hyperbolic lattices, which in turn gives a classification of lattice polarized K3 surfaces with finite automorphism group . \end{remark} \begin{example}[K3 surfaces with infinitely many symplectic automorphisms] Let $L$ be a lattice of rank $2$ with Gram matrix \[\left(\begin{matrix} 2na & nb \\ nb & 2nc \end{matrix} \right),\] where $n > 1$ and $4ac - b^2 < 0$ is not a perfect square. Then $L$ admits no classes $\delta$ of square $(-2)$, so if a K3 surface $S$ has $\NS(S) \cong L$, then $W_S$ is trivial and $\calC_S = \Amp(S)$. Therefore, the finite index subgroup of $\mathrm{O}^+(\NS(S))$ fixing the discriminant $A_{\NS(S)}$ is isomorphic to the group of symplectic automorphisms of $S$. This group is infinite cyclic and closely related to the group of units of a subring of $\mathbb{Q}\left(\sqrt{4ac - b^2}\right)$, we refer the interested reader to for details. \end{example} \subsection{The K\"{a}hler Cone} Now suppose that $S$ is a smooth K\"{a}hler surface (that is not necessarily projective). Then we have: \begin{definition} The \emph{K\"ahler cone} of $S$ is the open convex cone $\K(S)$ of all K\"ahler classes in $H^{1,1}(S,\R)$. \end{definition} The K\"ahler cone can be constructed in a very similar way to the ample cone on a projective surface (in fact, as we shall see, the two are very closely related). Begin by considering the set \[\{u \in H^{1,1}(S,\R) \mid \langle u,u \rangle > 0\},\] which is the analogue of $\NS(S)^+$ from Sect. . This set consists of two disjoint cones. All of the K\"ahler classes belong to one of them, the \emph{positive cone}, which will again be denoted $\calC_S$ (this seems like a confusing choice of terminology but, when $S$ is projective, the positive cone from Sect. is simply the intersection of $\calC_S$ with $\NS(S)$). Then we find: \begin{theorem} \textup{\cite[Cor. VIII.3.9]{bpv}} Let $S$ be a K3 surface and let $\Delta^+(S)$ be the set of classes in $\NS(S)$ which are represented by smooth rational curves on $S$. Then the K\"ahler cone of $S$ is given by the intersection between the positive cone $\calC_S$ and the set \[\{u \in H^{1,1}(S,\R) \mid \langle u,\delta \rangle > 0\ \mathrm{for}\ \mathrm{all}\ \delta \in \Delta^+(S) \}.\] \end{theorem} From this and Thm. , we see that if $S$ is projective then $\Amp(S) = \K(S) \cap \NS(S) \otimes \R$. Given this, most of the results on the ample cone that we saw in Sect. also hold for the K\"ahler cone. In particular, the action of the group $W_S$ (defined as before) extends to all of $H^2(S,\Z)$ and this action preserves the positive cone $\calC_S$. The closure of the K\"ahler cone $K(S)$ is then a fundamental domain for the action of $W_S$ on $\calC_S$ \cite[Prop. VIII.3.10]{bpv}. \begin{remark} It is easy to see that these results on the K\"ahler cone imply the corresponding results about the ample cone, so many references choose to focus on the K\"ahler cone first. However, we find the ample cone to be a conceptually simpler object to study, so we decided to reverse the order (in particular, the K\"ahler cone is always $20$ dimensional, whereas the dimension of $\Amp(S)$ depends upon $\NS(S)$, so we can find examples where $\Amp(S)$ is small enough to write down explicitly). \end{remark} \section{Further Reading} For the interested reader, more detailed information about K3 surfaces and the period map, including proofs of the Torelli Theorems, may be found in Chap. VIII of the book by Barth, Hulek, Peters and van de Ven . For a more in-depth discussion of the construction of moduli spaces of K3 surfaces and their compactifications, we recommend the article by Gritsenko, Hulek and Sankaran . Further information about the moduli space of polarized K3 surfaces and its compactifications (especially the Baily-Borel compactification) may also be found in the book by Scattone . An excellent overview of the theory of degenerations may be found in the survey paper by Friedman and Morrison . The best reference for the theory of lattice polarized K3 surfaces and their moduli is still probably Dolgachev's original paper . For more information on the theory of elliptic surfaces, Miranda's book is an excellent reference. Finally, readers interested in learning more about the ample and K\"{a}hler cones of K3 surfaces can consult the chapter on the ample cone and K\"{a}hler cone in Huybrechts' lecture notes \cite[Chap. 8]{lok3s}. \medskip \noindent {\small \textbf{Acknowledgements.} A part of these notes were written while A. Thompson was in residence at the Fields Institute Thematic Program on Calabi-Yau Varieties: Arithmetic, Geometry and Physics; he would like to thank the Fields Institute for their support and hospitality.} \section*{Appendix: Lattice Theory} \addcontentsline{toc}{section}{Appendix: Lattice Theory} In this appendix we present a short description of the lattice theory that is used in the preceding article. The main reference for this section will be . In this article, we use the word \emph{lattice} in the following sense. \begin{definition} A lattice is a pair $(L,\langle\cdot,\cdot\rangle)$ consisting of a finitely generated free $\mathbb{Z}$-module $L$ and an integral symmetric bilinear form $\langle \cdot,\cdot \rangle$ on $L$. \end{definition} Often we will suppress the bilinear form $\langle \cdot, \cdot \rangle$ and refer to a lattice simply as $L$. A lattice $L$ is called \emph{non-degenerate} if the $\R$-linear extension of the bilinear form $\langle \cdot, \cdot \rangle$ to the $\R$-vector space $L \otimes_{\Z} \R$ is non-degenerate. For the remainder of this appendix, we will assume that all lattices are non-degenerate. A lattice $L$ has \emph{signature} $(m,n)$ if, for some basis $u_1, \dots, u_{m+n}$ of $L \otimes_\mathbb{Z} \mathbb{R}$, we have \[ \langle u_i, u_j \rangle = \left\{ \begin{array}{cl} 1 & \mathrm{if}\ i = j \in\{ 1, \ldots, m\}, \\ -1 & \mathrm{if}\ i = j \in \{m+1,\ldots,m+n\}, \\ 0 & \mathrm{if}\ i \neq j. \end{array} \right.\] If $L$ is of signature $(m,0)$ we call it \emph{positive definite}, and if it has signature $(0,n)$ we say that it is \emph{negative definite}. If a lattice is neither positive nor negative definite, it is called \emph{indefinite}. If a lattice has signature $(m,1)$ we will call it \emph{hyperbolic}. Let $L$ be a lattice and $u_i$ a basis of $L$. Then the \emph{Gram matrix} of $L$ is the matrix of integers $g_{i,j} = \langle u_i,u_j \rangle$ and the \emph{discriminant} of $L$, denoted $\disc (L)$, is the absolute value of the determinant of the Gram matrix. Obviously the Gram matrix depends upon the basis chosen, but the discriminant is independent of basis. A lattice is called \emph{even} if for every $u$ in $L$, \[\langle u,u \rangle \equiv 0 \bmod 2.\] For instance, a root lattice of ADE type is a positive definite even lattice. When dealing with K3 surfaces, all relevant lattices are even. A lattice is called \emph{unimodular} if it has discriminant $1$. Up to isomorphism, there is a single even unimodular rank 2 lattice of signature $(1,1)$, which has Gram matrix for some basis given by \[\left( \begin{matrix} 0 & 1 \\ 1 & 0 \end{matrix} \right).\] This lattice is called the \emph{hyperbolic plane} and, depending on the author, is denoted $U$ or $H$. We will denote it by $H$. Now suppose that $L$ and $M$ are two lattices and that $L$ embeds into $M$. Then $L$ is said to be a \emph{sublattice} of $M$. This embedding is called \emph{primitive} if the quotient $M/L$ is torsion-free. Similarly, an element $u \in M$ is called \emph{primitive} if the sublattice of $M$ generated by $u$ is primitively embedded in $M$. Given a lattice $L$, we may define a second lattice $L^*$, called the \emph{dual lattice} of $L$, as follows. Consider the tensor product $L \otimes_\mathbb{Z} \mathbb{Q}$, with bilinear form induced by the $\mathbb{Q}$-linear extension of $\langle\cdot , \cdot \rangle$. Then define $L^*$ to be the subgroup of $L \otimes_\mathbb{Z} \mathbb{Q}$ made up of elements $v$ which satisfy $\langle v,u \rangle \in \mathbb{Z}$ for all $u \in L$, equipped with the integral binear form induced by $\langle \cdot,\cdot \rangle$. Note that $L$ is a sublattice of $L^*$. For even lattices $L$, we may use this to define a more refined version of the discriminant, called the \emph{discriminant lattice} of $L$. This is given by the finite group \[A_L:=L^*/L.\] This group is equipped with a quadratic form and a bilinear form as follows: take $u,v \in L^*$ and let $\overline{u},\overline{v}$ be their images in $A_L$, then define \[q_L(\overline{u}) = \langle u,u \rangle \bmod 2 \mathbb{Z}\] and \[b_L(\overline{u},\overline{v}) = \langle u,v \rangle \bmod \mathbb{Z}.\] Note that if $u,v \in L$, then the fact that $L$ is an even lattice implies that $q_L(\overline{u}) = 0$ and $b_L(\overline{u},\overline{v}) =0$, so $q_L$ and $b_{L}$ are well-defined. The group $A_L$ is finite and $|A_L| = \disc (L)$. The invariant $A_L$ is obviously finer than just the discriminant of the lattice, but its true strength is made evident by the following proposition of Nikulin. \begin{proposition}\textup{\cite[Cor. 1.13.3]{isbfa}} Let $L$ be an even indefinite lattice of signature $(m,n)$ and rank $m+n$, with discriminant lattice $A_L$. Let $\ell(L)$ denote the minimal number of generators of $A_L$. If $\ell(L) \leq m+n-2$, then any other lattice with the same rank, signature and discriminant lattice is isomorphic to $L$. \end{proposition} \subsection*{Overlattices} \addcontentsline{toc}{subsection}{Overlattices} Now assume that $L$ and $M$ are two even lattices of the same rank, such that $L$ embeds inside of $M$. Then we say that $M$ is an \emph{overlattice} of $L$. If we begin with a lattice $M$, then it is easy to compute all possible sublattices of maximal rank of $L$, but the problem of computing all possible overlattices of $L$ is more subtle. It is solved by the following theorem: \begin{theorem} \textup{\cite[Prop. 1.4.1]{isbfa}} Let $L$ be an even lattice. Then there is a bijection between subgroups $G$ of $A_L$ on which the form $q_L$ satisfies $q_L(u) = 0$ for all $u \in G$ and overlattices $L_G$ of $L$. Furthermore, the discriminant form of the lattice $L_G$ associated to the subgroup $G$ is given by the form $q_L$ restricted to $G^\perp/G$, where orthogonality is measured with respect to $b_L$. \end{theorem} The main practical use of this proposition is to determine when a specific lattice is primitively embedded in another. In particular, if a lattice $L$ has no non-trivial overlattices, then any embedding of $L$ into another lattice $M$ must be primitive. However, it can also be used to explicitly compute the possible overlattices of a given lattice, as illustrated by the next example. \begin{example} Let $L$ be the lattice $H \oplus (-E_7) \oplus (-E_7) \oplus (-A_3)$. Then $A_L$ is isomorphic to $\mathbb{Z}/2 \oplus \mathbb{Z}/2 \oplus \mathbb{Z}/4$ with generators $u,v,w$ respectively. It may be checked explicitly that \[q_L(u) = q_L(v) = \frac{1}{2}\] and \[q_L(w) = \frac{5}{4},\] and that $u,v,w$ are mutually orthogonal with respect to $b_L$. One checks easily that the only nontrivial element $Q$ in $A_L$ with $q_L(Q) = 0$ is $Q = u + v + 2w$, which has order $2$. Thus $L$ has a unique overlattice $L_G$ of index $2$, corresponding to the subgroup $G$ of $A_L$ generated by $Q$. One may construct $L_G$ concretely in the following way: let $\hat{Q}$ be some element of $L^*$ whose image in $A_L$ is $Q$, then $L_G$ can be identified as the sublattice of $L\otimes_\mathbb{Z} \mathbb{Q}$ spanned by $\hat{Q}$ and the image of $L$ in $L\otimes_\mathbb{Z} \mathbb{Q}$. However, it is often simpler to use Prop. to identify the overlattice $L_G$. The subgroup $G$ of $A_L$ generated by $Q$ has orthogonal complement generated by $Q$ and $v + w$. Modulo $G$, this group is cyclic of order four and \[q_L(v+ w) = -\frac{1}{4}.\] Therefore, the overlattice $L_G$ of $L$ associated to $G$ has rank $19$, signature $(1,18)$ and discriminant group of order $4$ with a generator satisfying $q_L(v+w) = -1/4$. Now, the lattice $M = H \oplus (-E_8) \oplus (-E_8) \oplus \langle -4 \rangle$ also has rank $19$, signature $(1,18)$ and discriminant group of order $4$ with generator $e$ satisfying $q_M(e) = -1/4$ so, by Prop. , the overlattice $L_G$ of $L$ must be isomorphic to the lattice $M$. \end{example} \begin{remark} Note that if we replaced the lattice $H \oplus (-E_7) \oplus (-E_7) \oplus (-A_3)$ with the lattice $(-E_7) \oplus (-E_7) \oplus (-A_3)$ then we could not use Prop. here, since the second lattice is not indefinite. \end{remark} \bibliography{books} \bibliographystyle{amsplain} |
1501.04052 | Title: Convexity of Energy-Like Functions: Theoretical Results and Applications
to Power System Operations
Abstract: Power systems are undergoing unprecedented transformations with the
incorporation of larger amounts of renewable energy sources, distributed
generation and demand response. All these changes, while potentially making
power grids more responsive, efficient and resilient, also pose significant
implementation challenges. In particular, operating the new power grid will
require new tools and algorithms capable of predicting if the current state of
the system is operationally safe. In this paper we study and generalize the
so-called energy function as a tool to design algorithms to test if a
high-voltage power transmission system is within the allowed operational
limits. In the past the energy function technique was utilized primarily to
access the power system transient stability. In this manuscript, we take a new
look at energy functions and focus on an aspect that has previously received
little attention: Convexity. We characterize the domain of voltage magnitudes
and phases within which the energy function is convex. We show that the domain
of the energy function convexity is sufficiently large to include most
operationally relevant and practically interesting cases. We show how the
energy function convexity can be used to analyze power flow equations, e.g. to
certify solution uniqueness or non-existence within the domain of convexity.
This and other useful features of the generalized energy function are described
and illustrated on IEEE 14 and 118 bus models.
Body: \begin{abstract} Power systems are undergoing unprecedented transformations with the incorporation of larger amounts of renewable energy sources, distributed generation and demand response. All these changes, while potentially making power grids more responsive, efficient and resilient, also pose significant implementation challenges. In particular, operating the new power grid will require new tools and algorithms capable of predicting if the current state of the system is operationally safe. In this paper we study and generalize the so-called energy function as a tool to design algorithms to test if a high-voltage power transmission system is within the allowed operational limits. In the past the energy function technique was utilized primarily to access the power system transient stability. In this manuscript, we take a new look at energy functions and focus on an aspect that has previously received little attention: \emph{Convexity}. We characterize the domain of voltage magnitudes and phases within which the energy function is convex. We show that the domain of the energy function convexity is sufficiently large to include most operationally relevant and practically interesting cases. We show how the energy function convexity can be used to analyze power flow equations, e.g. to certify solution uniqueness or non-existence within the domain of convexity. This and other useful features of the generalized energy function are described and illustrated on IEEE 14 and 118 bus models. \end{abstract} \section{Appendix} \subsection{Proof of Theorem } \begin{proof} We introduce an independent edge variable $\theta_k$ for each edge $k \in \E$. Let $\theta_{\E}=\{\theta_k:k\in\E\}$. Let $k_1,k_2$ denote the ``from'' and ``to'' ends of the edge $k$. We then have $\theta_{\E}=A\theta$. If the energy function is convex jointly in $\br{\rho,\theta,\theta_\E}$, then it must be jointly convex in $\br{\rho,\theta}$. Writing the energy function in terms of $\br{\rho,\theta,\theta_\E}$, we derive $E\br{\rho,\theta,\theta_{\E}} =-\sum_{i\in\Lo\cup\G}P_i \theta_i -\sum_{i\in\Lo} Q_i\rho_i +\sum_{k} \frac{1}{2}B_{k}\br{\expb{2\rho_{k1}}+\expb{2\rho_{k2}}} -\sum_{k}B_k\expb{\rho_{k1}+\rho_{k2}}\cos\br{\theta_k}$. In this expression only the first term depends on $\theta$. Further, the first two terms depend linearly on $\br{\rho,\theta}$ and are hence convex. Therefore, convexity of the energy function reduces to convexity of the second and third terms. We study the Hessian which can be broken into 4 sub-matrices: \[ \begin{pmatrix} \nabla^2_{\rho}E\br{\rho,\theta_{\E}} & \nabla^2_{\rho,\theta_{\E}}E\br{\rho,\theta_{\E}}\\ \tranb{\nabla^2_{\rho,\theta_{\E}}E\br{\rho,\theta_{\E}}} & \nabla^2_{\theta_{\E}}E\br{\rho,\theta_{\E}} \end{pmatrix}=\begin{pmatrix} M & N \\ \tran{N} & R\end{pmatrix}\] where $M\in\R^{|\Lo|\times |\Lo|},N\in\R^{|\Lo|\times |\E|},R\in\R^{|\E|\times|\E|}$. For the matrix to be positive semi-definite, we require that $R\succeq 0,M-N \inve{R} \tran{N}\succeq 0 $. It is easy to see that $R$ is diagonal, with $R_{kk}=B_{k}\expb{\rho_{k1}+\rho_{k2}}\cos\br{\theta_k}$. Further, \begin{align*} N_{ik}&=|A_{ik}|B_k\expb{\rho_{k_1}+\rho_{k_2}}\sin\br{\theta_k},i\in\Lo,k\in\E\\ M_{ij}&=-\sum_k |A_{ik}||A_{jk}|B_{k}\expb{\rho_{k_1}+\rho_{k_2}}\cos\br{\theta_k},i\neq j\\ M_{ii}&=2B_i\expb{2\rho_i}-\sum_k |A_{ik}|B_{k}\expb{\rho_{k_1}+\rho_{k_2}}\cos\br{\theta_k}. \end{align*} Let $O=M-N\inve{R}\tran{N} \in \R^{|\Lo|\times|\Lo|}$ Then, \begin{align*} O_{ij}& =-\sum_k |A_{ik}||A_{jk}|B_{k}\expb{\rho_{k_1}+\rho_{k_2}}\br{\cos\br{\theta_k}+\frac{\sin^2\br{\theta_k}}{\cos\br{\theta_k}}}\\ &=-\sum_k |A_{ik}||A_{jk}|B_{k}\expb{\rho_{k_1}+\rho_{k_2}}\frac{1}{\cos\br{\theta_k}} \end{align*} for $i\neq j$ and \begin{align*} O_{ii}& =2B_i\expb{2\rho_i}\\ & \quad-\sum_k |A_{ik}|B_{k}\expb{\rho_{k_1}+\rho_{k_2}}\br{\cos\br{\theta_k}+\frac{\sin^2\br{\theta_k}}{\cos\br{\theta_k}}}\\ & = 2B_i\expb{2\rho_i}-\sum_k |A_{ik}|B_{k}\expb{\rho_{k_1}+\rho_{k_2}}\frac{1}{\cos\br{\theta_k}} \end{align*} Note also that \begin{align*} &L_{ij}=\frac{O_{ij}}{\expb{\rho_i}\expb{\rho_j}}=-\sum_k |A_{ik}||A_{jk}|B_{k}\frac{1}{\cos\br{\theta_k}}, i\neq j\\ &L_{ii}=\frac{O_{ii}}{\expb{2\rho_i}}=2B_i-\sum_k |A_{ik}|\expb{\rho_{k_1}+\rho_{k_2}-2\rho_i}B_{k}\frac{1}{\cos\br{\theta_k}} \end{align*} Then $L\succeq 0$ if and only if $O \succeq 0$ since $L=\mathrm{diag}\br{\expb{-\rho}}O\mathrm{diag}\br{\expb{-\rho}}$. Finally, note that $L\succeq 0$ if and only if the following matrix is positive semidefinite: \begin{align*} &\left[2B_i-\sum_{j\in\G,(i,j)\in\E}B_{ij}\expb{\rho_j-\rho_i}\frac{1}{\cos\br{\theta_{ij}}}\right]_{ii}^{\Lo}\\ & \,-\sum_{(i,j)\in\E,i,j\in\Lo}\frac{B_{ij}}{\cos\br{\theta_{ij}}} \brs{\begin{pmatrix}\expb{\rho_j-\rho_i} & 1 \\ 1 & \expb{\rho_i-\rho_j}\end{pmatrix}}_{ij}^{\Lo} \end{align*} To see that this is a convex constraint, we only need to observe that \[\frac{1}{\cos\br{\theta_{ij}}}\brs{\begin{pmatrix}\expb{\rho_j-\rho_i} & 1 \\ 1 & \expb{\rho_i-\rho_j}\end{pmatrix}}_{ij}\] is $\succeq$-convex in $\br{\rho_j-\rho_i,\theta_{ij}}$ (Lemma ). The terms in the diagonal matrix are all concave functions of $\br{\rho,\theta}$ and hence the rerms are $\succeq$-concave. Finally, plugging in $\theta_{ij}=\theta_i-\theta_j$, one observes that the convexity is preserved and hence the energy function is convex over $\C$. Further, at $\br{\rho,\theta}=0$, the condition reduces to: \[M\succeq 0, M_{ii}=2B_i-\sum_{j\sim i} B_{ij}=\sum_{j\sim i} B_{ij}, M_{ij}=-B_{ij}\] Since $M$ is symmetric and diagonally dominant, it must be positive semi-definite. Hence, $0\in\C$. Further, since $\C$ is convex $\br{\alpha\rho,\alpha\theta}\in\C\quad \forall \alpha\in [0,1]$. Therefore, $\C\subset \D$. For the converse, we look at the tree networks. We assume that the tree is connected, else the energy function can be decomposed into a sum of independent functions on each connected component and the following arguments apply. First, note that for a connected tree with $n$ buses, we find $n-1$ edges. Further, note that $\theta_{\Sb}=0$. Let $\tilde{\theta}$ denote the vector of phases at all buses except at the slack bus. Then, there is a one-to-one correspondence between $\tilde{\theta}$ and $\theta_\E$, i.e., there exists an $(n-1)\times (n-1)$ matrix $\tilde{A}$ (submatrix of $A$ formed by deleting the column corresponding to the slack bus) such that $\ \theta_\E=\tilde{A}\tilde{\theta}$. Further, note that $\tilde{A}$ is invertible. Hence, $\theta=\inv{\tilde{A}}\theta_\E$ and the Hessian of $E$ wrt $\br{\rho,\theta}$ is equal to \begin{align*} \begin{pmatrix} \nabla^2_{\rho}E\br{\rho,\theta} & \nabla^2_{\rho,\theta_{\E}=\tilde{A}\theta}E\br{\rho,\theta_{\E}}\inve{\tilde{A}}\\ A^{-T}\tran{{\nabla^2_{\rho,\theta_{\E}}E\br{\rho,\theta_{\E}}}} & {\tilde{A}}^{-T}\nabla^2_{\theta_\E=\tilde{A}\theta}E\br{\rho,\theta,\theta_\E}{\inve{\tilde{A}}} \end{pmatrix} \end{align*} Using Schur-complements and the invertibility of $\tilde{A}$, it is easy to see that the positive semi-definiteness of this matrix is equivalent to that of the matrix in the first part of the proof. The positive semi-definiteness of the bottom right block requires that $|\theta_{\E}|\leq\frac{\pi}{2}$. If this inequality is strict, the block is positive definite and then the analysis in the first part is necessary and sufficient. The cases where this block is singular can be handled by a continuity argument. Hence, in this case, $\C\supseteq\D$, and combining the result from the first part, we arrive at $\D=\C$. \end{proof} \begin{lemma} The matrix-valued function \[f(x,y)=\frac{1}{\cos(y)}\begin{pmatrix}\expb{x} & 1 \\ 1 & \expb{-x}\end{pmatrix}\] is $\succeq$-convex, that is $ f\br{\lambda x_1+(1-\lambda)x_2,\lambda y_1+(1-\lambda)y_2}\preceq \lambda f\br{x_1,y_1}+(1-\lambda)f\br{x_2,y_2}$ \end{lemma} \begin{proof} $f$ is $\succeq$-convex if and only if $\tr{f\br{x,y}R}$ is convex for all $R\succeq 0$ \citep{boyd2009convex}. Let $R=\begin{pmatrix}a & b \\ b & c\end{pmatrix}\succeq 0$ be an arbitrary $2\times 2$ positive semi-definite matrix. Then, we have \begin{align*} &\tr{f\br{x,y}R}=\frac{a\expb{x}+c\expb{-x}+2b}{\cos\br{y}}\\ & = \frac{a\expb{x}+c\expb{-x}-2\sqrt{ac}}{\cos\br{y}}+2\frac{b+\sqrt{ac}}{\cos\br{y}} \end{align*} Since $R\succeq 0$, $|b|\leq\sqrt{ac}$ and hence the second term is convex (it is the inverse of a positive concave function, $\cos$). The first term can be rewritten as $\frac{\powb{|\sqrt{a}\expb{x/2}-\sqrt{c}\expb{-x/2}|}{2}}{\cos\br{y}}$. It is easy to check that $|\sqrt{a}\expb{x/2}-\sqrt{c}\expb{-x/2}|$ is twice differentiable at all values of $x$ and its second derivative is equal to $|\sqrt{a}\expb{x/2}-\sqrt{c}\expb{-x/2}|$. Hence, this function is convex. Therefore, the function $|\sqrt{a}\expb{x/2}-\sqrt{c}\expb{-x/2}|^2/\cos(y)$ is convex by the composition rules (since $x^2/t$ is convex in $x,t$, increasing in $x$ and decreasing in $t$). Thus, $f$ is $\succeq$-convex as long as $|y|\leq \frac{\pi}{2}$. \end{proof} \subsection{Lossy Networks} Here we show that all of our results extend to the special class of lossy networks, where all the transmission lines have the same/fixed value of the reactance-to-resistance ratio: $\frac{X_{ij}}{R_{ij}}=\frac{G_{ij}}{B_{ij}}=\kappa$. Further, we assume that all buses are \PQ~buses, except for the slack bus. In this case, let us restate PF equations through the following linear combination of the PF equations written in their original form \begin{subequations} \begin{align} & P_i+\kappa Q_i=\sum_{j \neq i} \br{\kappa^2+1}\expb{\rho_i+\rho_j}B_{ij}\sin\br{\theta_{ij}} ,i\in\Lo\\ & \kappa P_i- Q_i=\sum_{j \neq i} \br{\kappa^2+1}\expb{\rho_i+\rho_j}B_{ij}\cos\br{\theta_{ij}}\quad \forall i \in \Lo\\ & \theta_{\Sb}=\rho_{\Sb}=0 \end{align} \end{subequations} Then the PF equations \eqref{eq:pfLossy} can be rewritten in a variational form: \begin{subequations} \begin{align} E\br{\rho,\theta} &=\sum_{i\in\Lo} -\br{\br{P_i+\kappa Q_i}\theta_i+\br{\kappa P_i-Q_i}\rho_i}\nonumber\\ &\quad +\sum_{\br{i,j}\in\E} B_{ij}\br{\kappa^2+1}\br{\frac{\expb{2\rho_i}+\expb{2\rho_j}}{2}}\nonumber \\ &\quad -\sum_{\br{i,j}\in\E} B_{ij}\br{\kappa^2+1}\expb{\rho_i+\rho_j}\cos\br{\theta_i-\theta_j} \\ &\eqref{eq:pqaLossy}\equiv \frac{\partial E}{\partial \theta_i}=0 , \eqref{eq:pqbLossy}\equiv \frac{\partial E}{\partial \rho_i}=0 \end{align} \end{subequations} We arrive at the following statement. \begin{theorem} The energy function for the lossy system with constant $\frac{R}{X}$ ratio \eqref{eq:ELossy} is convex over the domain: \begin{align} &|\theta_i-\theta_j|\leq \frac{\pi}{2}\quad\forall (i,j)\in\E\\ &\sum_{i\in\Lo} \left[2B_i-\sum_{j\in \{S\}}B_{ij}\frac{\expb{\rho_{ji}}}{\cos\br{\theta_{ij}}}\right]_{ii}^{\Lo}\nonumber \\ & -\sum_{(i,j)\in\E,i,j\in\Lo}\frac{B_{ij}}{\cos\br{\theta_{ij}}} \brs{\begin{pmatrix}\expb{\rho_{ji}} & 1 \\ 1 & \expb{\rho_{ij}}\end{pmatrix}}_{ij}^{\Lo}\succeq 0 \end{align} \end{theorem} \begin{proof} Almost identical to the proof of theorem . \end{proof} \title{Bare Demo of IEEEtran.cls for Journals} \author{Michael~Shell,~\IEEEmembership{Member,~IEEE,} John~Doe,~\IEEEmembership{Fellow,~OSA,} and~Jane~Doe,~\IEEEmembership{Life~Fellow,~IEEE} \thanks{M. Shell is with the Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, 30332 USA e-mail: (see http://www.michaelshell.org/contact.html).} \thanks{J. Doe and J. Doe are with Anonymous University.} \thanks{Manuscript received April 19, 2005; revised September 17, 2014.}} \markboth{Journal of \LaTeX\ Class Files,~Vol.~13, No.~9, September~2014} {Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for Journals} \maketitle \begin{abstract} The abstract goes here. \end{abstract} \begin{IEEEkeywords} IEEEtran, journal, \LaTeX, paper, template. \end{IEEEkeywords} \IEEEpeerreviewmaketitle \section{Introduction} \IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file'' for IEEE journal papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \section*{Acknowledgment} The authors would like to thank... \ifCLASSOPTIONcaptionsoff \newpage \fi \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to \LaTeX}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \begin{IEEEbiography}{Michael Shell} Biography text here. \end{IEEEbiography} \begin{IEEEbiographynophoto}{John Doe} Biography text here. \end{IEEEbiographynophoto} \begin{IEEEbiographynophoto}{Jane Doe} Biography text here. \end{IEEEbiographynophoto} \section{Conclusions and Path Forward} This manuscript presents novel descriptions for the domain of convexity of the energy function of a lossless power transmission network. Since the energy function provides a variational description of the power flow equations, this allows us to develop an algorithm, based on a convex optimization, solving the PF equations. This enables various applications, including detecting if the PF equations have a solution in a certain domain. There is also a growing family of important power system applications which should benefit from these convexity and PF solvability results. These applications include (a) Optimal Power Flow, (b) State and Topology Estimation, (c) Distance to Insolvability/Failure, (d) Optimal Load Shedding etc. Another direction for future work is to extend these results to lossy networks. A special case, assuming that all lines within the network are of the same grade (the value of $X_{ij}/R_{ij}=\kappa \quad \forall\br{i,j}\in\E$), is discussed in Appendix . In the future work we plan to exploit continuity arguments to establish existence of solutions for networks with $\frac{R}{X}$ values roughly constant. It should also be possible to quantify the degree of deviations that can be tolerated. \section{Acknowledgments} This work has emerged from discussions in July of 2014 at Los Alamos with Scott Backhaus and Ian Hiskens, whom the authors are thankful for guidance, comments, encouragement and criticism. We thank Dan Molzahn and Enrique Mallada for helpful comments on various aspects of this work. We thank Florian Dorfler for pointers to useful references and comments on this manuscript. The work at LANL was carried out under the auspices of the National Nuclear Security Administration of the U.S. Department of Energy at Los Alamos National Laboratory under Contract No. DE-AC52-06NA25396. This material is based upon work partially supported by the National Science Foundation award \# 1128501, EECS Collaborative Research ``Power Grid Spectroscopy" under NMC. \maketitle \input{Abstract.tex} \input{Intro.tex} \input{TechIntro.tex} \input{EnergyFunctionConvexity.tex} \input{Conclusions.tex} \bibliography{Ref} \bibliographystyle{alpha} \input{Appendix.tex} \newcommand{\Sen}[4]{\mathrm{S}_{#2,#1}\br{{#4}}} \newcommand{\Dcal}{\hat{\cal D}} \newcommand{\HJB}{{ HJB}} \newcommand{\DGs}{{ DGs}} \newcommand{\DG}{{ DG}} \newcommand{\LDGs}{{ LDGs}} \newcommand{\LDG}{{ LDG}} \newcommand{\Pname}{passive dynamics} \newcommand{\Ppol}{passive policy} \newcommand{\LMDP}{{ LMDP}} \newcommand{\LMDPs}{{ LMDPs}} \newcommand{\MDPs}{{ MDPs}} \newcommand{\MDP}{{ MDP}} \newcommand{\MGs}{{ MGs}} \newcommand{\MG}{{ MG}} \newcommand{\LMGs}{{ LMGs}} \newcommand{\LMG}{{ LMG}} \newcommand{\LD}{{ LD}} \newcommand{\LDs}{{ LDs}} \newcommand{\Pol}{Pol} \newcommand{\Dyn}{Dyn} \newcommand{\OptPol}{OP} \newcommand{\BE}{{ BE}} \newcommand{\BEs}{{ BEs}} \newcommand{\Cost}{Co} \newcommand{\FH}{{\rm FH}} \newcommand{\FE}{{\rm FE}} \newcommand{\IH}{{\rm IH}} \newcommand{\PIC}{{\bf PIC}} \newcommand{\LGs}{{\bf $\alpha$-RLCs}} \newcommand{\LG}{{\bf $\alpha$-RLC}} \newcommand{\Ren}{Renyi~} \newcommand{\Pipasname}{Null Policy} \newcommand{\CCDa}[1]{~\mathrm{CCS}_1\left({#1}\right)} \newcommand{\CCDb}[1]{~\mathrm{CCS}_2\left({#1}\right)} \newcommand{\Sop}{\mathrm{SUBOPT}} \newcommand{\CON}{\rm{CON}} \newcommand{\NCON}{\rm{NCON}} \newcommand{\Aut}{\rm{AUT}} \newcommand{\OPT}{\rm{OPT}} \newcommand{\D}{\mathcal{D}} \newcommand{\la}[1]{\lambda\left({#1}\right)} \newcommand{\lai}[2]{\lambda_{#2}\left({#1}\right)} \newcommand{\lam}[1]{\lambda_{\min}\left({#1}\right)} \newcommand{\lama}[1]{\lambda_{\max}\left({#1}\right)} \newcommand{\sva}[1]{\sigma\left({#1}\right)} \newcommand{\svai}[2]{\sigma_{#2}\left({#1}\right)} \newcommand{\svam}[1]{\sigma_{\min}\left({#1}\right)} \newcommand{\svama}[1]{\sigma_{\max}\left({#1}\right)} \newcommand{\ind}[1]{\mathsf{I}\left[{#1}\right]} \newcommand{\brs}[1]{\left[{#1}\right]} \newcommand{\opt}[1]{{#1}^\ast} \newcommand{\optt}[1]{{{#1}_c}^\ast} \DeclareMathOperator*{\subsetm}{\subset^m} \DeclareMathOperator{\sign}{\mathrm{sign}} \newcommand{\sgn}[1]{\mathrm{sgn}\br{{#1}}} \DeclareMathOperator{\LogSumExp}{\Psi} \DeclareMathOperator*{\mini}{\text{Minimize}} \DeclareMathOperator*{\maxi}{\text{Maximize}} \DeclareMathOperator*{\argmin}{{ \sf{argmin}}} \DeclareMathOperator*{\argmax}{{\sf{argmax}}} \DeclareMathOperator{\ntimes}{\otimes} \newcommand{\logp}[2]{\LogSumExp_{#1}\left[{#2}\right]} \newcommand{\logpa}[3]{\LogSumExp^{#3}_{#1}\left[{#2}\right]} \newcommand{\fpert}[3]{{\bar{#1}}\br{#3}} \newcommand{\fpertr}[3]{{\tilde{#1}}_{#2}\br{{#3}}} \newcommand{\logpaw}[4]{{#4}_{{#3}}\br{{#2}}} \newcommand{\expb}[1]{\exp\left({#1}\right)} \newcommand{\logb}[1]{\log\left({#1}\right)} \newcommand{\norm}[1]{\left\lVert {#1} \right\rVert} \newcommand{\Reg}[1]{R\left({#1}\right)} \newcommand{\ord}[2]{{#1}_{\left[{#2}\right]}} \newcommand{\abs}[1]{\left| {#1} \right|} \newcommand{\powb}[2]{{\left({#1}\right)}^{#2}} \newcommand{\pbar}{\bar{p}} \newcommand{\One}{\mathbf{1}} \newcommand{\tran}[1]{{#1}^T} \newcommand{\Set}[1]{\left\{{#1}\right\}} \newcommand{\inner}[2]{\left\langle {#1},{#2} \right\rangle} \newcommand{\Rad}[1]{\mathrm{R}\br{#1}} \DeclareMathOperator{\der}{\mathrm{d}} \DeclareMathOperator{\pder}{\partial} \newcommand{\diff}[1]{\der{#1}} \newcommand{\pdiff}[1]{\pder{#1}} \newcommand{\DiffGen}[1]{\DiffG\left[{#1}\right]} \newcommand{\Diff}[2]{\frac{\partial {#1}}{\partial {#2}}} \newcommand{\DDiff}[2]{\frac{\partial^2 {#1}}{\partial {#2} \tran{\partial {#2}} }} \newcommand{\Rnn}{\R^{n \times n}} \newcommand{\Snn}{\mathcal{S}^{n \times n}} \newcommand{\Pspace}[1]{\mathcal{P}\left[{#1}\right]} \newcommand{\Poss}[1]{{{#1}}^{\R^+}} \newcommand{\Fspace}[1]{\mathcal{F}\left[{#1}\right]} \newcommand{\meso}{\mu_0} \newcommand{\mes}{\mu} \newcommand{\mesp}{\nu} \DeclareMathOperator*{\Es}{\mathrm{E}} \DeclareMathOperator*{\Covs}{\mathrm{Cov}} \DeclareMathOperator*{\Vars}{\mathrm{Var}} \DeclareMathOperator{\vari}{\mathbb{\sigma}} \DeclareMathOperator{\KLs}{\mathrm{KL}} \newcommand{\N}{\mathcal{N}} \newcommand{\reg}{\mu} \newcommand{\Vari}[1]{\mathrm{Var}\left({#1}\right)} \newcommand{\Covb}[1]{\mathrm{Cov}\left({#1}\right)} \newcommand{\Var}[2]{\Vars_{{#1}}\left({#2}\right)} \newcommand{\Gauss}[2]{\N\left({#1},{#2}\right)} \newcommand{\ub}{\mathbf{u}} \newcommand{\ExP}[2]{\Es_{{#1}}\left [{#2}\right ]} \newcommand{\Ent}[1]{H\left [{#1}\right ]} \newcommand{\Prob}[2]{\Pr_{{#1}}\left [{#2}\right ]} \DeclareMathOperator{\supo}{\mathrm{supp}} \newcommand{\KL}[2]{\KLs\left({#1 }\parallel {#2}\right)} \newcommand{\Div}[2]{\mathrm{Div}\left({#1 }\parallel {#2}\right)} \newcommand{\Dalpha}[2]{\Dal\left({#1 }\parallel {#2}\right)} \newcommand{\UN}{\tilde{\mathcal{N}}} \newcommand{\UNQ}{\UN_Q} \newcommand{\supp}[1]{\supo\left [{#1}\right ]} \newcommand{\Prjo}[1]{\mathrm{Proj}_0\left({#1}\right)} \newcommand{\Proj}[2]{\mathrm{Proj}_{{#1}}\left({#2}\right)} \newcommand{\Cyc}[2]{{#1}_{\langle {#2} \rangle}} \newcommand{\Cycv}[1]{\mathrm{Cyc}_{\langle {#1} \rangle}} \DeclareMathOperator{\Drm}{\mathbb{D}} \DeclareMathOperator{\Dal}{\mathbb{D}_\alpha} \DeclareMathOperator{\ovec}{\mathbf{\omega}} \newcommand{\Darg}[3]{\Drm_{#1}\left({#2} \parallel {#3}\right)} \DeclareMathOperator{\DiffG}{\mathcal{L}} \newcommand{\Gcal}{\mathcal{G}} \newcommand{\Gb}{\mathbf{G}} \newcommand{\ic}{\mathbf{j}} \newcommand{\detb}[1]{\det\left({#1}\right)} \newcommand{\tranb}[1]{{\left({#1}\right)}^T} \newcommand{\br}[1]{\left({#1}\right)} \newcommand{\sqb}[1]{\left[{#1}\right]} \newcommand{\inv}[1]{{\left(#1\right)}^{-1}} \newcommand{\inve}[1]{{#1}^{-1}} \newcommand{\invet}[1]{{#1}^{-T}} \newcommand{\vecb}[1]{\mathrm{vec}\left({#1}\right)} \newcommand{\diag}{\mathrm{diag}} \newcommand{\diagb}[1]{\mathrm{diag}\left({#1}\right)} \newcommand{\tr}[1]{\mathrm{tr}\left( {#1} \right)} \newcommand{\stt}{\left[\begin{smallmatrix} \s_t\\ \s_{t+1} \\ 1\end{smallmatrix}\right]} \newcommand{\risk}{\alpha} \newcommand{\nf}{r} \newcommand{\nc}{n_u} \newcommand{\ns}{n_s} \DeclareMathOperator{\Np}{N_+} \DeclareMathOperator{\Nn}{N_-} \newcommand{\nn}{p} \newcommand{\nm}{\mathrm{m}} \newcommand{\nh}{\mathrm{k}} \newcommand{\costu}{R} \newcommand{\costn}{\tilde{\ell}} \newcommand{\costf}{\costt_f} \newcommand{\costtraj}{\mathcal{L}} \newcommand{\costr}{\costt_r} \newcommand{\costw}[2]{\costt\left({#1};{#2}\right)} \newcommand{\costx}{Q} \newcommand{\costt}{\ell} \newcommand{\cost}{\costt_t} \newcommand{\costJ}{\mathcal{J}} \newcommand{\costfin}{\costt_f} \newcommand{\Qcost}{\mathcal{Q}_t} \newcommand{\costU}{\mathcal{R}} \newcommand{\costq}{\mathcal{Q}} \newcommand{\costfq}{q} \newcommand{\Qb}{\mathbf{Q}} \newcommand{\qb}{\mathbf{q}} \newcommand{\Lm}{\mathcal{L}} \newcommand{\Lop}{\tilde{L}} \newcommand{\R}{\mathbb{R}} \DeclareMathOperator{\co}{\bar{c}} \newcommand{\s}{x} \newcommand{\snew}{\s^\prime} \DeclareMathOperator{\snewy}{\mathbf{y}} \newcommand{\sv}{\mathbf{s}} \DeclareMathOperator{\y}{\mathbf{y}} \DeclareMathOperator{\Traj}{\mathbf{X}} \newcommand{\noisetraj}{\pmb{\epsilon}} \newcommand{\noiseutraj}{{\bf \noiseu}} \newcommand{\trajpr}{\traj^\prime} \newcommand{\trajp}[1]{\traj_{+}^{({#1})}} \newcommand{\trajn}[1]{\traj_{-}^{({#1})}} \newcommand{\traj}{\mathbf{x}} \newcommand{\uc}{u_c} \newcommand{\ug}{\mathit{u}} \newcommand{\ugb}{\mathbf{u}} \newcommand{\ucb}{\ugb_\mathrm{c}} \newcommand{\usc}{\ug_{\mathrm{net}}} \newcommand{\uca}{u} \newcommand{\uopp}{\ug_\mathrm{a}} \newcommand{\uoppb}{\ugb_\mathrm{a}} \newcommand{\uoth}{\tilde{u}} \newcommand{\uopt}{\uc^\ast} \newcommand{\Uspace}{\mathcal{U}} \newcommand{\Uspaceopp}{\mathcal{U}_{\mathrm{a}}} \newcommand{\Xspace}{\mathcal{X}} \newcommand{\yg}{\mathbf{y}} \newcommand{\ytraj}{\mathbf{y}} \newcommand{\yb}{\bar{y}} \newcommand{\noiseu}{\omega} \newcommand{\noise}{\omega} \newcommand{\noisew}{\epsilon} \DeclareMathOperator{\T}{\mathcal{T}} \DeclareMathOperator{\Tfin}{N_f} \DeclareMathOperator{\Tex}{N_e} \DeclareMathOperator{\Th}{N} \newcommand{\Nh}{N} \newcommand{\Po}{\Pi} \newcommand{\Pipas}{\Pi^{0}} \newcommand{\ppr}{P_\mathrm{prior}} \newcommand{\Ut}[2]{\pic\left({#1}|{#2}\right)} \newcommand{\Pas}[2]{\Pipas\left({#1}|{#2}\right)} \newcommand{\PolOpt}[1]{\Po^*\left({#1}\right)} \DeclareMathOperator{\pic}{\Po_c} \DeclareMathOperator{\picb}{\mathbf{\Po_c}} \DeclareMathOperator{\pig}{\Po} \DeclareMathOperator{\pigb}{\Po} \DeclareMathOperator{\piopp}{\Po_\mathrm{a}} \DeclareMathOperator{\pioppb}{\mathbf{\Po}_\mathrm{a}} \DeclareMathOperator{\piopts}{{\pic}^\ast} \DeclareMathOperator{\pibopts}{{\picb}^\ast} \DeclareMathOperator{\pigopts}{{\pig}^\ast} \DeclareMathOperator{\pigbopts}{{\pigb}^\ast} \DeclareMathOperator{\pioppopts}{{\piopp}^\ast} \DeclareMathOperator{\pioppbopts}{{\pioppb}^\ast} \newcommand{\pigs}[2]{\pig_{#2}\left({#1}\right)} \newcommand{\piopt}[2]{\piopts\left({#1};{#2}\right)} \newcommand{\pibopt}[2]{\pibopts\left({#1};{#2}\right)} \newcommand{\pigopt}[2]{\pig^\ast_{#2}\left({#1}\right)} \newcommand{\pigbopt}[2]{\pig^{\ast}_{{#2}}\left({#1}\right)} \newcommand{\pioppopt}[2]{\pioppopts_{{#2}}\left({#1}\right)} \newcommand{\pioppbopt}[2]{\pioppbopts_{{#2}}\left({#1}\right)} \newcommand{\piop}[1]{\piopts\left({#1}\right)} \newcommand{\pigop}[1]{\pigopts\left({#1}\right)} \newcommand{\pioppop}[1]{\pigopts_\mathrm{a}\left({#1}\right)} \newcommand{\Polw}[3]{\Po^{#3}\left({#1}|{#2}\right)} \newcommand{\Polws}[1]{\Po^{#1}} \newcommand{\PolwT}[2]{\Po^{#2}\left({#1}\right)} \newcommand{\PolPasT}[1]{\Pipas\left({#1}\right)} \newcommand{\PolPas}[2]{\Pipas\left({#1}|{#2}\right)} \newcommand{\PolDet}[2]{\Pi_{{#2}}\left({#1}\right)} \newcommand{\valt}[2]{\mathit{v}_{#2}\left({#1}\right)} \newcommand{\valtw}[3]{\mathit{v}^{{#3}}_{{#2}}\left({#1}\right)} \newcommand{\valts}[1]{\mathit{v}_{#1}} \newcommand{\val}[1]{\mathit{v}\left({#1}\right)} \newcommand{\vals}{\mathit{v}} \newcommand{\zal}[1]{\mathit{z}\left({#1}\right)} \newcommand{\zalt}[2]{\za_{#2}{\left(#1\right)}} \newcommand{\zalts}[1]{\za_{#1}} \newcommand{\zals}{\za} \DeclareMathOperator{\za}{\mathit{z}} \DeclareMathOperator{\vt}{\tilde{v}} \DeclareMathOperator{\fct}{a} \DeclareMathOperator{\Bct}{B} \DeclareMathOperator{\Cct}{C} \DeclareMathOperator{\Dct}{D} \DeclareMathOperator{\Detf}{f} \newcommand{\Pt}[2]{\Ptrans\left({#1}|{#2}\right)} \newcommand{\Pts}[1]{\Ptrans\left({#1}\right)} \DeclareMathOperator{\Ptrans}{\mathbb{P}} \newcommand{\Detff}[4]{\dynf\left({#1},{#2},{#3},{#4}\right)} \newcommand{\Pn}{\mathbb{P}_{\noisew}} \newcommand{\PnS}{\tilde{\Sigma}} \newcommand{\dynf}{\mathcal{F}} \newcommand{\feats}{\phi} \newcommand{\feat}[2]{\feats\left({#1},{#2}\right)} \DeclareMathOperator{\Kd}{\mathcal{K}} \DeclareMathOperator{\K}{K} \newcommand{\Kb}{\mathbf{K}} \newcommand{\Acal}[2]{F\left({#1}\right)} \newcommand{\Bcal}[2]{G\left({#1}\right)} \newcommand{\Aca}{F} \newcommand{\Bca}{G} \newcommand{\FP}[1]{\rho\left({#1}\right)} \newcommand{\Ab}{\mathcal{S}} \newcommand{\Htwo}{\mathcal{H}_2} \newcommand{\Hinf}{\mathcal{H}_\infty} \newcommand{\LQR}{\mathrm{LQR}} \newcommand{\C}{\mathcal{C}} \newcommand{\M}{\mathcal{M}} \newcommand{\Ad}{\tilde{A}} \newcommand{\FHinf}[1]{q_\infty} \newcommand{\FHtwo}[1]{q_2} \newcommand{\FHone}[1]{q_1} \newcommand{\FCtwo}[1]{q^c_2\left({#1}\right)} \newcommand{\FCinf}[1]{q^c_\infty\left({#1}\right)} \DeclareMathOperator{\Adts}{\tilde{\mathcal{A}}} \DeclareMathOperator{\Ats}{\tilde{A}} \DeclareMathOperator{\A}{A} \DeclareMathOperator{\Bd}{\mathcal{B}} \DeclareMathOperator{\B}{B} \newcommand{\Sigmab}[1]{\Sigma_{\mathrm{big}}\left({#1}\right)} \newcommand{\SigmaAltb}[1]{\tilde{\Sigma}_{\mathrm{big}}\left({#1}\right)} \DeclareMathOperator{\Sigmain}{\Sigma_{in}} \DeclareMathOperator{\Sigmaint}{\tilde{\Sigma_{in}}} \DeclareMathOperator{\Sigmat}{\tilde{\Sigma}} \newcommand{\Unoise}{\Sigma} \newcommand{\Unoisei}{S} \newcommand{\mup}{\overline{m}} \newcommand{\md}{\underline{m}} \section{Energy Function and Convexity} In this section, we study the convexity of the energy function \eqref{eq:Energyrho}. We first discuss why convexity is an interesting and useful property, and then describe our results characterizing the domain of convexity of the energy function \eqref{eq:Energyrho}. Before we proceed, let us define the domain of convexity precisely. \begin{definition} The domain of convexity of the energy function is defined as: \[\D=\left\{\br{\rho,\theta}:\nabla^2 E\br{\alpha \rho,\alpha \theta}\succeq 0 \quad \forall \alpha\in[0,1]\right\}\] \end{definition} \begin{remark} In general, there could be several (disconnected) regions over which the energy function is convex . We require that for any $\br{\rho,\theta}\in\D$, the energy function is convex at every point on the line segment connecting the origin to $\br{\rho,\theta}$. This ensures that we pick the region of convexity that includes the origin. \end{remark} \subsection{Why is Convexity Interesting?} The major contribution of this paper is the characterization of the convexity domain of the energy function \eqref{eq:Energyrho}. We now discuss why this characterization is important and interesting. The principal reason is that solutions of the PF equations are stationary points of the energy function \eqref{eq:pfvar}. We will show, in section , that one can construct a convex optimization problem that will either find a solution of the PF equations within the convexity domain, or otherwise certify that there exist no solutions within the convexity domain. In this Section, we justify restricting ourselves to PF solutions contained in $\D$, as the solutions in $\D$ show some additional desirable properties. \subsubsection{Asymptotic Stability} The dynamics of the power system, under certain reasonable assumptions, reduce to the so-called ``swing dynamics'', see e.g. . In this model, each generator is modeled as a constant voltage source behind a transient reactance. For each generator, there is an ``internal'' node (modeled as a constant voltage source subjected to the second order angular dynamics) and a ``terminal'' node (a node with $0$ active and reactive power injections). We denote the set of internal generator nodes by $\G_0$. The swing equation at an internal node $i\in\G_0$ is \begin{align} M_i\ddot{\theta_i}+D_i\dot{\theta_i}=-\frac{\partial E\br{\rho,\theta}}{\partial \theta_i} \end{align} At all other nodes $i\in\G\cup\Lo$, we have the \emph{algebraic} active and reactive power balance equations \eqref{eq:pf}. The dynamic state variables are $\{\dot{\theta_i}:i\in\G_0\}\cup\{\theta_i:i\in\G_0\}$. \begin{thm} Let $\br{\rho,\theta}$ be an equilibrium of the swing dynamics such that $\br{\rho,\theta}\in\D$ and satisfy all the algebraic equations \eqref{eq:pf}. Then, $\br{\rho,\theta}$ is asymptotically stable, that is, linearizing \eqref{eq:pvdyn} around $\br{\rho,\theta}$ results in a stable linear system. \end{thm} \begin{proof} This follows from Lemma 5.1 in , which is a stronger result. \end{proof} \subsubsection{Existence of Solutions} The following result is contained in our recent manuscript : \begin{thm} Suppose that the network is a tree. Then, the PF equations \eqref{eq:pf} have a solution if and only if a solution is contained within $\D$. \end{thm} Thus, at least for trees, it is sufficient to look for solutions of the PF equations contained in $\D$. Preliminary numerical evidence (section ) suggests that this observation also extends to mesh networks. Further, as we shall illustrate numerically in Section , the solutions within $\D$ are the ``desirable'' solutions (they have small phase differences between neighboring buses and sufficiently large voltage magnitudes at all \PQ~buses). \subsection{Characterization of Domain of Convexity of the Energy Function} We now derive our main results on convexity of the energy functions. Since stationary points of the energy functions correspond to solutions of the PF equations, which are known to have multiple isolated solutions, the energy function is not globally convex. Thus, we focus on searching for restricted domains within which the energy functions are convex. In particular, we will characterize a subset of $\D$ such that phase differences between neighbors are smaller than $90\deg$, since this is typical for normal power systems operations. It was shown in \citep{bent2013synchronization} that this restriction is actually sufficient for convexity in the case when all buses are \PV. However, when the network also contains \PQ~buses additional conditions are required. \begin{thm} The energy function $E\br{\rho,\theta}$ is jointly convex in $\br{\rho,\theta}$ over the convex domain $\C\subseteq\D$ given by \begin{align} &|\theta_i-\theta_j|\leq \frac{\pi}{2}\quad\forall (i,j)\in\E\\ &\sum_{i\in\Lo} \left[2B_i-\sum_{j\in\G\cup\{\Sb\}}B_{ij}\frac{\expb{\rho_{ji}}}{\cos\br{\theta_{ji}}}\right]_{ii}^{\Lo}\nonumber \\ & \,-\sum_{(i,j)\in\E,i,j\in\Lo}\frac{B_{ij}}{\cos\br{\theta_{ij}}} \brs{\begin{pmatrix}\expb{\rho_{ji}} & 1 \\ 1 & \expb{\rho_{ij}}\end{pmatrix}}_{ij}^{\Lo}\succeq 0 \end{align} Further, suppose that the network has a tree topology. Then, the domain of function convexity, $\D$, equals $\C$. Also, the energy function is \emph{strictly convex} on $\Int\br{\C}$. \end{thm} \begin{proof} See appendix section . \end{proof}A number of Remarks are in order. \begin{remark}{\it Simple special cases}\\ When no \PQ~nodes are connected to each other, the second term of the semi-definite inequality \eqref{eq:ConvC2} is an empty sum. Thus, the semi-definite inequality is equivalent to \[2B_i=2\br{\sum_{j\sim i} B_{ij}} \geq \sum_{j\in\G\cup\{\Sb\},j\sim i}B_{ij}\frac{\Vset_j\expb{-\rho_i}}{\cos\br{\theta_i-\theta_j}}\] Since $B_i=\sum_{j\sim i} B_{ij}$, it suffices that $\Vset_i\expb{-\rho_j}\leq 2\cos\br{\theta_i-\theta_j}$ for this to be true, which is satisfied if $V_j\geq .7 \Vset_i$, $|\theta_i-\theta_j|\leq \frac{\pi}{4}$. This is a fairly reasonable restriction in practice. A similar condition was discussed in . However, it was imposed only at the equilibrium point where the energy function reaches its minimum, as a sufficient condition to guarantee (small deviation) asymptotic stability. The asymptotic stability follows from our results as well, which guarantees that the equilibrium point lies within the domain of convexity of the energy function and it is hence asymptotically stable. Also, if there are no \PQ~nodes, the matrix inequality condition in \eqref{eq:ConvC2} is empty and hence the convexity condition reduces to requiring that all phases differences (over existing lines) are smaller than $\frac{\pi}{2}$ (see ). When \PQ-\PV~connections are present, the semidefinite inequality does not simplify. However, in section , we show numerically that the conditions imposed by the convexity domain are not very restrictive and most practical power flow solutions lie within the convexity domain. \end{remark} \begin{remark}{\it Conservatism} \\ For the tree networks, we have shown that the convexity conditions are necessary and sufficient (assuming that we are interested only in the domain with phase differences over all the lines smaller than $\frac{\pi}{2}$). For meshed networks, these conditions are sufficient but, possibly, not necessary. This is due to the fact that our convexity analysis has treated the edge variables $\theta_{ij}$ as independent for each edge $(i,j)\in\E$ while in reality the variables are coupled since $\theta_{ij}=\theta_i-\theta_j$. Analysis of the gap is left for future work. We present some numerical results indicating that the gap is small in section . \end{remark} \subsection{Solving Power Flow Equations via Convex Optimization} Although energy function was developed initially as a tool to assess dynamic stability, the convexity results also yields consequences useful in the context of a number of other important power system applications. We choose to emphasize here one of these applications: Solution of the PF equations, leaving the discussion of other potential applications for future work. As we have already seen, the energy function technique provides a variational characterization of the PF equations \eqref{eq:pfvar}. Therefore, finding solution of the PF equations is equivalent to finding a stationary point of the energy function. \begin{corollary} Let $S\subset \Int\br{\C}$. Then, the power flow equations \eqref{eq:pf} have a solution $\br{\opt{\rho},\opt{\theta}}\in S$ if and only if the following optimization problem has its (unique) optimal solution in $S$: \begin{align} \min_{\br{\rho,\theta}\in \C,\theta_{\Sb}=0,\rho_{\Sb}=0} E\br{\rho,\theta} \end{align} \end{corollary} \begin{proof} If $\br{\opt{\rho},\opt{\theta}}\in S$ solves \eqref{eq:pf}, it satisfies $\nabla_{\rho}E\br{\opt{\rho},\opt{\theta}}=0,\nabla_{\theta}E\br{\opt{\rho},\opt{\theta}}=0$. Thus, it also satisfies the KKT conditions for the optimization problem \eqref{eq:optE}, forcing the respective Lagrange multipliers to be $0$. Hence, the solution is optimal for \eqref{eq:optE}. Conversely, suppose that $\br{\opt{\rho},\opt{\theta}}\in S\subset\mathrm{int}\br{\C}$ solves \eqref{eq:optE}. Then, by complementary slackness, the Lagrange multipliers are $0$ and hence the KKT conditions reduce to $\nabla_{\rho}E\br{\rho,\theta}=0,\nabla_{\theta}E\br{\rho,\theta}=0$. Thus, $\br{\opt{\rho},\opt{\theta}}\in\C$ also solves \eqref{eq:pf}. Further, by strict convexity of the energy function $E\br{\rho,\theta}$ on $\Int\br{\C}$, there can be at most one stationary point of $E$ in $\Int\br{\C}$, and hence, the power flow solution, if it exists, is unique. \end{proof} \section{Related Work} In this section, we discuss and contrast our work with related previous work. \subsection{Principal Singular Surfaces} The work closest in spirit to our approach is presented in the series of papers . The authors of consider lossless networks where all buses are \PV~buses. They define Singular surfaces, which are closed surfaces where $\detb{\nabla^2 E}=0$. The Principal Singular Surface (PSS) is the unique surface that encloses the origin $\theta=0$ and the Principal Region is the region enclosed by it. The principal region coincides with our definition of $\D$ when all nodes are \PV~nodes, apart from the additional requirement that the phase differences are smaller than $\frac{\pi}{2}$. Our results characterizing $\C$ can be seen as constructing convex inner approximations to the principal region. This analysis was extended in , where a number of counterintuitive properties of the set of solutions of the power flow equations were shown , even in the simple case when all buses are \PV~buses. Based on these results, it was conjectured in this paper that the principal region is nonconvex in general. Further, there are examples of networks for which the only solutions to the power flow equations are unstable solutions. However, it was conjectured that if there is a stable solution, there must be one in the principal region. Extending this analysis and verifying the conjectures for our general setting with \PQ~nodes is a direction of future work. \subsection{Convex Relaxations of OPF} The authors of the series of recent papers have studied \emph{convex relaxations} of Optimal Power Flow (OPF) over tree networks. Various conditions were uncovered under which the convex relaxations of the OPF problems are exact, in the sense that an optimal solution of the original OPF problem can be recovered from the solution of the convex relaxation. In this work, we primarily deal with the PF equations, not OPF. However, for lossless networks our approach applies to arbitrary topologies and can potentially be extended to the OPF formulation, following the logic first time sketched in : \begin{align*} \max_{P,Q \in T} \min_{\br{\rho,\theta}\in S} & -\sum_{i\in\G\cup\Lo}P_i \theta_i -\sum_{i\in\Lo} Q_i\rho_i \nonumber \\ & +\sum_{\br{i,j}\in\E} B_{ij}\br{\frac{\expb{2\rho_i}+\expb{2\rho_j}}{2}}\nonumber\\ & -\sum_{\br{i,j}\in\E} B_{ij}\expb{\rho_i+\rho_j}\cos\br{\theta_{ij}}\nonumber \\ & \quad -\lambda \costt\br{P,Q} \end{align*} where $\cost\br{P,Q}$ is a convex cost on the injections and $T,S$ are convex operational constraint sets on the injections and voltages, respectively and $\lambda>0$ is a positive scaling factor. This is a convex-concave saddle point problem, as long as $S\subset \C$ and can be solved efficiently using ideas from duality . If the constraints in $S$ are not binding at the optimal solution, the inner minimization effectively enforces the PF equations \eqref{eq:pf}, since the minimization corresponds to setting the gradient of the energy function to $0$. The outer maximization then optimizes a cumulative cost consisting of the negative injection cost with an extra term added -- energy function value at the optimum over line flows (corresponding to solution of the PF). We can choose $\lambda$ large enough so that the energy function term can be neglected for the outer maximization, so that we are effectively solving the OPF problem: \begin{align*} \min_{P,Q,\rho,\theta} & \costt\br{P,Q}\\ \text{Subject to } & \text{\eqref{eq:pf}}, \br{P,Q}\in T,\br{\rho,\theta} \in S \end{align*} Further developments on this line of work, along with conditions when our minimax formulation is equivalent to the OPF problem above, will be pursued in subsequent work. \subsection{Conditions on Existence of solutions to PF equations} Several papers have studied conditions for existence of solutions to the Power Flow Equations . In , the authors propose a sufficient condition for the insolvability of power flow equations based on a convex relaxation. However, our approach differs from these in the following important ways: \begin{itemize} \item To solve the PF equations in $\C$, we provide necessary and sufficient conditions, that is, our approach finds a solution in $\C$ if and only if there exists a solution in $\C$. \item Our approach is algorithmic, that is, we provide an algorithm (based on convex optimization) that is guaranteed to find the solution efficiently (in polynomial time). \item If there are additional operational constraints $\br{\rho,\theta}\in S$, with $S\subset \C$, e.g. correspondent to line protection and/or line flow limits, we can additionally answer the question of whether there exists a PF solution with $\br{\rho,\theta}\in S$. This is an important contribution, since most of the time system operators are interested in finding power flow solutions that additionally satisfy operational constraints. \end{itemize} In , the authors also propose an algorithm based on a contraction mapping consideration. However, the algorithm only works in a small ball around the origin in the $\br{P,Q}$ space. Our results are stated in terms of a nonlinear convex constraint in $\br{\rho,\theta}$ space. Precisely understanding the set of $\br{P,Q}$ for which the solution $\br{\rho,\theta}\in\C$, and conversely the set of $\br{\rho,\theta}$ for which $\br{P,Q}$ lies in a certain ball, is still an open problem, even for the special case where all buses are \PV~buses. This setting was studied and the results were exteded and connected with results on synchronization in coupled oscillators in a series of recent papers . The authors provide distinct sufficient and necessary conditions on the injections for the existence of power flow solutions with phase differences satisfying certain bounds. The authors also note that obtaining non-conservative sufficient conditions is still an open problem, and show numerically that a necessary condition is ``almost'' sufficient for most power systems. Extending this analysis to the setting with \PQ~buses and obtaining non-conservative sufficient conditions on the injections that guarantee existence of power flow solutions is an interesting direction for future work. \section{Numerical Illustrations} This Section presents numerical experiments illustrating theoretical results resented above. \subsection{Existence of Solutions: 2-bus Network} Here we analyze the toy case of a two bus network with one generator and one load. The generator also serves as the slack bus. Thus, the only variables are the voltage phase $\theta$ and voltage magnitude $\expb{\rho}$ at the load bus, and the parameters are the active power demand $P$ and the reactive power demand $Q$ at the load bus. The results, plotted in the figure show the power flow solutions shown on the heat (iso-line) map of the energy function. There are two solutions for small loads. The convexity domain is the region inside the dashed green arc. The solution plotted as a green dot is a dynamically stable solution in the interior of the convexity domain. The solution plotted as a red dot is a dynamically unstable solution outside the convexity domain. As the loads are increased, the two solutions move closer to each other. At a certain critical value, the two solutions hit the boundary of the convexity domain and then vanish, thus showing that, beyond this critical loading level, there are no longer any solutions to the PF equations. This simple experiment supports our conjecture that in general, if there is a solution to the PF equations, there must be one found within the convexity domain. As stated earlier, this result was proven for networks with tree topology . In the special case when all nodes are the \PV~buses, this result was conjectured (without proof) in , with the caveat that one is only looking for \emph{dynamically stable} solutions. We plan to investigate this conjecture for the general case with \PQ~buses in further work. \subsection{Theoretical versus Numerical Domain of Convexity} In this section, we compare numerically these regions of the energy function and the reduced energy function convexity. In order to be able to visualize these regions, we restrict ourselves to a 3 bus toy system with one PV bus and two \PQ buses. Bus 1 is the PV bus (also the slack bus): its voltage magnitude $V_1$ is fixed to 1 pu and its phase is selected as a reference, $\theta_1=0$. Buses 2 and 3 are \PQ buses with variable voltage magnitude $V_2=\expb{\rho_2},V_3=\expb{\rho_3}$ and voltage phase $\theta_2,\theta_3$. $V_2,V_3$ are determined by solving the reactive PF equations \begin{align*} & Q_2=B_{12}\br{V_2^2-V_2\cos\br{\theta_2}}+B_{23}\br{V_2^2-V_2V_3\cos\br{\theta_2-\theta_3}} \\ & Q_3=B_{13}\br{V_3^2-V_3\cos\br{\theta_3}}+B_{23}\br{V_3^2-V_2V_3\cos\br{\theta_2-\theta_3}} \end{align*} Active power injections are chosen as $P_1=3.7,P_2=-1.7,P_3=-2$ (these numbers are in the normalized units) and the reactive power demands are $Q_2=-1.05,Q_3=-1.24$. The susceptances on the lines are $B_{12}=26.88,B_{13}=26.88$. We consider two cases for $B_{23}$ that lead to different network topologies: In the first case buses $2,3$ are not connected so $B_{23}=0$ and in the second case we choose $B_{23}=16.67$. Given $\theta$, we can solve the reactive power flow equations \eqref{eq:pqb} for $\rho$: We denote the solution by $\rho\br{\theta}$. We study the convexity domain of the reduced energy function $E\br{\rho\br{\theta},\theta}$ so that we can plot the results (since the reduced energy function is only a function of two variables, $\br{\theta_1,\theta_2}$). For different values of $\theta_2,\theta_3\in\left[-\frac{\pi}{3},\frac{\pi}{3}\right]$, we compute the value of the energy function $E$ by solving the reactive PF equations for $\rho_2,\rho_3$ (using the Newton-Raphson iterative algorithm) and then plug the results into $E\br{\rho,\theta}$. For any value of $\theta_2,\theta_3$, we also estimate the Hessian $\nabla^2_{\theta}E\br{\rho(\theta),\theta}$ by means of numerical differentiation. This gives us the exact region over which the reduced energy function, $E\br{\rho(\theta),\theta}$, is a convex function of $\theta$. We compare this to the domain $\C$ of the energy function (and thus reduced energy function) convexity predicted by our theorems. Since the theorems only predict joint convexity of $E\br{\rho,\theta}$, we deduce reduced convexity by finding the set of $\theta$ such that $\br{\rho\br{\theta},\theta}\in \C$. Fig.~ shows the actual and predicted regions of the reduced energy function convexity. For the 3-bus case considered here, our theoretical results predict the region of convexity exactly. In Fig.~, we plot the convexity domain for the case $B_{23}=0$. All other Figures show the case of $B_{23}>0$ with a scaled-up injection. The scaling is 3 and 5.5 and 6 in Fig.~, Fig.~ and Fig.~ respectively. As the loading increases, the region of convexity shrinks and eventually we reach the critical load where the region of convexity becomes empty. At this point, there is no longer a stable power flow solution within the given domain, since the energy function must be convex around a stable equilibrium point. These results confirm that, at least for the 3-bus case, our theorems provide an exact description for the region of convexity. \subsection{Power Flow In Larger Networks: Existence of Solutions} In Section , we argued that one justification for seeking solutions only within the energy function's domain of convexity is that there is an evidence to suggest that if there are no solutions within the Convexity Domain, then there are no solutions anywhere. We have a proof of this for the tree case and we conjecture that the result holds in general. In order to examine this hypothesis, we have studied PF solutions for two IEEE test networks: The IEEE 14 bus network and the IEEE 118 bus network. We first modify these networks to become lossless. (we simply set conductances to $0$ for all the transmission lines.) We start with the nominal base load profiles for these networks and gradually scale up the active power injections at all the buses (except the slack bus) by a factor $\kappa$ and scale up the reactive power injections by $\delta\kappa$ at all the \PQ~buses. In order to verify insolvability of the PF equations, we use the technique developed in along with the code implementing this technique in the MATPOWER package . For a fixed value of $\delta$, we increase $\kappa$ until the technique from detects insolvability of the power flow equation. Our convex solver does not find a PF solution when the optimal solution to \eqref{eq:optE} lies on the boundary of $\C$ or when the norm of the gradient of energy function with respect to $\br{\rho,\theta}$ is non-zero at the optimum. As $\kappa$ increases, we plot the norm of the gradient of the energy function at the optimum (a non-zero value indicates that there are no solutions in the convexity domain) and the value of the insolvability test (marking the result by $1$ in the case of insolvability and $0$ otherwise). For the IEEE 14 test case, the results are shown in Figures , and . One observes that for $\delta=1$, the scaling at which our convex approach fails to find a solution is exactly the point at which the technique from detects insolvability. For smaller values of $\delta$, there is a slight gap between the two, but the values are still fairly close. For the IEEE 118 bus test case, the results are shown in Figures , and . We observe here (again) that the point at which detects insolvability is close to the point at which our convex approach fails to find a power flow solution. The gap here is larger than the 14 bus case, but it is still rather small (difference of $.1$ in $\kappa$, which amounts to a $2.5\ The results overall show that our convex approach finds solutions to the power flow equations in almost all the cases when a solution exists. There is a potential gap close to the boundary of insolvability of the PF equations. We further observed in all the cases when our algorithm reported no solution, the Newton-Raphson based approaches implemented in MATPOWER also failed to find a solution. Based on this observation, we conjecture that the gap is actually due to the approximate nature of the test (based on a sufficient but not necessary guarantees for the infeasibility). \subsection{Operational Constraints and Convexity} The convexity condition \eqref{eq:ConvC1}\eqref{eq:ConvC2} is a complicated nonlinear matrix inequality. In this Section, we describe the simple sufficient conditions for this inequality to be satisfied by computing bounds on the phase differences $|\theta_i-\theta_j$ and voltage magnitude ratios $\expb{|\rho_i-\rho_j|}=\max\br{\frac{V_i}{V_j},\frac{V_j}{V_i}}$. We fix a bound on the log-voltage differences $\expb{|\rho_i-\rho_j|}\leq b_\rho$ and compute a bound on the phase differences $|\theta_i-\theta_j|\leq b_\theta$ so that \[\{\br{\rho,\theta}:\expb{|\rho_i-\rho_j|}\leq b_\rho,|\theta_i-\theta_j|\leq b_\theta\quad \forall \br{i,j}\in\E\}\subseteq \C\] It is easy to see that the phase differences may be different for different lines, so we compute the tightest bound among all the lines. For the 14 bus case, the results are plotted in Fig.~. The results show that if we allow voltage magnitude ratios across lines to be $\expb{|\rho_i-\rho_j|}\leq 1.5$, the phase differences can still be as large as $50\deg$. This is very reasonable for practical power systems, where these bounds are seldom exceeded. For the 118 bus case, the results are plotted in Fig.~. Again, if we allow for voltage differences of up to $\expb{|\rho_i-\rho_j|}\leq 1.5$, phase differences can still be as large as $45\deg$. To conclude, our experimental results show that with fairly lax constraints on the voltage magnitude ratios and phase differences across neighboring buses, we are guaranteed to fall within the domain of convexity. In other words, most practical solutions to the PF equations lie within the domain of convexity. \title{Bare Advanced Demo of IEEEtran.cls\\ for Computer Society Journals} \author{Michael~Shell,~\IEEEmembership{Member,~IEEE,} John~Doe,~\IEEEmembership{Fellow,~OSA,} and~Jane~Doe,~\IEEEmembership{Life~Fellow,~IEEE} \IEEEcompsocitemizethanks{\IEEEcompsocthanksitem M. Shell is with the Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, 30332.\protect\\ E-mail: see http://www.michaelshell.org/contact.html \IEEEcompsocthanksitem J. Doe and J. Doe are with Anonymous University.} \thanks{Manuscript received April 19, 2005; revised September 17, 2014.}} \markboth{Journal of \LaTeX\ Class Files,~Vol.~13, No.~9, September~2014} {Shell \MakeLowercase{\textit{et al.}}: Bare Advanced Demo of IEEEtran.cls for Journals} \IEEEtitleabstractindextext{ \begin{abstract} The abstract goes here. \end{abstract} \begin{IEEEkeywords} Computer Society, IEEEtran, journal, \LaTeX, paper, template. \end{IEEEkeywords}} \maketitle \IEEEdisplaynontitleabstractindextext \IEEEpeerreviewmaketitle \ifCLASSOPTIONcompsoc \IEEEraisesectionheading{\section{Introduction}} \else \section{Introduction} \fi \IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file'' for IEEE Computer Society journal papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \ifCLASSOPTIONcompsoc \section*{Acknowledgments} \else \section*{Acknowledgment} \fi The authors would like to thank... \ifCLASSOPTIONcaptionsoff \newpage \fi \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to {\LaTeX}}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \begin{IEEEbiography}{Michael Shell} Biography text here. \end{IEEEbiography} \begin{IEEEbiographynophoto}{John Doe} Biography text here. \end{IEEEbiographynophoto} \begin{IEEEbiographynophoto}{Jane Doe} Biography text here. \end{IEEEbiographynophoto} \title{Bare Demo of IEEEtran.cls for Conferences} \author{\IEEEauthorblockN{Michael Shell} \IEEEauthorblockA{School of Electrical and\\Computer Engineering\\ Georgia Institute of Technology\\ Atlanta, Georgia 30332--0250\\ Email: http://www.michaelshell.org/contact.html} \and \IEEEauthorblockN{Homer Simpson} \IEEEauthorblockA{Twentieth Century Fox\\ Springfield, USA\\ Email: homer@thesimpsons.com} \and \IEEEauthorblockN{James Kirk\\ and Montgomery Scott} \IEEEauthorblockA{Starfleet Academy\\ San Francisco, California 96678--2391\\ Telephone: (800) 555--1212\\ Fax: (888) 555--1212}} \maketitle \begin{abstract} The abstract goes here. \end{abstract} \IEEEpeerreviewmaketitle \section{Introduction} This demo file is intended to serve as a ``starter file'' for IEEE conference papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \section*{Acknowledgment} The authors would like to thank... \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to \LaTeX}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \title{Bare Demo of IEEEtran.cls for Computer Society Conferences} \author{\IEEEauthorblockN{Michael Shell} \IEEEauthorblockA{School of Electrical and\\Computer Engineering\\ Georgia Institute of Technology\\ Atlanta, Georgia 30332--0250\\ Email: http://www.michaelshell.org/contact.html} \and \IEEEauthorblockN{Homer Simpson} \IEEEauthorblockA{Twentieth Century Fox\\ Springfield, USA\\ Email: homer@thesimpsons.com} \and \IEEEauthorblockN{James Kirk\\ and Montgomery Scott} \IEEEauthorblockA{Starfleet Academy\\ San Francisco, California 96678-2391\\ Telephone: (800) 555--1212\\ Fax: (888) 555--1212}} \maketitle \begin{abstract} The abstract goes here. \end{abstract} \IEEEpeerreviewmaketitle \section{Introduction} This demo file is intended to serve as a ``starter file'' for IEEE conference papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \ifCLASSOPTIONcompsoc \section*{Acknowledgments} \else \section*{Acknowledgment} \fi The authors would like to thank... \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to \LaTeX}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \title{Bare Demo of IEEEtran.cls for Journals} \author{Michael~Shell,~\IEEEmembership{Member,~IEEE,} John~Doe,~\IEEEmembership{Fellow,~OSA,} and~Jane~Doe,~\IEEEmembership{Life~Fellow,~IEEE} \thanks{M. Shell is with the Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, 30332 USA e-mail: (see http://www.michaelshell.org/contact.html).} \thanks{J. Doe and J. Doe are with Anonymous University.} \thanks{Manuscript received April 19, 2005; revised September 17, 2014.}} \markboth{Journal of \LaTeX\ Class Files,~Vol.~13, No.~9, September~2014} {Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for Journals} \maketitle \begin{abstract} The abstract goes here. \end{abstract} \begin{IEEEkeywords} IEEEtran, journal, \LaTeX, paper, template. \end{IEEEkeywords} \IEEEpeerreviewmaketitle \section{Introduction} \IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file'' for IEEE journal papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \section*{Acknowledgment} The authors would like to thank... \ifCLASSOPTIONcaptionsoff \newpage \fi \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to \LaTeX}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \begin{IEEEbiography}{Michael Shell} Biography text here. \end{IEEEbiography} \begin{IEEEbiographynophoto}{John Doe} Biography text here. \end{IEEEbiographynophoto} \begin{IEEEbiographynophoto}{Jane Doe} Biography text here. \end{IEEEbiographynophoto} \title{Bare Demo of IEEEtran.cls\\ for Computer Society Journals} \author{Michael~Shell,~\IEEEmembership{Member,~IEEE,} John~Doe,~\IEEEmembership{Fellow,~OSA,} and~Jane~Doe,~\IEEEmembership{Life~Fellow,~IEEE} \IEEEcompsocitemizethanks{\IEEEcompsocthanksitem M. Shell is with the Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, 30332.\protect\\ E-mail: see http://www.michaelshell.org/contact.html \IEEEcompsocthanksitem J. Doe and J. Doe are with Anonymous University.} \thanks{Manuscript received April 19, 2005; revised September 17, 2014.}} \markboth{Journal of \LaTeX\ Class Files,~Vol.~13, No.~9, September~2014} {Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for Computer Society Journals} \IEEEtitleabstractindextext{ \begin{abstract} The abstract goes here. \end{abstract} \begin{IEEEkeywords} Computer Society, IEEEtran, journal, \LaTeX, paper, template. \end{IEEEkeywords}} \maketitle \IEEEdisplaynontitleabstractindextext \IEEEpeerreviewmaketitle \IEEEraisesectionheading{\section{Introduction}} \IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file'' for IEEE Computer Society journal papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \ifCLASSOPTIONcompsoc \section*{Acknowledgments} \else \section*{Acknowledgment} \fi The authors would like to thank... \ifCLASSOPTIONcaptionsoff \newpage \fi \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to \LaTeX}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \begin{IEEEbiography}{Michael Shell} Biography text here. \end{IEEEbiography} \begin{IEEEbiographynophoto}{John Doe} Biography text here. \end{IEEEbiographynophoto} \begin{IEEEbiographynophoto}{Jane Doe} Biography text here. \end{IEEEbiographynophoto} \title{Bare Demo of IEEEtran.cls for \textsc{Transactions on Magnetics}} \author{\IEEEauthorblockN{Michael Shell\IEEEauthorrefmark{1}, Homer Simpson\IEEEauthorrefmark{2}, James Kirk\IEEEauthorrefmark{3}, Montgomery Scott\IEEEauthorrefmark{3}, and Eldon Tyrell\IEEEauthorrefmark{4},~\IEEEmembership{Fellow,~IEEE}} \IEEEauthorblockA{\IEEEauthorrefmark{1}School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332 USA} \IEEEauthorblockA{\IEEEauthorrefmark{2}Twentieth Century Fox, Springfield, USA} \IEEEauthorblockA{\IEEEauthorrefmark{3}Starfleet Academy, San Francisco, CA 96678 USA} \IEEEauthorblockA{\IEEEauthorrefmark{4}Tyrell Inc., 123 Replicant Street, Los Angeles, CA 90210 USA} \thanks{Manuscript received December 1, 2012; revised September 17, 2014. Corresponding author: M. Shell (email: http://www.michaelshell.org/contact.html).}} \markboth{Journal of \LaTeX\ Class Files,~Vol.~13, No.~9, September~2014} {Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for Journals} \IEEEtitleabstractindextext{ \begin{abstract} The abstract goes here. \end{abstract} \begin{IEEEkeywords} IEEEtran, journal, \LaTeX, magnetics, paper, template. \end{IEEEkeywords}} \maketitle \IEEEdisplaynontitleabstractindextext \IEEEpeerreviewmaketitle \section{Introduction} \IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file'' for IEEE \textsc{Transactions on Magnetics} journal papers produced under \LaTeX\ using IEEEtran.cls version 1.8a and later. I wish you the best of success. \hfill mds \hfill September 17, 2014 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \section*{Acknowledgment} The authors would like to thank... \ifCLASSOPTIONcaptionsoff \newpage \fi \begin{thebibliography}{1} \bibitem{IEEEhowto:kopka} H.~Kopka and P.~W. Daly, \emph{A Guide to \LaTeX}, 3rd~ed.\hskip 1em plus 0.5em minus 0.4em\relax Harlow, England: Addison-Wesley, 1999. \end{thebibliography} \begin{IEEEbiography}{Michael Shell} Biography text here. \end{IEEEbiography} \begin{IEEEbiographynophoto}{John Doe} Biography text here. \end{IEEEbiographynophoto} \begin{IEEEbiographynophoto}{Jane Doe} Biography text here. \end{IEEEbiographynophoto} \section{Introduction} Power systems are experiencing revolutionary changes. Integration of renewable generation, distributed generation, smart metering, direct or price-based load-control capabilities have contributed to the revolution, but are also posing significant operational challenges by making the power system inherently stochastic and inhomogeneous. With these changes, the system operators will no longer have the luxury of large positive and negative reserves. Moreover, operating the future power grid will require developing new computational tools that can assess the system state and its operational margins more accurately and efficiently than current approaches. Specifically, these new techniques need to go beyond linearized methods of analysis and ensure that the power system is safe even in the presence of large disturbances and uncertainty. In this paper, we study the energy function approach and argue that this classic tool can be generalized to help solve a number of existing and emerging problems in operating modern power systems. Introduced more than 50 years ago by Aylett , energy functions have been used to develop direct methods of the power systems stability and analysis. Essentially, energy functions are physically motivated Lyapunov functions for certain simplified models of the power system dynamics. In the 70s and 80s, many results were derived for stability analysis of power systems,first via approximate Kron-reduction techniques and then using more accurate structure-preserving energy functions . Developments of this period are summarized in . Since then, much of the work has focussed on developing effective heuristics to produce non-conservative estimates of the region of power system stability (see and references therein). Interest in these approaches declined in the 1990s and 2000s, partially due to the difficulty of generalizing these approaches to more detailed power system models, but also due to advances in computing technologies that made stability analysis based on real-time dynamic simulations feasible. Our results are also linked to more recent lines of research related to (a) necessary and sufficient conditions guaranteeing existence and uniqueness power flow solutions ; (b) analysis of the fixed point characterizations of the power flow equations and analytical approximation schemes for power flow solutions ; (c) semidefinite relaxations of optimal power flow and various extensions (see for example). We argue in this manuscript that energy functions can be viewed as a unifying tool for several operational problems in power systems. This is based on the fact that for lossless systems (resistance of transmission lines are ignored) stationary points of energy functions are solutions of the Power Flow (PF) equations. Hence, many traditional operational problems dependent on the notion of the (static) PF equations, such as the PF analysis itself, the Critical Loadability Problem (CLP) and Optimum Power Flow (OPF) problem,can be studied using the energy function methodology. This viewpoint, first presented in , is extended here both in terms of theory and applications. Moreover, we show that this viewpoint extends to the more general cases where the energy-like functions are not true Lyapunov functions, but simply functions whose stationary points correspond to PF solutions. This extension will allow us to deal with lossy systems that have fixed resistance to reactance ratio for all transmission lines. In order to obtain efficient formulations, we study a property of the energy functions that, to the best of our knowledge, has previously received little attention: \emph{convexity}. Since power systems are designed to operate at a stable equilibrium point, the energy (Lyapunov) functions characterizing their behavior are obviously convex at least in a small vicinity of the stable equilibrium. However, to the best of our knowledge, the exact domain of convexity has never been characterized. The classical papers by Bergen and Hill on the structure preserving model of power systems, extended in to account for effects of the reactive power flows and variable voltages, form a solid base for our analysis. We study various assumptions under which the energy function is {\bf\it provably convex}. It is well-known that when all nodes are the so-called \PV~ nodes (active power consumption/production and voltages are fixed/known), the energy function is convex as long as all phase differences are smaller than $90$ degrees. In this paper, we generalize this result to the networks with \PQ~nodes (i.e. nodes where active and reactive power injections/consumptions are fixed while voltage magnitudes are variable). Our main technical result is a description of the domain over which the energy function is convex. We provide a nonlinear but convex matrix inequality condition on the voltage phases and magnitudes that guarantees convexity of the energy function. Moreover, we show that this condition provides an \emph{exact} characterization of the domain of convexity of the energy function for the case of acyclic networks. For mesh networks, in general, the matrix inequality condition provides an inner approximation of the convexity domain. The condition for convexity is quite natural and captures the intuition that if voltage magnitudes and voltage phases at neighboring nodes are ``not too far'' from each other then the underlying energy function is convex. The rest of this paper is organized as follows. In section , we describe the mathematical background and power system models. In section , we first analyze convexity of the energy function jointly in voltage magnitudes and phases, and then describe how the convexity can be used to design a provably convergent algorithm solving the PF equations (Section ). In section , we discuss related work and describe the advantages of the energy function methodology. In section , we provide numerical illustrations of our approach on the IEEE 14 and 118 bus networks. Section contains conclusions and a brief discussion of the future work. Finally, Appendices and are reserved for the proof of the main theorem and a generalization to a special class of lossy networks. \newcommand{\ygu}{y} \newcommand{\opta}[1]{{#1}^{\ast}_{\risk}} \newcommand{\E}{\mathcal{E}} \newcommand{\G}{\mathrm{pv}} \newcommand{\Vs}{\mathcal{V}} \newcommand{\Lo}{\mathrm{pq}} \newcommand{\Loo}{\mathrm{pq0}} \newcommand{\Lon}{\mathrm{pq1}} \newcommand{\Vl}{\underline{V}} \newcommand{\Vset}{v} \newcommand{\Vdif}{\zeta} \newcommand{\Vpow}{\beta} \newcommand{\Bt}{\tilde{B}} \newcommand{\Qt}{\tilde{Q}} \newcommand{\af}{\alpha} \newcommand{\df}{\delta} \newcommand{\Vu}{\overline{V}} \newcommand{\V}{\mathcal{V}} \newcommand{\PQ}{$(P,Q)$~} \newcommand{\PV}{$(P,V)$~} \newcommand{\KE}{KE} \newcommand{\Et}{\tilde{E}} \newcommand{\Aa}{A^1} \newcommand{\Ac}{A^2} \newcommand{\Int}{\mathrm{Int}} \newcommand{\Sb}{\mathcal{S}} \usepackage{verbatim} \usepackage{algorithm} \usepackage{bm} \usepackage{algpseudocode} \usepackage[all]{xy} \usepackage{url} \usepackage{amsmath,amsthm,amssymb} \usepackage{graphicx} \usepackage{subcaption} \newtheorem{thm}{Theorem}[section] \newtheorem{theorem}{Theorem}[section] \newtheorem{lem}[thm]{Lemma} \newtheorem{Thm}{Theorem} \newtheorem{corollary}{Corollary} \newtheorem{MyLem}{Lemma} \newtheorem{lemma}{lemma} \theoremstyle{definition} \newtheorem{definition}{Definition} \newtheorem{Def}{definition} \theoremstyle{remark} \newtheorem{remark}{Remark} \begin{comment} \input{Duality.tex} We present numerical results applying this method to the 6 bus network shipped with MATPOWER (named ``case6ww''). We modify this network to have $0$ resistance and ignore shunt impedances etc, to get a network with purely inductive transmission lines. The network has 3 \PV and 3 \PQ buses. We investigate the following question:\\ Our convexity results require the reactive power demand to grow with voltage. If we make this approximation even though it is incorrect, do we get approximate but reliable solutions to the power flow equations? We investigate this by starting with the model with the given real and reactive power demands (for which a solution to the power flow equations exists and can be found by the default MATPOWER AC power flow solver). We then increase all the loads by a constant factor $\beta$. For a grid of values of $\beta \in [1,1.3]$, we solve the power flow equations using the default MATPOWER AC solver, the MATPOWER DC solver and a convex approach based on our results using an interior point mehtod (OPTI MATLAB interface to IPOPT ). We compute the $\ell_2$ error in the power flow equations (normalized by the $\ell_2$ norm of $P$,$Q$) in the solution found by the three approaches as a function of $\beta$ and plot the results in figure . The results from figure show that at small load scaling factors (upto $1.06$), the MATPOWER AC solver does well and is able to find accurate solutions. In this regime, our solver is slightly less accurate but much more accurate than the DC approximation. On increasing the loading further, the MATPOWER solver fails completely and gives highly inaccurate solutions. Our solver degrades more gracefully and is still able to find reasonably accurate solutions. It remains more accurate than both the DC and AC solver for all loading factors above, but the three solvers converge at very high loading factors. This shows that our approach, although requiring a potentially artificial assumption of reactive demand dependence on voltage, can find reasonably accurate solutions when in cases where traditional solvers (AC and DC) fail or are inaccurate. In fact, in practice the assumption of constant $P$ and $Q$ is an approximation and reactive power dependence of the form $Q_i\br{V_i}\propto V_I^\alpha$ with $\alpha \in [1.3,1.6]$ is quite common, so our approximation may be even more accurate than the results presented here. \section{Other Applications} In this section we briefly mention other possible applications of the energy function convexity results. \subsection{Convex AC Optimum Power Flow} Let $P,Q$ denote the vectors of active and reactive power injections and we denote the energy function parameterized by these as $U\br{\rho,\theta;P,Q}$. Further, define a modified energy function that blows up outside the domain of convexity $\br{\rho,\theta}\in\C$. \[\tilde{U}\br{\rho,\theta;P,Q} = \begin{cases}E\br{\rho,\theta|P,Q} & \text{ if} \br{\rho,\theta}\in\mathrm{int}\br{\C}(P,Q) \\ \infty & \text{otherwise} \end{cases}\] Then formulate and solve the following new optimization problem \[\min_{\br{\rho,\theta}\in S,P,Q} \lambda c(P,Q)+\tilde{U}\br{\rho,\theta;P,Q}\] where $c\br{P,Q}$ is an exogenously given cost on the injection/consumption and $S$ is a set specifying operational constraints on $\br{\rho,\theta}$ to be contained within $\mathrm{int}\br{\C}$. If we optimize with respect to $\br{\rho,\theta}$ keeping $P,Q$ fixed, we recover the active and reactive power balance equations (assuming a solution exists within $S$). Further, by choosing $\lambda$ sufficiently large, one can guarantee that the optimization with respect to $P,Q$ is not affected by the second term except to ensure that the active and reactive power balance equations are satisfied. Observe that we have just formulated the AC Optimum Power Flow problem in terms of the energy function. As stated this optimization is not jointly convex in $\br{\rho,\theta,P,Q}$, however, one can additionally exploit ideas from the duality theory to reformulate it as a jointly convex optimization problem. For the simple case where voltage magnitudes are held fixed at all buses, related covexification strategy, based on the dual optimization formulation of PF problem introduced in , was discussed in . \subsection{Convex Model Predictive Control (MPC)} The convexity results can also be used to obtain a convex formulation of model predictive control scheme for frequency and voltage regulation. Consider an optimal control problem that seeks to adjust active and reactive power injections at generators and loads in order to minimize frequency deviations while also maintaining operational constraints on voltage magnitudes and phases (for example see ). The dynamical system under consideration here can be modeled by the DAE model given by \eqref{eq:swing} and \eqref{eq:PQReactive}. After discretizing time, one can enforce the dynamical constraints implicitly in the same way as discussed in the previous subsection. Again, using ideas from duality, the MPC optimization problem can be recast as a tractable convex optimization problem. Moreover, the resulting convex optimization will inherit the separable structure of the network and hence can be potentially solved in a distributed manner using algorithms like ADMM . In future work, we will explore building on recent attempts to explore distributed optimization algorithms for power flow problems , in order to get efficient distributed algorithms for our convex formulations of power flow and related problems. \subsection{Instantons} As the penetration of intermittent sources of energy in the grid increases (through increased adoption of renewable generation like wind and Photovoltaics), an important problem is to characterize likely configurations of uncertain generation that would cause violation of operational constraints on voltage magnitudes and phases. This problem was studied using the DC power flow approximation in where the problem was shown to be tractable through reduction to a small number of convex optimizations. The failure may be defined as line overload or nodal voltage leaving its safe range. More concretely, for overload on a line $(i,j)\in\E$, one would like to solve the following optimization problem: \begin{eqnarray} \max_w {\cal P}(w) \mbox{ s.t. } \Biggl\{\begin{array}{c} \mbox{PF Eqs. have a solution}\\ \mbox{line flow over $(i,j)$}\\ \mbox{is at the threshold} \end{array}\Biggr. \end{eqnarray} where $w$ is the vector of uncertain generation and ${\cal P}(w)$ is its probability density. As stated, the optimization \eqref{instanton} is not convex. However, as discussed in the previous sections, we can recast the power flow solution as the solution to a convex optimization of the energy function over voltages and phases. One can then generalizing the method of (using the dual problem and restating \eqref{instanton} as single stage optimization over wind, voltage magnitudes and phases) one can derive a fully convex equivalent of the AC instanton problem \eqref{instanton}). Further details on this convex AC-instanton method will be pursued in future work. \end{comment} \section{Reduced Convexity} In section , we examined various conditions under which the joint energy function $E\br{\rho,\theta}$ is jointly convex in $\br{\rho,\theta}$. However, as the numerical results (section )indicate), these results are conservative and convexity of $E\br{\rho,\theta}$ is only sufficient but not necessary for convexity of $E\br{\opt{\rho}\br{\theta},\theta}$. In this section, we examine the convexity of $E\br{\opt{\rho},\theta}$ directly. However, it will be convenient for us to work directly with the voltage variables $V$ rather than their logarithms $\rho$, so we will state our results in terms of $E\br{V,\theta}$ and $E\br{\opt{V}\br{\theta},\theta}$. \subsection{A Variational Representation for the Reduced Energy Function} In this section, we develop a variational representation for the reduced energy function. As a simple corollary, we can also define a convex optimization problem that finds the solutions to the reactive power equations (the voltage magnitudes) given the voltage phases and a simple condition for uniqueness of the solution for voltage magnitudes within a given domain. Before we proceed, we define some notation. Let $M\in\R^{n\times n}$ be such that \begin{align} M_{ij}=\frac{B_{ij}}{\sum_{j\neq i} B_{ij}} \end{align} Let $M_{\Lo}$ be the $|\Lo|\times|\Lo|$ submatrix with row and column indices in $\Lo$. and $M_{\G}$ denote the $|\Lo|\times|\G|$submatrix with row indices indices in $\Lo$ and column indices in $\G$. Further, we define the normalized reactive power consumptions by \begin{align} \tilde{Q}_{i}=\frac{Q_{i}}{\sum_{j\neq i} B_{ij}} \end{align} Further, we will assume that $Q_i>0$ at all \PQ~nodes, that is, we require that all \PQ~nodes consume reactive power. The more general case can also be dealt with, but we will focus on this case for now as the results are slightly simpler. \begin{thm} Suppose that each load is connected to at least one generator. The reduced energy function can be expressed as follows: \begin{align} &\Et\br{\theta} =\nonumber\\ &\min_{V_{\Lo}} \sum_{i\in\Lo} Q_i\logb{V_i}-\frac{1}{2}\sum_{\substack{i\in\Lo\\j\in\G}}B_{ij}V_i\Vset_j\cos\br{\theta_i-\theta_j}-\frac{1}{2}\sum_{i,j\in\G}B_{ij}\Vset_i\Vset_j\cos\br{\theta_i-\theta_j} \\ & \text{Subject to } \nonumber \\ & \br{\sum_{j\neq i}B_{ij}}V_i^2-V_i\br{\sum_{j\neq i} B_{ij}V_j\cos\br{\theta_i-\theta_j}}+Q_i \leq 0 \quad\forall i\in\Lo\\ & V_i=\Vset_i \quad \forall i \in \G \\ & V_i > \frac{2Q_i}{\sum_{j\in\G} B_{ij}\Vset_j\cos\br{\theta_i-\theta_j}}\forall i \in \Lo \end{align} as long as the above optimization problem is feasible and $|\theta_i-\theta_j|\leq \frac{\pi}{2}$ for all $(i,j)\in\E$. \end{thm} \begin{proof} We first show that the domain is compact. For $i\in\G$, $V_i$ is fixed by an equality constraint. Thus, each $V_i,i\in\Lo$ is bounded below by $0$ and above by lemma . Thus, the domain is compact and the optimization problem must have a finite minimum that is attained, if the problem is feasible. Further, the objective is decreasing in $V_i$ for each $i\in\Lo$. Indeed, differentiating the objective with respect to $V_i$, we get \[\frac{Q_i}{V_i}-\frac{1}{2}\sum_{j\in\G}B_{ij}\Vset_j\cos\br{\theta_i-\theta_j}\] which is smaller than $0$ (using \eqref{eq:ConstEqTwo}). Thus, the objective is decreasing in $V_i$. Also, if constraint \eqref{eq:ConstEqTwo} is satisfied for some value $V_i$, it is satisfied for all larger values. We will now argue that constraint \eqref{eq:SolveQuad} must be tight at the optimum. The key observation is that for a given $i$, constraint $i$ is tight at two values of $V_i$ (roots of the quadratic equation) with the inequality being satisfied strictly at all intermediate values. We will argue that at the optimum, the constraint must be tight with $V_i$ being equal to the larger root. Suppose that this is not the case, that is let $\opt{V}$ be a (locally) optimal solution with constraint \eqref{eq:SolveQuad} not being tight at the optimum for some $i\in\Lo$. Then, one can increase $\opt{V}_i$ infinitesimally and reduce the objective while maintaining feasibility with respect to all the other constraints (since the other constraints are all satisfied more easily as $\opt{V}_i$ increases). This is a contradiction to $\opt{V}$ being a locally optimal solution. Finally, it is easy to see that \eqref{eq:SolveQuad} with inequality replaced by equality simply reduces to the reactive power flow equations. Now, using lemma , the objective is equal to the reduced energy function at the optimum. \end{proof} \begin{remark} The above representation gives us a way of computing voltage magnitudes (at \PQ buses) that solve the the reactive power equations given voltage phases, as long as the problem is feasible. Although this problem is nonconvex, as long as it is feasible, \emph{any} locally optimal solution is a solution to the reactive power equations. Below, we show that this can in fact be done via \emph{convex optimization}. \end{remark} \begin{corollary} Let $c_i>0,i\in\Lo$ be an arbitrary set of positive numbers. A solution to the reactive power flow equations can always be found by solving the following convex optimization problem \begin{align} &\min_{\zeta_{\Lo}} -\sum_{i\in\Lo} c_i\zeta_i \\ & \text{Subject to } \nonumber \\ & \br{\sum_{j\neq i}B_{ij}}\zeta_i-\br{\sum_{j\neq i} B_{ij}\sqrt{\zeta_i\zeta_j}\cos\br{\theta_i-\theta_j}}+Q_i \leq 0 \quad\forall i\in\Lo\\ & \zeta_i=\Vset_i^2 \quad \forall i \in \G \end{align} as long as the above optimization problem is feasible and $|\theta_i-\theta_j|\leq \frac{\pi}{2}$ for all $(i,j)\in\E$ and setting $V_i=\zeta_i^2$. Further, by choosing all possible $c_i>0$, one can find the set of all solutions that are not dominated by any other (that is, all solutions such that no other solution to the reactive power flow equations has equal or higher voltage magnitudes at all \PQ buses). \end{corollary} \begin{proof} The objective and constraints are easily seen to be convex in $\zeta_\Lo$. Further, the objective is strictly decreasing in $\zeta_i$ for each $i$. Hence, by the same argument as the previous theorem, any optimal solution must satisfy \eqref{eqc:SolveQuad} with equality (assuming the problem is feasible) at an optimal solution. Thus, any optimal solution must be a solution to the reactive power flow equations. Further, this is a vector-convex optimization problem (see section 4.7). Thus, any vector $\opt{\zeta}_{\Lo}$ such that there is no $\zeta_{\Lo}\neq \opt{\zeta}_{\Lo}$ in the feasible set such that $\opt{\zeta}_{\Lo}\leq \zeta_{\Lo}$ is the solution to \eqref{eqc:RedEnergy} for some choice of $c_i>0$. This completes the proof. \end{proof} \begin{remark} Also note that these theorems did not require the assumption that $Q_i>0$ but the results on convexity of the reduced energy function will. \end{remark} \subsection{Convexity of the Reduced Energy Function} The representation from theorem shows that the reduced energy function is the solution to an optimization problem with an objective that is convex in $\theta$ and a convex constraint set on $\theta$. However, the objective and constraints are not convex in $V_i$ and hence we are unable to say anything about the convexity of the reduced energy function. In the following theorem, under additional assumptions, we show that the above optimization problem can be rewritten to show that the reduced energy function is convex. We will require a condition on the impedances and reactive power demands in order to get convexity. The least conservative way to state this condition is in terms of the solution to an optimization problem, which we define below: \begin{align} \overline{V_i}\br{B,Q}& =\max_{V_{\Lo}} V_i \\ &\text{Subject to }\nonumber \\ & \br{\sum_{j\neq i}B_{ij}}V_i^2-\sum_{j\neq i} B_{ij}V_iV_j+Q_i \leq 0 \quad\forall i\in\Lo \nonumber\\ & V_i=\Vset_i \quad \forall i \in \G \nonumber \end{align} This optimization problem can be solved as a convex optimization problem in the transformed variables $\zeta_i=V_i^2$. \newcommand{\Vd}{\zeta} \begin{thm} Choose any $\Vpow \in (0,1)$. Let $c_{ij}>0$ be a set of positive numbers such that $Q_i=\sum_{j\neq i} c_{ij}B_{ij}$ for each $i\in\Lo$. Suppose that \begin{align} \overline{V_i}\br{B,Q}\leq\sqrt{\frac{1+\Vpow}{1-\Vpow}}\sqrt{\tilde{Q}_i} \quad\forall i \in \Lo \end{align} The reduced energy function can be expressed as \begin{align} &\Et\br{\theta} =\nonumber\\ &\min_{\Vd_{\Lo}} \sum_{i\in\Lo} \Vpow Q_i\logb{\Vd_i}-\frac{1}{2}\sum_{\substack{i\in\Lo\\j\in\G}}B_{ij}{\Vd_i}^{\Vpow}\Vset_j\cos\br{\theta_i-\theta_j}-\frac{1}{2}\sum_{i,j\in\G}B_{ij}\Vset_i\Vset_j\cos\br{\theta_i-\theta_j} \\ & \text{Subject to } \nonumber \\ & \br{\sum_{j\neq i}B_{ij}}\zeta_i^\beta+Q_i\zeta_i^{-\beta}\leq \sum_{j\neq i} B_{ij}\zeta_j^\beta\cos\br{\theta_i-\theta_j} \quad \forall i\in\Lo\\ & \Vd_i={\Vset_i}^{\frac{1}{\Vpow}} \quad \forall i \in \G \\ & {\Vd_i}^{\Vpow} > \frac{2Q_i}{\sum_{j\in\G} B_{ij}\Vset_j\cos\br{\theta_i-\theta_j}}\quad\forall i \in \Lo\\ & \cos\br{\theta_i-\theta_j} \geq \frac{c_{ij}}{\Vd^\Vpow}+\sqrt{\powb{\frac{c_{ij}}{\Vd^\Vpow}}{2}+\Vpow}\quad\forall i \in \Lo \end{align} if the above optimization problem is feasible. Further, $\Et\br{\theta}$ is convex over the domain \[\left\{\theta:|\theta_i-\theta_j|\leq\arccos\br{\sqrt{\Vpow}}\forall (i,j)\in\E,i,j\in\Lo,|\theta_i-\theta_j|\leq\arccos\br{\alpha}\forall\,(i,j)\in\E\right\}\] \end{thm} \begin{proof} This problem is exactly equivalent to \eqref{eq:RedEnergy} with $\Vd_i=V_i^{\frac{1}{\Vpow}}$, except for the additional constraint \eqref{eq2:ConstEqThree}. Further, if this additional constraint is satisfied for some $\zeta_{\Lo}$, it is also satisfied for all larger $\zeta_{\Lo}$. Thus, the proof from theorem is still valid and the above optimization problem, if feasible, attains its optimum and the constraint \eqref{eq2:SolveQuad} is tight at the optimum. Now, on the given domain, it is easy to see that constraints \eqref{eq2:ConstEq},\eqref{eq2:ConstEqTwo},\eqref{eq2:ConstEqThree} are all convex constraints in $\br{\rho_{\Lo},\theta}$ by lemma . The last term in the objective is convex in $\theta$ and independent of $\zeta$. The other terms can be rewritten as: \[\sum_{i\in\Lo} \frac{1}{2}\sum_{j\in\G} B_{ij}\br{2\beta c_{ij}\logb{\zeta_i}-\zeta_i^\Vpow \cos\br{\theta_i-\theta_j}}\] Each term in the above sum jointly convex in $\br{\rho_{\Lo},\theta}$ as long as constraint \eqref{eq2:ConstEqThree} is satisfied (by lemma ). Thus, the only thing remaining is the constraint \eqref{eq2:SolveQuad}. By lemma , the RHS is concave in $\br{\rho_{\Lo},\theta}$ as long as \begin{align} \zeta_i^\Vpow \leq \sqrt{\tilde{Q}_i}\sqrt{\frac{1+\Vpow}{1-\Vpow}} \end{align} With $V_i=\powb{\zeta_i}{\frac{1}{\Vpow}}$, the constraint set defining the optimization problem \eqref{eq:Bound} is a relaxation of the constraints imposed in this problem, and hence $\overline{V_i}\br{B,Q}$ is an upper bound on $\zeta_i^\beta$ for any $\zeta_{\Lo}$ in the feasible set. Hence, the constraint \eqref{eq2:SolveQuad} is convex as well. Thus, the optimization problem is jointly convex in $\br{\zeta_{\Lo},\theta}$ and hence minimizing with respect to $\zeta_{\Lo}$ over the jointly convex domain produces a convex function of $\theta$. Hence the theorem. \end{proof} \begin{remark} We tested the condition \eqref{eq:ConvCondRed} for some test networks in MATPOWER and found that for most cases, one can choose $\beta$ to satisfy \eqref{eq:ConvCondRed} such that $\arccos\br{\sqrt{\beta}}=10-20\deg$. This is a fairly stringent requirement and indicates that our analysis can be improved, although it is encouraging that the condition covers a non-trivial set of phases. The other set of constraints are typically satisfied under normal operating conditions, so the main source of conservatism in the convexity region estimate is \eqref{eq:ConvCondRed}, which in turn comes from requiring convexity of the LHS of \eqref{eq2:SolveQuad}. Thus, if we can deal with this more precisely (for example by using a convex relaxation and arguing tightness), we might be able to improve the estimates of convexity. \end{remark} \section{Modeling Power Systems} In order to guarantee the existence of an energy function, we will need to make the standard assumptions used in the structure preserving models , which are fairly well justified for the high-voltage transmission network: \begin{itemize} \item[(1)] All transmission lines are purely reactive (network is lossless), i.e. the network resistive power losses are considered small and thus ignored. \item[(2)] Nodal active and reactive power injections/consumptions are constant. \end{itemize} \newcommand{\neb}{\sim} \subsection{Notation} The transmission network is modeled as a graph $\br{\Vs,\E}$ where $\Vs$ is the set of nodes and $\E$ is the set of edges. In power systems parlance, the nodes are called buses and the edges are called lines (transmission lines). We shall use these terms interchangeably in this manuscript. Nodes are denoted by indices $i=1,\ldots,n=|\Vs|$ and edges by ordered pairs of nodes $\br{i,j}$. We pick an arbitrary orientation for each edge, so that for an edge between $i$ and $j$, only one of $\br{i,j}$ and $\br{j,i}$ is in $\E$. If there is an edge between buses $i$ and $j$, we write $i \neb j,j\neb i$. Sometimes, it will be convenient to denote edges by a single index $k$. In this notation, $k_1$ and $k_2$ denote the ``from'' and ``to'' ends of the edge $k$. Define the edge-incidence matrix $A\in \R^{|\Vs|\times |\E|}$ as follows: \begin{align} A_{ik} = \begin{cases} 1 \text{ if } k\in\E,i\in\Vs,k_1=i\nonumber\\ -1\text{ if } k\in\E,j\in\Vs,k_2=j\in\Vs\nonumber\\ 0 \text{ otherwise } \nonumber \end{cases} \end{align} The set of vertices includes all load buses as well as buses representing generators (a single bus, or terminal and internal bus per generator). This allows us to develop a unifying framework that works for different load and generator models. Edges correspond to transmission lines. Each transmission line is characterized by its admittance $G_{ij}-\ic B_{ij}$ ($\ic=\sqrt{-1}$). However, since we assume that transmission lines are lossless, we set $G_{ij}=0$ and thus each line is characterized by a single real and positive parameter $B_{ij}$ (the negated line susceptance). Let $V_i, \theta_i, P_i$ and $Q_i$ denote voltage (magnitude), phase, active and reactive injection at the bus $i$ respectively. Let $\rho_i=\logb{V_i}$. Let $\theta_{ij}=\theta_i-\theta_j,\rho_{ij}=\rho_i-\rho_j$. Buses are of three types: \begin{itemize} \item\underline{\PV buses} where active power injection and voltage are fixed, while voltage phase and reactive power adjust as conditions (e.g. power flows) are variables. The set of \PV buses is denoted by $\G$. \item\underline{\PQ buses} where active and reactive power injections are fixed, while voltage phase and magnitude are variables. The set of \PQ buses is denoted by $\Lo$. \item\underline{Slack bus}, a reference bus at which the voltage magnitude and phase are fixed, and the active and reactive power injections are free variables. The slack bus is denoted by $\Sb$. \end{itemize} $\theta$ and $V$ (with no subscript) denote the vectors of phases at all buses except the slack bus and voltage magnitudes at all $\PQ$ buses. $\rho,V$ are indexed by $\{i:i\in \Lo\}$. For every bus $i$ define $ B_i=\sum_{j\sim i}B_{ij}$. Let $S$ be an ordered set of indices. We use the notation \[M=\brs{\begin{pmatrix}a & b\\ c & d\end{pmatrix}}_{ij}^{S}\] where $i,j\in S$, to denote an $|S| \times |S|$ matrix whose rows and columns are indexed by the entries of $S$, with $M_{ii}=a,M_{ij}=b,M_{ji}=c,M_{jj}=d$ and all other entries equal to $0$. Similarly, we use the notation $M=[a]_{ii}^{S}$ to denote an $|S|\times |S|$ matrix with the $(i,i)-th$ entry equal to $a$ and all other entries equal to $0$. Given a vector $x$ with indices $i\in S$, denote by $\mathrm{diag}\br{x}$ the $|S|\times|S|$ matrix with diagonal entries $x_i$. Given a set $\C\subset\R^n$, $\Int\br{\C}$ denotes the interior of the set. \subsection{Background} In this section, we introduce two power systems concepts that are critical for what follows: PF equations and energy functions. \subsubsection{Power Flow Equations} PF Equations model the flow of power over the power system network. They are a set of coupled nonlinear equations that follow from Kirchoff's laws applied to the AC power network. Circuit elements in the standard power systems models are all linear, if one ignores discrete elements like phase shifters and tap-changing transformers. Even though Ohm's law and Kirchhoff's law are linear in voltages and currents, power is their product and hence quadratic. Since we are interested in power, e.g., to determine power generation, specify demand, minimize line loss, PF equations are nonlinear. PF equations assume that the network is balanced, that is, the net sum of powers consumptions, injections and power dissipated is zero. This relies on the assumption that at the time-scale where the PF are solved (every 5 minutes or so), the system is in a quasi-steady state, i.e. the dynamic disturbances have been resolved through actions of the automatic control schemes represented by the automatic voltage regulators, power system stabilizers and primary and secondary frequency control systems. Since our main concern here is to study properties of the PF equations, we simply state them without derivation. At each node in the power system, there are four variables: voltage magnitude ($V_i$), voltage phase ($\theta_i$), active power injection ($P_i$), reactive power injection ($Q_i$). When solving the PF equations, at each node, two of these variables are fixed and the other two are free. This leads to $2n$ equations in $2n$ variables, where $n$ is the number of buses. Typically, power systems have two types of nodes: \PV~nodes and \PQ~nodes. At the \PV~buses, the active power injection $P_i$ and the voltage magnitude $V_i=\expb{\rho_i}$ are fixed. The fixed voltage magnitude is denoted $\Vset_i$, to distinguish it from variable voltages at other nodes. Generators are typically modeled as \PV~buses, since they have voltage regulators that inject reactive power to maintain voltages at fixed values. We can choose $\Vset_i=1$ for each $i\in\G$, since any non-unit voltage set-point can be absorbed into the susceptances $B_{ij}$ in the active and reactive power flow equations. At the \PQ~buses, the active power injection $P_i$ and the reactive power injection $Q_i$ are held fixed. Loads are typically \PQ~nodes, where $P_i$ and $Q_i$ represent the active and reactive power demands (assumed constant). Usually, the variables $Q_i$ at the \PV~buses are not considered explicitly. These variables can be obtained from the reactive power balance equations by plugging in there the values of $\br{\rho,\theta}$. Finally, we have a special bus (slack bus) which absorbs all the power imbalances in the network. The injections $P_{\Sb},Q_{\Sb}$ at this bus are free variables whose values are adjusted to guarantee the power balance and $\theta_{\Sb}=\rho_{\Sb}=0$. The power balance equations at all the buses, a set of coupled nonlinear equations in $\br{\rho,\theta}$, together constitute the PF equations: \begin{subequations} \begin{align} P_i & =\sum_{j \neb i} B_{ij}\expb{\rho_i+\rho_j}\sin\br{\theta_{ij}} ,i\in\G\cup\Lo\\ \rho_i & =0 ,i\in\G\\ Q_i &=\sum_{j \neb i} B_{ij}\br{\expb{2\rho_i}-\expb{\rho_i+\rho_j}\cos\br{\theta_{ij}}} ,i\in\Lo\\ \theta_{\Sb} &=0 ,\rho_{\Sb}=0. \end{align} \end{subequations} \subsubsection{Energy Functions} Energy functions were first introduced in the context of first integral analysis of the power system swing dynamic equations . The energy function consists of two terms -- kinetic and potential. It is also a Lyapunov function for the swing dynamics. In this paper, we are not explicitly concerned with the dynamics, so the most useful property of the energy functions is that the stationary points of the energy function potential part map to solutions of the PF equations. The PF equations \eqref{eq:pf} can be re-written in the following variational form \begin{subequations} \begin{align} \text{\eqref{eq:pva} : } 0 & =\frac{\partial E}{\partial \theta_i} \quad \forall i\in\Lo\cup\G\\ \text{ \eqref{eq:pqb} : } & 0= \frac{\partial E}{\partial \rho_i}\quad \forall i\in\Lo\\ & \rho_i=0\quad \forall i\in\G, \rho_{\Sb}=\theta_{\Sb}=0 \end{align} \end{subequations} where $E$ is called the energy function: \begin{align} E\br{\rho,\theta} & = -\sum_{i\in\G\cup\Lo}P_i \theta_i -\sum_{i\in\Lo} Q_i\rho_i \nonumber\\ & \, + \sum_{\br{i,j}\in\E} B_{ij}\br{\frac{\expb{2\rho_i}+\expb{2\rho_j}}{2}}\nonumber\\ &\, -\sum_{\br{i,j}\in\E} B_{ij}\expb{\rho_i+\rho_j}\cos\br{\theta_{ij}} \end{align} Note that $\rho_i$ is a variable for $i \in \Lo$ and $\rho_i=0$ for $i \in \G$, $\rho_{\Sb}=0,\theta_{\Sb}=0$. |
1501.04053 | Title: Stochastic Local Interaction (SLI) Model: Interfacing Machine Learning
and Geostatistics
Abstract: Machine learning and geostatistics are powerful mathematical frameworks for
modeling spatial data. Both approaches, however, suffer from poor scaling of
the required computational resources for large data applications. We present
the Stochastic Local Interaction (SLI) model, which employs a local
representation to improve computational efficiency. SLI combines geostatistics
and machine learning with ideas from statistical physics and computational
geometry. It is based on a joint probability density function defined by an
energy functional which involves local interactions implemented by means of
kernel functions with adaptive local kernel bandwidths. SLI is expressed in
terms of an explicit, typically sparse, precision (inverse covariance) matrix.
This representation leads to a semi-analytical expression for interpolation
(prediction), which is valid in any number of dimensions and avoids the
computationally costly covariance matrix inversion.
Body: \title{Stochastic Local Interaction (SLI) Model: Interfacing Machine Learning and Geostatistics} \author{Dionissios T. Hristopulos} \email{dionisi@mred.tuc.gr} \affiliation{School of Mineral Resources Engineering, Technical University of Crete, Chania 73100, Greece} \date{\today} \begin{abstract} Machine learning and geostatistics are powerful mathematical frameworks for modeling spatial data. Both approaches, however, suffer from poor scaling of the required computational resources for large data applications. We present the Stochastic Local Interaction (SLI) model, which employs a local representation to improve computational efficiency. SLI combines geostatistics and machine learning with ideas from statistical physics and computational geometry. It is based on a joint probability density function defined by an energy functional which involves local interactions implemented by means of kernel functions with adaptive local kernel bandwidths. SLI is expressed in terms of an explicit, typically sparse, precision (inverse covariance) matrix. This representation leads to a semi-analytical expression for interpolation (prediction), which is valid in any number of dimensions and avoids the computationally costly covariance matrix inversion. \end{abstract} \pacs{02.50.-r, 02.50.Tt, 05.45.Tp, 89.60.-k} \keywords{machine learning, kernel regression, geostatistics, big data, sparse methods} \maketitle \newpage \section{Introduction} Big data is expected to have a large impact in the geosciences given the abundance of remote sensing and earth-based observations related to climate~. A similar data explosion is happening in other scientific and engineering fields~. This trend underscores the need for algorithms that can handle large data sets. Most current methods of data analysis, however, have not been designed with size as a primary consideration. This has inspired statements such as: ``Improvements in data-processing capabilities are essential to make maximal use of state-of-the-art experimental facilities''~. Machine learning can extract information and ``learn'' characteristic patterns in the data. Thus, it is expected to play a significant role in the era of big data research. The application of machine learning methods in spatial data analysis has been spearheaded by Kanevski~. Machine learning and geostatistics are powerful frameworks for spatial data processing. A comparison of their performance using a set of radiological measurements is presented in~. The question that we address in this work is whether we can combine ideas from both fields to develop a computationally efficient framework for spatial data modeling. Most data processing and visualization methods assume complete data sets, whereas in practice data often have gaps. Hence, it is necessary to fill missing values by means of imputation or interpolation methods. In geostatistics, such methods are based on various flavors of stochastic optimal linear estimation (kriging)~. In machine learning, methods such as $k$-nearest neighbors, artificial neural networks, and the Bayesian framework of Gaussian processes are used~. Both geostatistics and Gaussian process regression are based on the theory of random fields and share considerable similarities~. The Gaussian process framework, however, is better suited for applications in higher than two dimensions. A significant drawback of most existing methods for interpolation and simulation of missing data is their poor scalability with the data size $N$, i.e., the $O(N^3)$ algorithmic complexity and the $O(N^2)$ memory requirements: An $O(N^p)$ dependence implies that the respective computational resource (time or memory) increases with $N$ as a polynomial of degree at most equal to $p$. Improved scaling with data size can be achieved by means of local approximations, dimensionality reduction techniques, and parallel algorithms. A recent review of available methods for large data geostatistical applications is given in~. These approaches employ clever approximations to reduce the computational complexity of the standard geostatistical framework. Local approximations involve methods such as maximum composite likelihood~ and maximum pseudo-likelihood~. Another approach involves covariance tapering which neglects correlations outside a specified range~. Dimensionality reduction includes methods such as fixed rank kriging which models the precision matrix by means of a fixed rank matrix $r \ll N$~. Markov random fields (MRFs) also take advantage of locality using factorizable joint densities. The application of MRFs in spatial data analysis was initially limited to structured grids~. However, a recently developed link between Gaussian random fields and MRFs via stochastic partial differential equations (SPDE) has extended the scope of MRFs to scattered data~. We propose a Stochastic Local Interaction (SLI) model for spatially correlated data which is based by construction on local correlations. SLI can be used for the interpolation and simulation of incomplete data in $d$-dimensional spaces, where $d$ could be larger than $3$. The SLI model incorporates concepts from statistical physics, computational geometry, and machine learning. We use the idea of local interactions from statistical physics to impose correlations between ``neighboring'' locations by means of an explicit precision matrix. The local geometry of the sampling network plays an important role in the expression of the interactions, since it determines the size of local neighborhoods. On regular grids, the SLI model becomes equivalent to a Gaussian MRF with specific structure. For scattered data, the SLI model provides an alternative to the SPDE approach that avoids the preprocessing cost involved in the latter. The SLI model extends previous research on Spartan spatial random fields~ to an explicitly discrete formulation and thus enables its application to scattered data without the approximations used in~. SLI is based on a joint probability density function (pdf) determined from local interactions. This is achieved by handling the irregularity of sampling locations in terms of kernel functions with locally adaptive bandwidth. Kernel methods are common in statistical machine learning~\citep{Vapnik00} and in spatial statistics for the estimation of the variogram and the covariance function~\citep{Elogneetal07,Garcia04,Hall94}. The remainder of the article is structured as follows: Section~ briefly introduces useful definitions and terminology. In Section~ we construct the SLI model, propose a computationally efficient parameter estimation approach, and formulate an explicit interpolation expression. In Section~ we investigate SLI interpolation using different types of simulated and real data in one, two and four dimensional Euclidean spaces. Section~ discusses potential extensions of the current SLI version and connections with machine learning. Finally, in Section~ we present our conclusions and point to future research. \section{Background Concepts and Notation} \subsection{Definition of the problem to be learned} \paragraph{Sampling grid} The set of sampling points is denoted by $\Sn=\{\bfs_{1},\ldots, \bfs_{N}\}$, where $\bfs_{i},$ $i=1,\ldots,N$ are vectors in the Euclidean space $\R^d$ or in some abstract feature space that possesses a distance metric. In Euclidean spaces, the domain boundary is defined by the convex hull, $\Hconv,$ of $\Sn.$ \paragraph{Sample and predictions} The sample data are denoted by the vector $\xsam \equiv (\Xo_{1}, \ldots, \Xo_{N})^{T}$, where the superscript ``$T$'' denotes the transpose. Interpolation aims to derive estimates of the observed field at the nodes of a regular grid $\mathcal{G} \subset {\mathbb Z}^{d}$, or at validation set points which may be scattered. The estimates (predictions) will be denoted by $\hXo(\bfz_p)$, $p=1,\ldots,P$, i.e., $\xpre =(\hXo_1, \ldots, \hXo_{P})^{T}$. \paragraph{Spatial random field model} The data $\xsam$ are assumed to represent samples from a spatial random field (SRF) $X_{i}(\om)$, where the index $i=1, \ldots, N$ denotes the spatial location $\bfs_i \in \Sn$. The expectation over the ensemble of probable states is denoted by $\E[X_{i}(\om)]$, and the autocovariance function is given by $C_{i,j} := \E[X_{i}(\om)\, X_{j}(\om)] - \E[X_{i}(\om)]\, \E[X_{j}(\om)].$ The pdf of \textit{Gibbs SRFs} can be expressed in terms of an energy functional $ H(\xsam;\bmthe) $, where $\bmthe$ is a set of \textit{model parameters}, according to the Gibbs pdf \citep[p. 51]{Winkler95} \begin{equation} f_{\rm X} (\xsam;\bmthe) = \frac{\e^{ - H(\xsam;\bmthe)} }{Z(\bmthe)}. \end{equation} \noindent The constant $ Z(\bmthe) $, called the \textit{partition function}, is the pdf normalization factor obtained by integrating $ \e^{ -H(\xsam;\bmthe) } $ over all the probable states $\xsam$. \subsection{From continuum spaces to scattered data} The formulation based on~\eqref{eq:gibbspdf} has its origins in statistical physics, and it has found applications in pattern analysis~ and Bayesian field theory, e.g.~. In statistics, this general model belongs to the exponential family of distributions that have desirable mathematical properties~. Our group used the exponential density in connection with a specific energy functional to develop Spartan spatial random fields (SSRF's)~\citep{dth03b,dthsel07,dthetal08,dthsel09}. In Section~ we construct an explicitly discrete model motivated by SSRFs which adapts local interactions to general sampling networks and prediction grids by means of kernel functions. \subsection{Kernel weights} Let $K(\bfr)$ be a non-negative-valued kernel that is either compactly supported or decays exponentially fast at large distances (e.g., the Gaussian or exponential function). We define kernel weights associated with the sampling points $\bfs_{i}$ and $\bfs_{j}$ as follows \beq K_{i,j} \doteq K\left(\frac{\bfs_{i} - \bfs_{j}}{h_{i}}\right) = K\left(\frac{\|\bfs_{i} - \bfs_{j}\|}{h_{i}}\right), \eeq \noindent where $\|\bfs_{i} - \bfs_{j}\|$ is the distance (Euclidean or other) between two points $\bfs_{i}$ and $\bfs_{j} $, whereas $h_i$ is the respective \emph{kernel bandwidth} that adapts to local variations of the sampling pattern. The kernel weight $K_{j,i}$ is defined in terms of a bandwidth $h_j$. Hence, $K_{i,j} \neq K_{j,i}$ if the bandwidths $h_i$ and $h_j$ are different. Examples of kernel functions are given in Table~. \begin{table} \centering \caption{ Definitions of kernel functions used in Section~ below. The first three have compact support. Notation: $u = \| \bfr \|/h$ where $\| \bfr \| $ is the distance and $h$ the bandwidth; $\mathbbm{1}_{A}(u)$ is the {indicator function} of the set $A$, i.e., $\mathbbm{1}_{A}(u)=1, \, u \in A$ and $\mathbbm{1}_{A}(u)=0, \,u \notin A.$ } \centering \begin{tabular}{lll} \hline Triangular & $K(u)=(1-u)\, \mathbbm{1}_{|u| \le 1}(u)$ \\ Tricube & $K(u)=(1-u^3)^3 \, \mathbbm{1}_{|u| \le 1}(u)$ \\ Quadratic & $K(u)=(1-u^2) \, \mathbbm{1}_{|u| \le 1}(u)$ \\ Gaussian & $K(u)=\exp(-u^2)$ \\ Exponential & $K(u)=\exp(-|u|)$ \\ \hline \end{tabular} \end{table} Let $D_{i,[k]}(S_{N})$ denote the distance between ${\bfs}_{i}$ and its $k$-nearest neighbor in $\Sn$ $(k=0$ corresponds to zero distance). We choose the local bandwidth associated with ${\bfs}_{i}$ according to \beq h_{i} = \mu\, D_{i,[k]}(S_{N}), \eeq \noindent where $\mu > 1$ and $k>1$ are model parameters. In several case studies involving Euclidean spaces of dimension $d=1, 2, 3, 4$, we determined that $k=2$ (second nearest neighbors) performs well for compactly supported kernels and $k=1$ (nearest neighbors) for infinitely supported kernels. Using $k=2$ for compact kernels avoids zero bandwidth problems which result from $k=1$ for collocated sampling and prediction points. Since the sampling point configuration is fixed, $\mu$ and $D_{i,[k]}(S_{N})$ determine the local bandwidths. $D_{i,[k]}(S_{N})$ depends purely on the sampling point configuration, but $\mu$ also depends on the sample values. For compactly supported kernels setting $k=1$ only makes sense if $\mu >1$; otherwise $h_{i} = D_{i,[k=1]}(S_{N})$ implying that the kernel vanishes even for the nearest-neighbor pairs and thus fails to implement interactions. \subsection{Kernel averages} For any two-point function $\Phi(\cdot)$, we use a local-bandwidth extension of the Nadaraya-Watson kernel-weighted average over the network of sampling points~ \[ \langle \Phi(\cdot) \rangle_{\bfh} = \frac{\sum_{i=1}^{N} \sum_{j=1}^{N} K_{i,j} \, \Phi(\cdot)} {\sum_{i=1}^{N} \sum_{j=1}^{N} K_{i,j} }, \] \noindent where $\bfh = (h_{1}, \ldots, h_{N})^T$ is the vector of local bandwidths. The function $\Phi(\cdot)$ represents the distance between two points $\|\bfs_{i} - \bfs_{j}\|$ or the difference $x_{i} - x_{j}$ of the field values, or any other function that depends on the locations or the values of the field. The kernel average is normalized so as to preserve unity, i.e., $\langle 1 \rangle_{\bfh} = 1$ for all possible point configurations. \section{The Stochastic Local Interaction (SLI) Model} The SLI joint pdf is determined below by means of the energy functional~\eqref{eq:fgc-disc}. This leads to a precision matrix which is explicitly defined in terms of local interactions and thus avoids the covariance matrix inversion. The prediction of missing data is based on maximizing the joint pdf of the data and the predictand, which is equivalent to minimizing the corresponding energy functional. This leads to the mode predictor~\eqref{eq:SLI-mode-pred}, which involves a calculation with linear algorithmic complexity. \subsection{The energy functional} Consider a sample $\xsam$ on an unstructured sampling grid with sample mean $\mx$. We propose the following energy functional $\Hx(\xsam;\bftheta)$ \begin{align} \Hx (\xsam ;\bftheta) & = \frac{1}{2\,\lambda \, } \left[ {\mcal{S}_0}(\xsam) + \, \alpha_1 \, {\mcal{S}_1}(\xsam;\bfh_1) \right. + \left. \alpha_{2} \, {\mcal{S}_2}(\xsam;\bfh_2) \right], \end{align} \noindent where $\bftheta=(\mx, \alpha_1, \alpha_2, \lambda, \mu, k)$ is the SLI parameter vector and the parameters $\mu, k$ are defined in Section~ above. The terms $\mcal{S}_0(\xsam)$, ${\mcal{S}_1}(\xsam;\bfh_1)$, and ${\mcal{S}_2}(\xsam;\bfh_2)$ correspond to the averages of the square fluctuations, the square gradient and the square curvature in a Euclidean space of dimension $d$. The latter two are given by kernel-weighted averages that involve the \emph{field increments} $\Xo_{i,j} = x_{i} - x_{j}$. \beq {\mcal{S}_0}(\xsam) = \frac{1}{N}\,{\sum_{i=1}^{N} (\Xo_{i}-\mx)^{2}}, \eeq \begin{equation} {\mcal{S}_1}(\xsam;\bfh_1) = \cone \, \langle \Xo_{i,j}^{2} \rangle_{\bfh_1}, \end{equation} \begin{subequations} \begin{align} {\mcal{S}_2}(\xsam;\bfh_2) = c_{2,1} \, \langle \Xo_{i,j}^{2} \rangle_{\bfh_2} - c_{2,2} \, \langle \Xo_{i,j}^{2} \rangle_{\bfh_3} - c_{2,3} \, \langle \Xo_{i,j}^{2} \rangle_{\bfh_4}, \\ \mbox{where} \; c_{2,1}= 4d(d+2), \; c_{2,2} = 2d(d-1), \; c_{2,3} = d. \end{align} \end{subequations} The $c_{2,j}$ $(j=1,2,3)$ values in $\mcal{S}_2$ are motivated by discrete approximations of the square gradient and curvature~. We use two vector bandwidths, $\bfh_1$ and $\bfh_2$, to determine the range of influence of the kernel function around each sampling point for the gradient ${\mathcal{S}_1}(\bXn; \bfh_1)$ and curvature ${\mathcal{S}_2}(\bXn; \bfh_2)$ terms respectively. Additional bandwidths used in~\eqref{eq:S2} for ${\mathcal{S}_2}(\bXn; \bfh_2)$ are defined by $\bfh_3=\sqrt{2}\,\bfh_2$, $\bfh_4=2\,\bfh_2.$ These definitions are motivated by the formulation of SSRFs~. \subsection{SLI parameters and permissibility} To obtain realistic kernel bandwidths, $k$ should be a positive integer larger than one, and $\mu $ should be larger than one. The parameter $\mu_{\rm X}$ is set equal to the sample mean. The coefficients $\alpha_1, \alpha_2$ control the relative contributions of the mean square gradient and mean square curvature terms. The coefficient $\lambda$ controls the overall amplitude of the fluctuations. Finally, $\mu$ and $k$ control the bandwidth values as described in Section~. The SLI energy functional~\eqref{eq:fgc-disc} is permissible if $\Hx (\xsam ;\bftheta) \ge 0$ for all $\xsam$, a condition which ensures that $ \e^{ -H(\xsam;\bmthe) } $ is bounded and thus the existence of the partition function in~\eqref{eq:gibbspdf}. Assuming that $\mcal{S}_2 \ge 0$ ($\mcal{S}_0$ and $\mcal{S}_1$ are always non-negative by construction), a sufficient permissibility condition, independently of the distance metric used, is $\alpha_1, \alpha_2, \lambda >0$. In all the case studies that we have investigated, however, we have not encountered permissibility problems so long as $\alpha_1, \alpha_2, \lambda >0$. Intuitively, the justification for the permissibility of~\eqref{eq:fgc-disc} is that the first average, i.e., $\langle \Xo_{i,j}^{2} \rangle_{\bfh_2}$ in~\eqref{eq:S2} has a positive sign and is multiplied by $c_{2,1}$, which is significantly larger (especially as $d$ increases) than the coefficients $c_{2,2}$ and $c_{2,3}$ multiplying the negative-sign averages $\langle \Xo_{i,j}^{2} \rangle_{\bfh_3}$ and $\langle \Xo_{i,j}^{2} \rangle_{\bfh_4}$. This property is valid for geodesic distances on the globe and for other metric spaces as well. \subsection{Precision matrix representation} We express~\eqref{eq:fgc-disc} in terms of the \emph{precision matrix} $\hat{J}_{i,j}(\bftheta)$ ($i,j =1, \ldots, N$) \begin{equation} \Hx (\xsam ;\bftheta) = \frac{1}{2} (\xsam - \bmx)^{T} \, {\mathbf J}(\bftheta) \,(\xsam - \bmx). \end{equation} The symmetric precision matrix ${\mathbf J}(\bftheta)$ follows from expanding the squared differences in~\eqref{eq:fgc-disc}, leading to the following expression \begin{align} {\mathbf J}(\bftheta) = \frac{1}{\lambda } & \,\left\{ \frac{{\bf I}_{N}}{N} + \alpha_1 \, \cone \, {\mathbf J}_{1}({\bfh}_1) + \alpha_2 \,\left[ c_{2,1} \, {\mathbf J}_{2}({\bfh}_2) - c_{2,2} \, {\mathbf J}_{3}({\bfh}_3) - c_{2,3} \, {\mathbf J}_{4}({\bfh}_4)\right] \right\}, \end{align} \noindent where ${\bf I}_{N}$ is the $N\times N$ identity matrix: $[{\bf I}_{N}]_{i,j}=1$ if $i=j$ and $[{\bf I}_{N}]_{i,j}=0$ otherwise, and ${\mathbf J}_{q}({\bfh}_q), \; q=1,2,3,4$ are \emph{network matrices} that are determined by the sampling pattern, the kernel function, and the bandwidths. The index $q$ defines the gradient network matrix for $q=1$, whereas the values $q=2, 3, 4$ specify the curvature network matrices that correspond to the three terms in ${\mcal{S}_2}(\xsam;\bfh_2)$ given by~\eqref{eq:S2}. The elements of the network matrices ${\mathbf J}_{q}(\bfh_{q})$ are given by the following equations \begin{subequations} \begin{align} [{\mathbf J}_{q}({\bfh}_q)]_{i,j} & = - \w(h_{q;i}) -\w(h_{q;j}) + [{\bf I}_{N}]_{i,j} \, \sum_{l=1}^{N} \left[ u_{i,l}(h_{q;i}) + u_{l,i}(h_{q;l})\right], \\ \w(h_{q;i}) & = \frac{K\left( \frac{\bfs_{i} - \bfs_{j}}{h_{q,i}} \right)} {\sum_{i=1}^{N} \sum_{j=1}^{N} K\left( \frac{\bfs_{i} - \bfs_{j}}{h_{q,i}} \right)}, \quad q = 1, \ldots, 4. \end{align} \end{subequations} The network matrices defined by~\eqref{eq:network-mat} are symmetric by construction. It follows from~\eqref{eq:network-mat} that the row and column sums vanish, i.e., \beq \sum_{j=1}^{N} [{\mathbf J}_{q}({\bfh}_q)]_{i,j} = 0. \eeq \noindent Based on~\eqref{eq:Jtilde}, the diagonal elements are given by the following expression \beq [{\mathbf J}_{q}({\bfh}_q)]_{i,i} = \sum_{l=1,\neq i}^{N} \left[ u_{i,l}(h_{q;i}) + u_{l,i}(h_{q;l})\right]. \eeq \noindent Since the kernel weights are non-negative, it follows that the sub-matrices ${\mathbf J}_{q}({\bfh}_q)$ are \emph{diagonally dominant}, i.e., $\big| [{\mathbf J}_{q}({\bfh}_q)]_{i,i} \big| \ge \sum_{j \neq i} \big|[{\mathbf J}_{q}({\bfh}_q)]_{i,j} \big|$. It also follows from~\eqref{eq:prec-mat} and~\eqref{eq:sum-net-mat} that \begin{subequations} \beq \sum_{j=1}^{N} [{\mathbf J}(\bmthe)]_{i,j} = \frac{1}{N \lambda}. \eeq \end{subequations} \subsection{Parameter inference} We have experimented both with maximum likelihood estimation and leave-one-out cross validation. The former requires the calculation of the SLI partition function, which is an $O(N^3)$ operation for scattered data. For large data sets the $O(N^3)$ complexity is a computational bottleneck. Parameter inference by optimization of a cross validation metric is computationally more efficient, since it is at worst an $O(N^2)$ operation as we show below. The memory requirements for storing the precision matrix are $O(N^2)$ but can be significantly reduced by using sparse matrix structures. Let $\bftheta_{-\lambda}=(\alpha_{1},\alpha_{2},\mu,\mx)^{T}$ represent the parameter vector excluding $\lambda$. We use the following \emph{cross validation cost functional} \beq \Phi(\xsam;\bftheta_{-\lambda}) = \sum_{i=1}^{N} | \hat{\Xo}_{i}(\bftheta_{-\lambda}) - \Xo_{i} \,|, \eeq where $\hat{\Xo}_{i}(\bftheta_{-\lambda})$ is the SLI prediction at $\bfs_{i}$ based on the reduced sampling set $\Sn - \{ \bfs_{i} \}$ using the parameter vector $\bftheta_{-\lambda}$ which applies to all $i=1, \ldots, N$. The prediction is based on the interpolation equation~\eqref{eq:sli-mode-pred-J} below and does not involve $\lambda$ (see discussion in Section~). The optimal parameter vector excluding $\lambda$, i.e., $\bftheta_{-\lambda}$, is determined by minimizing the cost functional~\eqref{eq:cost}: \beq \bftheta^{\ast}_{-\lambda} = \arg \min_{\bftheta_{-\lambda}} \Phi(\xsam;\bftheta_{-\lambda}). \eeq \noindent If $\tilde{H}(\xsam;\bftheta_{-\lambda})$ is the energy estimated from~\eqref{eq:H-using-J} and~\eqref{eq:prec-mat} by setting $\lambda =1$, the optimal value $\lambda^{\ast}$ is obtained by minimizing the negative log-likelihood with respect to $\lambda$ leading to the following solution (see~) \beq \lambda^{\ast} = \frac{2\tilde{H}(\xsam;\bftheta_{-\lambda})}{N}. \eeq We determine the minimum of the cross validation cost functional~\eqref{eq:cost} using the {\sc Matlab} constrained optimization function \verb+fmincon+ with the \emph{interior-point} algorithm~. This function determines the local optimum nearest to the initial parameter vector. We use initial guesses for the parameters $\al_{1}, \al_{2}, \mu$, and we assume that the parameters are constrained between the lower bounds $[0.5, 0.5, 0.5 ]$ and the upper bounds $ [300, 300, 15 ]$. We investigated different initial guesses for the parameters which led to different local optima. We found, however, that the value of the cross validation function is not very sensitive on the local optimum. In the $4D$ case study presented in Section~, we also estimate for comparison purposes the global optimum using {\sc Matlab}'s global optimization tools. \subsection{Predictive SLI model} Let us now assume that the prediction point $\bfz_p$ is added to the sampling points. To predict the unknown value of the field at $\bfz_p$, we insert this point in the energy functional~\eqref{eq:H-using-J}, which is then given by Eq.~\eqref{eq:H-pred-J} below. Then, we determine the mode of the joint pdf~\eqref{eq:gibbspdf} with the prediction point inserted in the energy functional. Thus, we obtain a \emph{mode prediction equation} for $\hXo_{p}$ given by~\eqref{eq:sli-mode-pred-J} below. \subsubsection{Modification of kernel weights} Upon inclusion of $\bfz_p$, the weights~\eqref{eq:J-weights} of the network matrices~\eqref{eq:Jtilde} are modified as follows \beq \w(h_{q;i}) = \frac{K\left( \frac{\bfs_{i} - \bfs_{j}}{h_{q,i}} \right)} {\sum_{i,j} K\left( \frac{\bfs_{i} - \bfs_{j}}{h_{q,i}} \right) + \sum_{i} K\left( \frac{\bfs_{i} - \bfs_{p}}{h_{q,i}}\right) + \sum_{i} K\left( \frac{\bfs_{i} - \bfs_{p}}{h_{q,p}}\right) }, \eeq \noindent where $\sum_{i,j} := \sum_{i=1}^{N} \sum_{j=1}^{N} $. The first term in the denominator concerns interactions between sampling points. The second term involves local interactions between the sampling points and the prediction point which result from inserting the prediction point in the local neighborhoods of the sampling points, which control the bandwidths. Finally, the third term also involves interactions between the prediction point and the sampling points, but in this case the bandwidth is controlled by the former. Fig.~ below illustrates the difference between the second and third term in the context of the entire precision matrix. The index $q$ is used to distinguish between the weights linked to the gradient $(q=1)$ and the three weights $(q=2, 3, 4)$ linked to the curvature terms. The only difference between weights with different $q$ is the bandwidth. In the case of compactly supported kernels, different bandwidths imply that different numbers of pairs are involved in the summations, since a pair separated by a distance that exceeds the bandwidth does not contribute. Calculation of the predictand contributions in the denominator of~\eqref{eq:weights-pred} is an operation with computational complexity $O(N)$ compared to $O(N^2)$ for the interactions between sampling points. The latter term, however, is calculated once and used for all the prediction points. In addition to the weights that correspond to pairs of sampling points, there are weights for combinations of sampling and prediction points, i.e., \beq u_{p,j}(h_{q;p}) = \frac{K\left( \frac{\bfs_{p} - \bfs_{j}}{h_{q,p}} \right)} {\sum_{i,j} K\left( \frac{\bfs_{i} - \bfs_{j}}{h_{q,i}} \right) + \sum_{i} K\left( \frac{\bfs_{i} - \bfs_{p}}{h_{q,i}}\right) + \sum_{i} K\left( \frac{\bfs_{i} - \bfs_{p}}{h_{q,p}}\right) }, \eeq \noindent where $p=1, \ldots, P$, $j=1, \ldots, N$. The denominator of~\eqref{eq:weights-pred-pi} is identical to that of~\eqref{eq:weights-pred}. \subsubsection{SLI mode predictor} Using the precision matrix formulation, the energy functional including the predictand is given by \begin{align} {\hHx}(\xsam, \Xo_p ; \bftheta^{\ast}) = & \Hx(\xsam ; \bftheta^{\ast}) + J_{p,p}(\bftheta^{\ast})(\Xo_p - \mx)^2 + \sum_{i=1}^{N} (\Xo_{i} - \mx) \,J_{i,p}(\bftheta^{\ast})\,(\Xo_{p} - \mx) \nonumber \\ & + \sum_{i=1}^{N} (\Xo_{p} - \mx) \,J_{p,i}(\bftheta^{\ast})\,(\Xo_{i} - \mx). \end{align} The elements of the precision matrix that involve the prediction point are \begin{subequations} \begin{align} [{\mathbf J}_{q}({\bfh}_q)]_{p,p} = & \sum_{i=1}^{N} \left[ u_{i,p}(h_{q;i}) + u_{p,i}(h_{q;p})\right], \\ [{\mathbf J}_{q}({\bfh}_q)]_{i,p} = & - \left[ u_{i,p}(h_{q;i}) + u_{p,i}(h_{q;p})\right], \; i \neq p. \end{align} \end{subequations} Based on~\eqref{eq:prec-Jp-off} the symmetry property $J_{p,i}(\bftheta^{\ast}) =J_{i,p}(\bftheta^{\ast})$ follows. The coefficients $u_{i,p}(h_{q;i})$ and $u_{p,i}(h_{q;p})$ differ due to the different bandwidths used (in the former, the bandwidth is determined by the neighborhood of the sampling point $\bfs_{i}$, whereas in the latter by the neighborhood of $\bfz_{p}$.) A schematic illustration of terms in~\eqref{eq:H-pred-J} that involve the predictand is given in Fig.~. The left diagram corresponds to terms ``rooted'' at $\bfz_p$ (i.e., with coefficient $u_{p,i}(h_{q;p})$ that involves the bandwidth $h_p$), whereas the right hand side diagram corresponds to terms rooted at the sampling points, i.e., with coefficients $u_{i,p}(h_{q;i})$. The \emph{SLI mode predictor} is defined by the following equation \beq \hXo_p = \arg \min_{\Xo_p} {\hHx}(\xsam, \Xo_p ; \bftheta^{\ast}), \eeq \noindent where $ {\hHx}(\xsam, \Xo_p ; \bftheta^{\ast})$ is given by~\eqref{eq:H-pred-J}. Minimization of the energy with respect to $x_p$ leads to the following {mode estimator} \begin{align} \hXo_{p} & = \mx - \frac{\sum_{i=1}^{N} \, \left[ J_{i,p}(\bftheta^{\ast}) + J_{p,i}(\bftheta^{\ast}) \right] \, (\Xo_{i} - \mx)}{2\, J_{p,p}(\bftheta^{\ast})} \nonumber \\ & = \mx - \frac{\sum_{i=1}^{N} \, J_{p,i}(\bftheta^{\ast}) \, (\Xo_{i} - \mx)}{ J_{p,p}(\bftheta^{\ast})}, \end{align} \noindent where the precision matrix elements are given by~\eqref{eq:Jtilde} using the modified kernel weights~\eqref{eq:weights-pred} and~\eqref{eq:weights-pred-pi}. The SLI mode predictor can be generalized to $P$ prediction points as follows \begin{subequations} \begin{equation} \bhXo_{p} = \bmx - \mathbf{\tilde{J}}_{P,S}(\bftheta^{\ast}) \, (\bXo - \bmx) , \end{equation} \noindent where $\mathbf{\tilde{J}}_{P,S}(\bftheta^{\ast})$ is a $ P \times N$ matrix given by \begin{equation} [\mathbf{\tilde{J}}_{P,S}(\bftheta^{\ast})]_{p,i} = J_{p,i}(\bftheta^{\ast})/ J_{p,p}(\bftheta^{\ast}). \end{equation} \end{subequations} \subsubsection{Properties of SLI predictor} The SLI prediction~\eqref{eq:SLI-prediction} is \emph{unbiased} in view of the vanishing row sum property~\eqref{eq:sum-J} satisfied by the network matrices and the precision matrix. The SLI prediction~\eqref{eq:SLI-prediction} is independent of the parameter $\lambda$ which sets the amplitude of the fluctuations, because the transfer matrix $\mathbf{\tilde{J}}_{P,S}(\bftheta^{\ast})$ is given by the ratio of precision matrix elements. This property is analogous to the independence of the kriging predictor from the random field variance. Hence, leave-one-out cross validation does not determine the optimal value of $\lambda$, which is obtained from~\eqref{eq:lambda}. The SLI predictor is not necessarily an exact interpolator. In particular, let us consider a point $\bfs_{k}$, $k \in \{1, \ldots, N\}$, which is very close to $\bfz_p$. Based on~\eqref{eq:prec-Jp} and~\eqref{eq:sli-mode-pred-J}, $\hXo_{p} \rightarrow \Xo_k$ as $\bfz_p \rightarrow \bfs_k$ only if (i) $|u_{k,p}(h_p)| \gg |u_{i,p}(h_p)|$ and (ii) $|u_{k,p}(h_k)| \gg |u_{i,p}(h_i)|$ for all $ i \neq k$. Condition (i) materializes only for compactly supported kernels if $h_p \rightarrow 0$ which requires that the bandwidth be determined by the nearest neighbor distance. Condition (ii), on the other hand, requires that $\|\bfs_{k} - \bfz_{p} \|/h_{k} \ll \|\bfs_{i} - \bfz_{p} \|/h_{i}$ for $i \neq k.$ This condition holds approximately at best if the sample is sparse around $\bfz_{p}$. The computational complexity of the SLI predictor is $O(N^2 + P\, N)$. The $O(N^2)$ term is due to the double summation over the sampling points in~\eqref{eq:weights-pred}, which needs to be calculated only once. The remaining operations per each prediction point scale linearly with the sample size, hence the $O(P\, N)$ dependence. Based on the above, the dominant term (for fixed $P$) in the computational time scales as $O(N^2)$. In future work we will investigate approximating the double summation in the denominator of~\eqref{eq:J-weights} and~\eqref{eq:weights-pred-pi} with analytically evaluated double integrals over the kernel functions to increase the computational efficiency. \section{Case Studies} We first consider two synthetic data sets, the first consisting of a time series and the second of a four-dimensional test function. We then investigate a set of scattered real data in two spatial dimensions. \subsection{Time series with Mat\'{e}rn covariance function} We generate a time series of length $N=300$ from a random process with Mat\'{e}rn covariance $C(\tau) = \sigma^2 \, 2^{1-\nu} K_{\nu}(\tau/\xi) (\tau/\xi)^{\nu}/\Gamma(\nu)$, where $K_{\nu}(\cdot)$ is the modified Bessel function of order $\nu$, $\Gamma(\cdot)$ is the gamma function, $\sigma =10$, $\nu = 3.5$ is the smoothness index, and $\xi = 10$ is the correlation time. We use 60 randomly selected points as the training set (corresponding to an $80\ and the remaining 240 points as the validation set. The SLI optimal parameters using a quadratic kernel and $k=2$ are given by $\alpha_{1} \approx 29.30, \alpha_{2} \approx 191.02, \mu \approx 1.11, \lambda \approx 297.84$. The sparseness of the precision matrix is evident in Fig.~. The darkest areas correspond to negative infinity and reflect distances for which the precision matrix vanishes. The prediction performance is illustrated in the scatter plot of the SLI predictions versus the respective validation set values shown in Fig.~. The Pearson correlation coefficient between the validation values and the predictions is $0.89.$ The splitting of the time series into training and validation sets is shown in Fig.~ along with the SLI predictions and associated error bars. The SLI predictions capture well general features of the time series. However, in areas of low sampling density the SLI predictions smoothes excessively the fluctuations in the original series. The SLI performance is excellent for the same degree of thinning, if the length of the initial time series increases to $3\,000$. On the other hand, the prediction accuracy deteriorates for rougher random processes, such as a non-differentiable Mat\'{e}rn process with $\nu = 0.8$. \subsection{Four-dimensional deterministic test function} We consider the function $x(\bfs)$ \begin{equation} x(\bfs) = A\, {\rm e}^{-2 \| \bfs - {\mathbf a}\| } \, \prod_{i=1}^{4} s_{i} \,(1-s_{i}), \end{equation} \noindent where $A = 500$ and $\mathbf{a} = (0.3, 0.3, 0.3, 0.3)$, defined over the four-dimensional cube with unit length edges, i.e., for $\bfs \in [0, 1]^{4}$. We sample the function at $N=1\,000$ randomly selected points over the unit cube, and we generate a validation set of $N=1\,000$ points also by random selection. The SLI optimal parameters for the quadratic kernel with $k=2$ are given by $\alpha_{1} \approx 10.12, \alpha_{2} \approx 25.04, \mu \approx 1.64, \lambda \approx 0.0193$ starting with initial values $\alpha_{1}=10, \alpha_{2}=25, \mu=3$. Similar results in terms of cross validation performance are also obtained with different initial conditions that lead to different local optima. The cross validation measures for the parameters above are given by ME$=0.0046$, MAE$=0.0320$, RMSE$=0.0459$, $r$$=0.96$. The sparse structure of the precision matrix is illustrated in Fig.~ which displays the logarithm of the absolute value. The scatter plot of the validation values versus the respective SLI predictions is shown in Fig.~ and demonstrates very good agreement at most points. We repeat the experiment by adding Gaussian noise to the sample. The standard deviation of the noise is set to $\approx 10\ (in the simulations that we ran $x_{\max} \approx 1$). While the coefficients $\alpha_{1}$ and $\alpha_{2} $ remain practically unchanged, $\mu$ changes to $\approx 1.83$. The sparsity of the precision matrix is $\approx 76\ The respective cross validation measures are given by ME$=0.012$, MAE$=0.047$, RMSE$=0.061$, and $r$$=0.93$. We also used the {\sc Matlab} global optimizer \verb+GlobalSearch+ with the same initial parameter vector as above to determine the SLI model parameters. \verb+GlobalSearch+ uses a scatter-search algorithm to generate starting points (initial parameter guesses). The minimization is conducted using \verb+fmincon+ to determine the local minimum close to the current starting point. We use the lower and upper bounds defined in Section~ to constrain the space of the starting points. \verb+GlobalSearch+ investigates a set of 66 starting points, and convergence to a local minimum is achieved for all of them. The globally optimal SLI parameters are estimated as $\alpha_{1} \approx 1.50, \alpha_{2} \approx 224.62, \mu \approx 2.01, \lambda \approx 0.748$. The respective validation measures are given by ME$=0.0055$, MAE$=0.045$, RMSE$=0.063$, and $r=0.93$. These measures do not differ significantly from those obtained with the locally optimum solution. The MAE value is lower for the global optimum, which is expected since MAE reflects the value of the cost function~\eqref{eq:cost}. On the other hand, the RMSE obtained with the global optimum is slightly higher than that of the local optimum. This result indicates that quite different parameter vectors can lead to similar cross validation results. This behavior has also been observed with covariance models whose parameter vector involves more than the variance and the correlation length~. \subsection{Radioactivity data in two dimensions} This example focuses on daily means of radioactivity gamma dose rates over part of the Federal Republic of Germany. The data were provided by the German automatic radioactivity monitoring network for the Spatial Interpolation Comparison Exercise 2004 (SIC 2004)~. This data set is well studied and thus allows easy comparisons with other methods~. The $1\,008$ stations are partitioned into a \textit{training set} of $200$ randomly selected locations and a \textit{validation set} of $808$ locations where predictions are compared with the observations. Two different scenarios are investigated: A \textit{normal} data set corresponding to typical background radioactivity measurements, and an \textit{emergency} data set, in which a local release of radioactivity in the southwest corner of the monitored area was simulated using a dispersion process to obtain a few values with magnitudes around 10 times higher than the background. The rates are measured in nanoSievert per hour (nSv/h). The normal training set follows the Gaussian distribution with the minimum around 58 nSv/h and the maximum around 153~nSv/h. In the emergency training set there are two values $>1\,000$~nSv/h, with the maximum at $1\,499$~nSv/h. We compare the prediction performance of the SLI model against the $808$ values of the validation set. \subsubsection{Normal data} For normal data, the optimal SLI parameters based on the training set with a quadratic kernel and $k=2$ are given by $\alpha_{1} \approx 143, \alpha_{2} \approx 47.56, \mu \approx 2.64, \lambda \approx 3.24\times 10^3$. Figure~ illustrates the relative values of the bandwidths used. Higher values correspond to more isolated points in areas of low sampling density and along the boundaries of the convex hull of the domain. Figure~ presents the natural logarithm of the absolute value of the precision matrix. Overall, about $ 32 \ the total number of pairs yield nonzero precision values, implying that the sparsity of the precision matrix is $\approx 68\ The cross-validation results are tabulated in Table~. The cross validation measures (based on the validation set) obtained in a recent study by means of \emph{Ordinary Kriging} are: ME$=-1.36$, MAE$=9.29$, RMSE$=12.59$, $r$$=0.78$~. These values are in close agreement with the SLI results in Table~. Various geostatistical and machine learning methods have been applied to the SIC 2004 data (neural networks, geostatistics, and splines). Excluding the results of some poor performers, the cross validation measures obtained are in the following ranges~: {ME} $\in [-1.39 , -0.04]$ and $\in [0.20, 1.60]$, {MAE} $\in [9.05 , 12.10]$, {RMSE} $\in [12.43 , 15.90]$, and $r$ $\in [ 0.64 , 0.79]$. Hence, the SLI cross validation results are close to the best performers. Fig.~ presents a map of the radioactivity pattern generated by SLI and contrasts it with the map generated by bilinear interpolation. The SLI spatial pattern is smoother and thus appears more realistic. Its smoothing effect near the sample values, however, is more pronounced than that caused by bilinear interpolation. \begin{table}[h] \centering \caption{ Cross validation performance measures for SIC 2004 normal data. The second row presents the performance of the SLI predictor at the 808 validation set points. {ME:} Mean error (bias); {MAE:} Mean absolute error; {MARE:} Mean absolute relative error; {RMSE:} Root mean square error; {$r$:} Pearson correlation coefficient.} \begin{tabular}{lccccc} \hline {{SLI}}&{ME}&{MAE}&{MARE}&{RMSE}&{$r$}\\\hline {Validation set}&$-$1.30&9.30&0.09&12.62&0.78\\\hline \end{tabular} \end{table} We also conduct a stability analysis by removing one sampling point at a time and determining the optimal SLI model using leave-one-out cross validation with that point removed. The variation of the SLI parameters is shown in Fig.~; $\alpha_1, \alpha_2$ and $\mu$ are quite stable, whereas $\lambda$ shows more variability. The spikes in the plots of Fig.~ are exaggerated by using a narrow vertical range to better illustrate the parameter variability. For $\alpha_1$, $\alpha_2$ the maximum relative variation (with respect to the mean) ranges from a fraction of a thousandth (for $\alpha_1$) to few thousandths (for $\alpha_2$); $\mu$ shows stronger variations, whereas the strongest variation is exhibited by $\lambda$, since the latter is a scaling factor that determines the overall energy of the ensemble of points and compensates for variations in the other parameters. We believe that the parameter variations exhibited in Fig.~ are, at least partially due to extremely slow variation of the cost function over a region of the parameter space, a condition also observed in maximum likelihood estimation of spatial models with Mat\'{e}rn covariance~. This slow variation implies \emph{quasi-degeneracy} of the parameter vector; the quasi-degeneracy implies that vectors which are very far in parameter space may lead to very similar cost function values. More recently, the difficulties involved in nonlinear fits of multi-parametric models to data have been investigated in~. \subsubsection{Emergency data} For the SIC 2004 emergency data, cross validation results with different kernel functions (tricubic, exponential, and quadratic) are shown in Table~. The SLI parameters are initialized using the optimal values for the normal data. The last row of the table is based on estimation with a quadratic kernel function after removing the three highest values. The best results in Table~ are obtained with the quadratic kernel including all the data. The optimal SLI parameters are $\alpha_{1} \approx 143, \alpha_{2} \approx 47.56, \mu \approx 2.69, \lambda \approx 4.32\times 10^5$. The parameters, except for $\lambda$, are close to their normal case counterparts. The difference in $\lambda$ is due to the much higher variance of the emergency data set. The variation of the SLI parameters in leave-one-out cross validation exhibits similar patterns as for the normal data, except that more pronounced variations of $\lambda$ are observed when the extreme values are removed. The precision matrix has 23\,232 non-zero elements, implying a sparsity of $\approx 42\ in the normal case. This difference clearly illustrates the dependence of the precision matrix on the sample values in addition to the sampling pattern. SLI does not rely on estimating the variogram function, and thus it is not hindered by the presence of extreme values. On the other hand, geostatistical methods rely on the variogram function, which may not be reliably estimated in such cases~. The Pearson correlation coefficient is significantly lower than in the normal set due to underestimation of the extreme values, while the Spearman rank correlation coefficient is comparable to the normal case. The cross validation measures obtained in SIC 2004 are in the following intervals~: {ME} $\in [-11.10 , -0.12]$ and $\in [0.41, 19.71]$, {MAE} $\in [14.85 , 146.36]$, {RMSE} $\in [45.46 , 212.10]$, and $r$ $\in [ 0.02 , 0.86]$. The best performance in terms of both MAE and RMSE was obtained by means of a Generalized Regression Neural Network~. Looking at the scatter plot of MAE versus RMSE ---Fig.~6 in~--- the SLI performance is closer to the geostatistical and spline methods. \begin{table}[h] \centering \caption{ SLI cross validation performance measures for SIC 2004 emergency data. The second row presents the performance of the SLI predictor at the 808 validation set points. The first five cross validation measures are as described in the caption of Table~, and $r_S$ is the Spearman correlation coefficient.} \begin{tabular}{lcccccc} \hline {{SLI} Kernel function}&{ME}&{MAE}&{MARE}&{RMSE}&{$r$}& $r_S$\\\hline {Tricubic}: $(1-u^3)^3$ &5.78&24.22&0.20&81.33&0.25&0.57\\\hline {Exponential}: $ {\mathrm e}^{-u}$ &6.06&23.84&0.19&79.78&0.34&0.63\\\hline {Quadratic}&3.04&23.16&0.17&75.63&0.43 &0.77\\\hline {Quadratic (outliers removed)}&$-$8.28&16.46&0.10&81.41&0.27&0.77 \\\hline \end{tabular} \end{table} \section{Discussion} \subsection{Connections with machine learning} The SLI model is similar to $k$-Nearest Neighbors (KNN), since both methods employ an optimal neighborhood range. In the case of KNN a uniform optimal number of nearest neighbors is determined, and the estimate at an unmeasured point is simply the mean of its $k$ nearest neighbors. In SLI, a locally optimal neighborhood size is determined implying that the number of neighbors used in prediction varies locally. In addition, the estimate is a weighted mean of the neighbor values, in which the weights are determined by the kernel function and the bandwidths. In this respect, SLI is similar to the Nadaraya-Watson kernel regression method~ and to the Support Vector Machine algorithm~. SLI can also be viewed as a particular type of Gaussian process with a sparse inverse covariance kernel, which could be used as an alternative to the sparse Gaussian process framework to improve the computational efficiency of predictions~. In this study we formulated the SLI model using the spatial locations $\Sn$ as inputs and the respective values of the scalar field values as outputs. This framework is appropriate for scattered spatial data. It is possible, however, to use more general input variables instead of the spatial locations, so long as a suitable measure of distance can be defined. \subsection{Notes on implementation } We presented a ``plain vanilla'' version of the SLI model. Modifications that can increase the flexibility but also the complexity of the model are possible. The local kernel bandwidths are determined by fixing the neighbor order $k$ and using a uniform scaling parameter $\mu$. Alternatively, one can consider estimating $k$ from the data and using a locally varying $\mu$. With respect to the latter, potential gains should be weighted against the loss of computational efficiency that will result from the significant increase of the parameter vector size. While our estimate of $\mx$ is based on the sample mean, it is possible to estimate $\mx$ by means of the leave-one-out cross validation procedure. It is also possible to replace $\mx$ with a space-dependent trend function. The present version of the SLI model does not involve anisotropy. Nevertheless, anisotropy is important in cases such as the radioactivity emergency data~: the best performing method in SIC 2004 for this set was a general regression neural network with an anisotropic Gaussian kernel function. Similarly, in SLI it is possible to use weighted Euclidean distances or Minkowski metrics instead of the classical Euclidean distance~. SLI can also be extended to spherical surfaces, a case which is relevant for global geospatial data. In addition, the SLI model can capture correlations in higher-dimensional, abstract feature spaces equipped with a suitable distance. At this point there is no rigorous physical interpretation of the coefficients $\alpha_{1}$, $\alpha_{2}$ and $\lambda$. In general, higher values of $\alpha_{1}$ ($\alpha_{2}$) imply higher cost for gradient (curvature), whereas $\lambda$ controls the overall ``energy''. In the continuum case (i.e., for Spartan random fields) coefficients $\alpha_1$ and $\alpha_2$ are related to a rigidity coefficient and a characteristic length~. A similar correspondence can also be established for data distributed on rectangular grids. In contrast, such relations are not available for scattered data. Even in the continuum and grid cases, however, statistical measures such as the variance and the correlation length have a nonlinear dependence on the SSRF model parameters~. A reasonable initial value for $\mu$ is around $2$--$3$, to allow even compactly supported kernel functions to build local neighborhoods containing at least a few data points. For $\alpha_1$ and $\alpha_2$, we have used positive values between the arbitrary bounds of 0.5 and 300. Exploratory runs with different initial conditions can help to locate a reasonable starting point. Alternatively, a global optimization approach can be used as in Section~. We have opted for a cross-validation cost functional which is based on the mean absolute error. It is possible to use different cost functionals that involve a linear combination of validation measures such as the mean absolute error and the root mean square error. Most results for the case studies investigated above were obtained using an interior-point optimization method that searches for local minima of the cost function. In all of the cases that we have investigated (including data not presented herein), the local optimization led to reasonable cross validation measures which were comparable to those obtained with other methods. As we have shown in the case of $4D$ synthetic data, searching for global optima does not necessarily lead to significant performance improvement. The investigation of global optimization methods with different data sets, however, deserves further attention. \section{Conclusions} The SLI model presented above provides a bridge between geostatistics and machine learning. It is based on an exponential joint density which involves an energy functional with an explicit precision (inverse covariance) matrix. The latter is constructed by superimposing network sub-matrices that implement local interactions between neighboring field values in terms of kernel functions. The algorithmic complexity of SLI missing value estimation scales linearly with the sample size except for a global $O(N^2)$ term which is, however, computed once for all the prediction points. Hence, the leave-one-out cross-validation approach can be used to efficiently infer the SLI model parameters. For missing data on rectangular grids (ongoing research) the computational complexity of the SLI method can be simplified to linear scaling with $N$, because $\mcal{S}_{1}$ and $\mcal{S}_{2}$ can be calculated without kernel functions~. In addition, calculating and storing the large $N\times N$ distance matrix is not necessary in this case. In conclusion, the SLI model is a promising tool for the analysis of big spatial data. In future research we will investigate the extension of the model to space-time data. Finally, the {\sc Matlab} code used for the case studies in Section~ is available at the web address of the Geostatistics laboratory: . \section*{Acknowledgment} The research presented in this manuscript was funded by the project SPARTA 1591: ``Development of Space-Time Random Fields based on Local Interaction Models and Applications in the Processing of Spatiotemporal Datasets''. The project SPARTA is implemented under the ``ARISTEIA'' Action of the operational programme ``Education and Lifelong Learning'' and is co-funded by the European Social Fund and National Resources. \section*{References} \bibliographystyle{Chicago} \begin{thebibliography}{} \bibitem[\protect\citeauthoryear{Addair, Dodge, Walter, and Ruppert}{Addair et~al.}{2014}]{Addair14} Addair, T.~G., D.~A. Dodge, W.~R. Walter, and S.~D. Ruppert (2014). \newblock Large-scale seismic signal analysis with hadoop. \newblock {\em Computers \& Geosciences\/}~{\em 66}, 145--154. \bibitem[\protect\citeauthoryear{Adler}{Adler}{1981}]{Adler81} Adler, R.~J. (1981). \newblock {\em The Geometry of Random Fields}. \newblock New York: Wiley. \bibitem[\protect\citeauthoryear{Ahrens, Hendrickson, Long, Miller, Ross, and Williams}{Ahrens et~al.}{2011}]{Ahrens11} Ahrens, J., B.~Hendrickson, G.~Long, S.~Miller, R.~Ross, and D.~Williams (2011). \newblock Data-intensive science in the {US DOE}: Case studies and future challenges. \newblock {\em Computing in Science \& Engineering\/}~{\em 13\/}(6), 14--24. \bibitem[\protect\citeauthoryear{Atkenson, Moore, and Schaal}{Atkenson et~al.}{1997}]{Atkeson97} Atkenson, S.~G., A.~W. Moore, and S.~Schaal (1997). \newblock Locally weighted learning. \newblock {\em Artificial Intelligence Review\/}~{\em 11\/}(1-5), 11--73. \bibitem[\protect\citeauthoryear{Barndorff-Nielsen}{Barndorff-Nielsen}{2014}]{Barndorff14} Barndorff-Nielsen, O. (2014). \newblock {\em Information and Exponential Families in Statistical Theory}. \newblock West Sussex, UK: John Wiley \& Sons. \bibitem[\protect\citeauthoryear{Besag}{Besag}{1974}]{Besag74} Besag, J. (1974). \newblock Spatial interaction and the statistical analysis of lattice systems. \newblock {\em Journal of the Royal Statistical Society. Series B (Methodological)\/}~{\em 36}, 192--236. \bibitem[\protect\citeauthoryear{Chil{\`e}s and Delfiner}{Chil{\`e}s and Delfiner}{2012}]{Chiles12} Chil{\`e}s, J.~P. and P.~Delfiner (2012). \newblock {\em Geostatistics: Modeling Spatial Uncertainty\/} (2nd ed.). \newblock New York: Wiley. \bibitem[\protect\citeauthoryear{Cressie and Johannesson}{Cressie and Johannesson}{2008}]{Cressie08} Cressie, N. and G.~Johannesson (2008). \newblock Fixed rank kriging for very large spatial data sets. \newblock {\em Journal of the Royal Statistical Society: Series B (Statistical Methodology)\/}~{\em 70\/}(1), 209--226. \bibitem[\protect\citeauthoryear{Csat{\'o} and Opper}{Csat{\'o} and Opper}{2002}]{Csato02} Csat{\'o}, L. and M.~Opper (2002). \newblock Sparse on-line gaussian processes. \newblock {\em Neural Computation\/}~{\em 14\/}(3), 641--668. \bibitem[\protect\citeauthoryear{Du, Zhang, and Mandrekar}{Du et~al.}{2009}]{Du09} Du, J., H.~Zhang, and V.~S. Mandrekar (2009). \newblock Fixed-domain asymptotic properties of tapered maximum likelihood estimators. \newblock {\em The Annals of Statistics\/}~{\em 37\/}(6A), 3330--3361. \bibitem[\protect\citeauthoryear{Dubois}{Dubois}{1998}]{dubois98} Dubois, G. (1998). \newblock Spatial interpolation comparison 97: Foreword and introduction. \newblock {\em Journal of Geographic Information and Decision Analysis\/}~{\em 2\/}(2), 1--10. \bibitem[\protect\citeauthoryear{Dubois and Galmarini}{Dubois and Galmarini}{2006}]{dubois05} Dubois, G. and S.~Galmarini (2006). \newblock Spatial interpolation comparison ({SIC}) 2004: introduction to the exercise and overview of results. \newblock In G.~Dubois (Ed.), {\em Automatic Mapping Algorithms for Routine and Emergency Monitoring}, Volume EUR 21595 EN, pp.\ 7--18. Luxembourg, European Communities: Office for Official Publications of the European Communities. \bibitem[\protect\citeauthoryear{Elogne, Hristopulos, and Varouchakis}{Elogne et~al.}{2008}]{dthetal08} Elogne, S.~N., D.~Hristopulos, and E.~Varouchakis (2008). \newblock An application of {S}partan spatial random fields in environmental mapping: focus on automatic mapping capabilities. \newblock {\em Stochastic Environmental Research and Risk Assessment\/}~{\em 22\/}(5), 633--646. \bibitem[\protect\citeauthoryear{Elogne, Thomas, and Perrin}{Elogne et~al.}{2008}]{Elogneetal07} Elogne, S.~N., C.~Thomas, and O.~Perrin (2008). \newblock Nonparametric estimation of smooth stationary covariance functions by interpolation methods. \newblock {\em Statistical Inference and Stochastic Processes\/}~{\em 11\/}(2), 177--205. \bibitem[\protect\citeauthoryear{Farmer}{Farmer}{2007}]{Farmer07} Farmer, C.~L. (2007). \newblock Bayesian field theory applied to scattered data interpolation and inverse problems. \newblock In A.~Iske and J.~Levesley (Eds.), {\em Algorithms for Approximation}, pp.\ 147--166. Heidelberg: Springer-Verlag. \bibitem[\protect\citeauthoryear{Furrer, Genton, and Nychka}{Furrer et~al.}{2006}]{Furrer06} Furrer, R., M.~G. Genton, and D.~Nychka (2006). \newblock Covariance tapering for interpolation of large spatial datasets. \newblock {\em Journal of Computational and Graphical Statistics\/}~{\em 15\/}(3), 502--523. \bibitem[\protect\citeauthoryear{Garc\'ia-Soid\'an, Febrero-Bande, and Gonz\'alez-Manteiga}{Garc\'ia-Soid\'an et~al.}{2004}]{Garcia04} Garc\'ia-Soid\'an, P.~H., M.~Febrero-Bande, and W.~Gonz\'alez-Manteiga (2004). \newblock Nonparametric kernel estimation of an isotropic semivariogram. \newblock {\em Journal of Statistical Planning and Inference\/}~{\em 121\/}(1), 65--92. \bibitem[\protect\citeauthoryear{Geman and Geman}{Geman and Geman}{1984}]{Geman84} Geman, S. and D.~Geman (1984, Nov). \newblock Stochastic relaxation, {G}ibbs distributions, and the {B}ayesian restoration of images. \newblock {\em IEEE Transactions on Pattern Analysis and Machine Intelligence\/}~{\em PAMI-6\/}(6), 721--741. \bibitem[\protect\citeauthoryear{Giraldi and Bengio}{Giraldi and Bengio}{2006}]{Giralids05} Giraldi, N. and S.~Bengio (2006). \newblock Machine learning for automatic environmental mapping: when and how? \newblock In G.~Dubois (Ed.), {\em Automatic Mapping Algorithms for Routine and Emergency Monitoring}, Volume EUR 21595 EN, pp.\ 123--138. Luxembourg, European Communities: Office for Official Publications of the European Communities. \bibitem[\protect\citeauthoryear{Hall, Fisher, and Hoffman}{Hall et~al.}{1994}]{Hall94} Hall, P., N.~Fisher, and B.~Hoffman (1994). \newblock Properties of nonparametric estimators of autocovariance for stationary random fields. \newblock {\em Annals of Statistics\/}~{\em 22\/}(4), 2115--2134. \bibitem[\protect\citeauthoryear{Hristopulos}{Hristopulos}{2003}]{dth03b} Hristopulos, D. (2003). \newblock Spartan {G}ibbs random field models for geostatistical applications. \newblock {\em SIAM Journal of Scientific Computing\/}~{\em 24\/}(6), 2125--2162. \bibitem[\protect\citeauthoryear{Hristopulos and Elogne}{Hristopulos and Elogne}{2007}]{dthsel07} Hristopulos, D.~T. and S.~Elogne (2007). \newblock Analytic properties and covariance functions of a new class of generalized {G}ibbs random fields. \newblock {\em IEEE Transactions on Information Theory\/}~{\em 53\/}(12), 4667--4679. \bibitem[\protect\citeauthoryear{Hristopulos and Elogne}{Hristopulos and Elogne}{2009}]{dthsel09} Hristopulos, D.~T. and S.~N. Elogne (2009). \newblock Computationally efficient spatial interpolators based on {S}partan spatial random fields. \newblock {\em {IEEE} Transactions on Signal Processing\/}~{\em 57\/}(9), 3475--3487. \bibitem[\protect\citeauthoryear{Hristopulos and \v{Z}ukovi\v{c}}{Hristopulos and \v{Z}ukovi\v{c}}{2011}]{dth11} Hristopulos, D.~T. and M.~\v{Z}ukovi\v{c} (2011). \newblock Relationships between correlation lengths and integral scales for covariance models with more than two parameters. \newblock {\em Stochastic Environmental Research and Risk Assessment\/}~{\em 25\/}(1), 11--19. \bibitem[\protect\citeauthoryear{Kanevski and Maignan}{Kanevski and Maignan}{2004}]{Kanevski04} Kanevski, M. and M.~Maignan (2004). \newblock {\em Analysis and Modelling of Spatial Environmental Data}. \newblock Lausanne, Switzerland: EPFL Press. \bibitem[\protect\citeauthoryear{Kaufman, Schervish, and Nychka}{Kaufman et~al.}{2008}]{Kaufman08} Kaufman, C.~G., M.~J. Schervish, and D.~W. Nychka (2008). \newblock Covariance tapering for likelihood-based estimation in large spatial data sets. \newblock {\em Journal of the American Statistical Association\/}~{\em 103\/}(484), 1545--1555. \bibitem[\protect\citeauthoryear{Lemm}{Lemm}{2005}]{Lemm05} Lemm, J.~C. (2005). \newblock {\em Bayesian Field Theory}. \newblock Baltimore: Johns Hopkins University Press. \bibitem[\protect\citeauthoryear{Lindgren and Rue}{Lindgren and Rue}{2011}]{Rue10} Lindgren, F. and H.~Rue (2011). \newblock An explicit link between {G}aussian fields and {G}aussian {M}arkov random fields: The {SPDE} approach. \newblock {\em Journal of the Royal Statistical Society, Series B\/}~{\em 73\/}(4), 423–498. \bibitem[\protect\citeauthoryear{Nadaraya}{Nadaraya}{1964}]{Nadaraya64} Nadaraya, E.~A. (1964). \newblock On estimating regression. \newblock {\em Theory of Probability and its Applications\/}~{\em 9\/}(1), 141--142. \bibitem[\protect\citeauthoryear{Nychka, Bandyopadhyay, Hammerling, Lindgren, and Sain}{Nychka et~al.}{2014}]{Nychka14} Nychka, D., S.~Bandyopadhyay, D.~Hammerling, F.~Lindgren, and S.~Sain (2014). \newblock A multi-resolution gaussian process model for the analysis of large spatial data sets. \newblock {\em Journal of Computational and Graphical Statistics\/}~(just-accepted), 00--00. \bibitem[\protect\citeauthoryear{Rasmussen and Williams}{Rasmussen and Williams}{2006}]{Rasmussen06} Rasmussen, C.~E. and C.~K.~I. Williams (2006). \newblock {\em {Gaussian} Processes for Machine Learning}. \newblock Boston, MA: MIT Press. \bibitem[\protect\citeauthoryear{Rue and Held}{Rue and Held}{2005}]{Rue05} Rue, H. and L.~Held (2005). \newblock {\em {Gaussian} {M}arkov Random Fields: Theory and Applications}. \newblock Chapman and Hall/CRC. \bibitem[\protect\citeauthoryear{Spiliopoulos, Hristopulos, Petrakis, and Chorti}{Spiliopoulos et~al.}{2011}]{Spil11} Spiliopoulos, I., D.~T. Hristopulos, M.~P. Petrakis, and A.~Chorti (2011). \newblock A multigrid method for the estimation of geometric anisotropy in environmental data from sensor networks. \newblock {\em Computers \& Geosciences\/}~{\em 37\/}(3), 320--330. \bibitem[\protect\citeauthoryear{Steed, Ricciuto, Shipman, Smith, Thornton, Wang, Shi, and Williams}{Steed et~al.}{2013}]{Steed13} Steed, C.~A., D.~M. Ricciuto, G.~Shipman, B.~Smith, P.~E. Thornton, D.~Wang, X.~Shi, and D.~N. Williams (2013). \newblock Big data visual analytics for exploratory earth system simulation analysis. \newblock {\em Computers \& Geosciences\/}~{\em 61}, 71--82. \bibitem[\protect\citeauthoryear{Sun, Li, and Genton}{Sun et~al.}{2012}]{Sun12} Sun, Y., B.~Li, and M.~G. Genton (2012). \newblock Geostatistics for large datasets. \newblock In E.~Porcu, M.~Montero, J, and M.~Schlather (Eds.), {\em Advances and Challenges in Space-time Modelling of Natural Events}, Lecture Notes in Statistics, pp.\ 55--77. Springer Berlin Heidelberg. \bibitem[\protect\citeauthoryear{Timonin and Savelieva.}{Timonin and Savelieva.}{2006}]{Timonin06} Timonin, V. and E.~Savelieva. (2006). \newblock Spatial prediction of radioactivity using general regression neural network. \newblock In G.~Dubois (Ed.), {\em Automatic Mapping Algorithms for Routine and Emergency Monitoring}, Volume EUR 21595 EN, pp.\ 53--54. Luxembourg, European Communities: Office for Official Publications of the European Communities. \bibitem[\protect\citeauthoryear{Transtrum, Machta, and Sethna}{Transtrum et~al.}{2010}]{Transtrum10} Transtrum, M.~K., B.~B. Machta, and J.~P. Sethna (2010). \newblock Why are nonlinear fits to data so challenging? \newblock {\em Physical Review Letters\/}~{\em 104}, 060201. \bibitem[\protect\citeauthoryear{Van~Duijn et~al.}{Van~Duijn et~al.}{2009}]{VanDuijn09} Van~Duijn, M. et~al. (2009). \newblock A framework for the comparison of maximum pseudo-likelihood and maximum likelihood estimation of exponential family random graph models. \newblock {\em Social Networks\/}~{\em 31\/}(1), 52. \bibitem[\protect\citeauthoryear{Vapnik}{Vapnik}{2000}]{Vapnik00} Vapnik, V.~N. (2000). \newblock {\em The Nature of Statistical Learning}. \newblock New York: Springer Verlag. \bibitem[\protect\citeauthoryear{Varin, Reid, and Firth}{Varin et~al.}{2011}]{Varin11} Varin, C., N.~Reid, and D.~Firth (2011). \newblock An overview of composite likelihood methods. \newblock {\em Statistica Sinica\/}~{\em 21\/}(1), 5--42. \bibitem[\protect\citeauthoryear{Watson}{Watson}{1964}]{Watson64} Watson, G.~S. (1964). \newblock Smooth regression analysis. \newblock {\em Sankhya Ser. A\/}~{\em 26\/}(1), 359--372. \bibitem[\protect\citeauthoryear{Winkler}{Winkler}{1995}]{Winkler95} Winkler, G. (1995). \newblock {\em Image Analysis, Random Fields and Dynamic Monte Carlo Methods: A Mathematical Introduction}. \newblock New York: Springer Verlag. \bibitem[\protect\citeauthoryear{Wright}{Wright}{2005}]{Wright05} Wright, M. (2005). \newblock The interior-point revolution in optimization: history, recent developments, and lasting consequences. \newblock {\em Bulletin of the American Mathematical Society\/}~{\em 42\/}(1), 39--56. \bibitem[\protect\citeauthoryear{Wu, Zhu, Wu, and Ding}{Wu et~al.}{2014}]{Wu14} Wu, X., X.~Zhu, G.-Q. Wu, and W.~Ding (2014). \newblock Data mining with big data. \newblock {\em Knowledge and Data Engineering, IEEE Transactions on\/}~{\em 26\/}(1), 97--107. \bibitem[\protect\citeauthoryear{Yaglom}{Yaglom}{1987}]{Yaglom87} Yaglom, A.~M. (1987). \newblock {\em Correlation Theory of Stationary and Related Random Functions {I}}. \newblock New York: Springer Verlag. \bibitem[\protect\citeauthoryear{{\v{Z}}ukovi{\v{c}} and Hristopulos}{{\v{Z}}ukovi{\v{c}} and Hristopulos}{2009}]{zuk09} {\v{Z}}ukovi{\v{c}}, M. and D.~Hristopulos (2009). \newblock The method of normalized correlations: a fast parameter estimation method for random processes and isotropic random fields that focuses on short-range dependence. \newblock {\em Technometrics\/}~{\em 51\/}(2), 173--185. \bibitem[\protect\citeauthoryear{{\v{Z}}ukovi{\v{c}} and Hristopulos}{{\v{Z}}ukovi{\v{c}} and Hristopulos}{2013a}]{zuk13b} {\v{Z}}ukovi{\v{c}}, M. and D.~T. Hristopulos (2013a). \newblock A directional gradient-curvature method for gap filling of gridded environmental spatial data with potentially anisotropic correlations. \newblock {\em Atmospheric Environment\/}~{\em 77}, 901--909. \bibitem[\protect\citeauthoryear{{\v{Z}}ukovi{\v{c}} and Hristopulos}{{\v{Z}}ukovi{\v{c}} and Hristopulos}{2013b}]{zuk13} {\v{Z}}ukovi{\v{c}}, M. and D.~T. Hristopulos (2013b). \newblock Reconstruction of missing data in remote sensing images using conditional stochastic optimization with global geometric constraints. \newblock {\em Stochastic Environmental Research and Risk Assessment\/}~{\em 27\/}(4), 785--806. \end{thebibliography} \begin{appendix} \section{Minimization of NLL} For the pdf given by~(), the log-likelihood is given by \begin{equation} {\mathrm {LL}}(\xsam ;\bftheta) \doteq \ln{L(\xsam ;\bftheta)} = -\Hx(\xsam ;\bftheta) - \ln{Z(\bftheta)}. \end{equation} The partition function in~() is given by the multiple integral \begin{equation} Z(\bftheta) = \prod_{i=1}^{N} \, \int_{-\infty}^{\infty} d\Xo_i \, \exp\left( -\Hx(\xsam ;\bftheta) \right). \end{equation} The square gradient and square curvature terms do not depend on $\mx$ because they involve differences $x_{i}-x_{j}$. Hence, we can express~\eqref{eq:H-using-J} as follows \begin{equation} \Hx (\xsam ;\bftheta) = \frac{1}{2} \xsam ^{T} \, {\mathbf J}(\bftheta) \,\xsam + \frac{\mx^2}{\lambda } - \frac{2\mx \, \smx}{\lambda}, \end{equation} where $\smx$ is the sample mean. Maximizing the ${\mathrm {NLL}}$ with respect to $\mx$, using~\eqref{eq:en-est} for the energy functional, yields \[ \mx = \smx. \] Since this fixes the parameter $\mx$, we can use expression~\eqref{eq:H-using-J} for the energy functional. We apply the scaling transformation $\Hx(\xsam;\bftheta) = \tilde{H}(\xsam;\bftheta_{-\lambda})/\lambda$, where $\bftheta_{-\lambda}$ is the parameter vector except for $\lambda$ and $\tilde{H}(\xsam;\bftheta_{-\lambda})$ is $\lambda$-independent. The transformation $H(\cdot) \mapsto \tilde{H}(\cdot)$ is equivalent to $\Xo_{i} \mapsto y_{i} = (\Xo_{i} -\smx)/\sqrt{\lambda}$. Let us then define the scaled partition function $\tilde{Z}(\bftheta_{-\lambda})$ by means of \begin{align} \tilde{Z}(\bftheta_{-\lambda}) = & \prod_{i=1}^{N} \, \int_{-\infty}^{\infty} dy_i \, \exp\left( -\tilde{H}({\mathbf y};\bftheta_{-\lambda}) \right) \nonumber \\ = & \lambda^{-N/2} \prod_{i=1}^{N} \, \int_{-\infty}^{\infty} d\Xo_i \, \exp\left( -{H}(\xsam;\bftheta) \right) \nonumber \\ = & \lambda^{-N/2} \,{Z}(\bftheta) . \end{align} In light of the above transformations, the dependence of ${\mathrm {NLL}}$ on $\lambda$ takes the following explicit form \[ {\mathrm {NLL}}(\xsam;\bftheta) =\frac{\tilde{H}(\xsam;\bftheta_{-\lambda})}{\lambda} + \frac{N}{2} \ln\lambda + \ln\tilde{Z}(\bftheta_{-\lambda}). \] Hence, by minimizing NLL with respect to $\lambda$, i.e., $\frac{d {\mathrm {NLL}}(\xsam;\bftheta)}{d\lambda} = 0$, we obtain the following expression for the optimal $\lambda$: \begin{align} \lambda^{\ast} = \frac{2\tilde{H}(\xsam;\bftheta_{-\lambda})}{N}. \end{align} From the Gaussian joint pdf~\eqref{eq:H-using-J} it follows that \[ {\tilde{Z}(\bftheta_{-\lambda})} = (2\pi)^{N/2}\, \left\{\det\left[\tilde{J}(\bftheta_{-\lambda})\right] \right\}^{-1/2}, \] where $\tilde{J}(\bftheta_{-\lambda}) = \lambda \, {J}(\bftheta)$. We insert the optimal value $\lambda^{\ast}$ in ${\mathrm {NLL}}$ and use the expression above for the log-partition function which leads to \begin{align} {\mathrm {NLL}}(\xsam;\bftheta_{-\lambda}) = & \frac{N}{2} \ln \left( \frac{2\tilde{H}(\xsam;\bftheta_{-\lambda})}{N} \right) + \frac{N}{2}\ln(2\pi) - \frac{1}{2}\det\left[\tilde{J}(\bftheta_{-\lambda})\right]. \end{align} The NLL~\eqref{eq:NLL} is minimized numerically using the {\sc Matlab} constrained minimization function \verb+fmincon+. Constraints are used to ensure that the parameter values are positive. The log-determinant is calculated numerically using the singular value decomposition of the precision matrix. This is a procedure with numerical complexity $O(N^3)$ for a full rank matrix. For this reason, we use cross validation instead of maximum likelihood for parameter inference. \end{appendix} |
1501.04056 | Title: NLP Solutions as Asymptotic Values of ODE Trajectories
Abstract: In this paper, it is shown that the solutions of general differentiable
constrained optimization problems can be viewed as asymptotic solutions to sets
of Ordinary Differential Equations (ODEs). The construction of the ODE
associated to the optimization problem is based on an exact penalty formulation
in which the weighting parameter dynamics is coordinated with that of the
decision variable so that there is no need to solve a sequence of optimization
problems, instead, a single ODE has to be solved using available efficient
methods. Examples are given in order to illustrate the results. This includes a
novel systematic approach to solve combinatoric optimization problems as well
as fast computation of a class of optimization problems using analogic circuits
leading to fast, parallel and highly scalable solutions.
Body: \title{NLP Solutions as Asymptotic Values of ODE Trajectories} \author{Mazen Alamir \ \\ \ \\ CNRS/Gipsa-lab Control systems department.\\ Email : mazen.alamir@grenoble-inp.fr} \date{} \maketitle \begin{abstract} \noindent In this paper, it is shown that the solutions of general differentiable constrained optimization problems can be viewed as asymptotic solutions to sets of Ordinary Differential Equations (ODEs). The construction of the ODE associated to the optimization problem is based on an exact penalty formulation in which the weighting parameter dynamics is coordinated with that of the decision variable so that there is no need to solve a sequence of optimization problems, instead, a single ODE has to be solved using available efficient methods. Examples are given in order to illustrate the results. This includes a novel systematic approach to solve combinatoric optimization problems as well as fast computation of a class of optimization problems using analogic circuits leading to fast, parallel and highly scalable solutions. \end{abstract} \section{Introduction} Consider the following optimization problem with inequality constraints: \begin{eqnarray} \min_{x\in \mathbb{R}^{n}} f(x) \ \mbox{\rm under}\ c_i(x)\le 0\quad i\in \{1,\dots,n_c\} \end{eqnarray} where $f$ and $c_i$'s are scalar functions of the decision variable $x\in \mathbb{R}^{n}$. Let $f$ be the exact penalty induced function defined by: \begin{eqnarray} \bar f(x,\rho)=f(x)+\rho\cdot \psi(x) \end{eqnarray} where $\psi(x)$ is given by: \begin{eqnarray} \psi(x):=\sum_{i=1}^{n_c}\bigl[\max\{0,c_i(x)\}\bigr]^m\quad ;\quad m\in \mathbb{N} \end{eqnarray} A wide class of algorithms intends to solve () by solving a sequence of unconstrained optimization problems of the form \begin{eqnarray} \min_{x\in \mathbb{R}^n} \bar f(x,\rho_k) \end{eqnarray} for a varying (generally increasing) values of the weighting coefficient $\rho_k$. For each problem in the sequence, only $x$ is searched for while $\rho_k$ is kept constant . The series of unconstrained problems are sometimes replaced a series of problems with by box constraints as descent method with projection are easy to perform. \ \\ \ \\ The increase of $\rho$ is generally defined by $\rho_{k+1}\leftarrow r\times \rho_k$ with $r>1$. The sequence of precision parameter $\omega_k$ (to which the intermediate problems have to be solved) is also made such that $\omega_k\rightarrow 0$ in order to avoid solving with a uselessly high precision intermediate problems. It is then obvious that the efficiency of the resulting algorithms is tightly related to the choice of $r$ and the intermediate precision sequence $\{\omega_k\}_k$ since small values of $r>1$ leads to unnecessarily high number of intermediate problems while a too high values of $r>1$ leads to stiff problems that may lead to slow convergence (because this makes the solution of the box-constraint subproblem harder ) beside the fact that it breaks the {\em continuation} argument that underlies the whole scheme. The choice of $\omega_k$ corresponds to similar trade-offs that need to be carefully handled. Such issues are extensively studied in leading to non necessarily monotonic behavior of $\rho_k$ when solving the sequence of intermediate box constrained modified Lagrangian problems in order to avoid high number of iterations that result when $\rho_k$ is unnecessarily high. This recent study shows at least that monitoring the dynamic evolution of $\rho_k$ is not a trivial issue.\ \\ \ \\ In this paper, it is shown that simultaneous dynamics of $x$ and $\rho$ can be defined through a differential equation of the form: \begin{eqnarray} \dot x=F_1(x,\rho)\ ;\ \dot\rho=F_2(x,\rho) \end{eqnarray} such that solving () gives trajectories that asymptotically converge towards the set of solutions of (). \ \\ \ \\ Note that by doing so, the present paper does not propose a specific alternative algorithm to solve the NLP problem (). Rather, it enables all the efficient algorithms that are available through the huge literature on ODE integration to become candidate algorithms for (). Moreover all the computational background regarding many issues (such as parametric sensitivity , parallel computing , precision monitoring to cite but few items) become available for the constrained optimization problem paradigm. As such, the result of the present paper can be viewed as a starting point for future investigation. More interestingly, it is shown briefly in this paper that expressing the fact that solving () can be done by integrating ODEs enables (for some specific problems) to built analogic circuits that can achieve the task in fast, parallel and massively scalable way. \ \\ \ \\ This paper is organized as follows: first section gives the definitions and notation used throughout the paper. Section states the working assumptions that are needed to derive the main results of the paper. These results are stated and proved in section while section gives some examples of application of the paper results. Finally the paper ends with section that summarizes the contribution and gives hints for further investigations. \subsection{Definition \& Notation} \noindent Throughout the paper, the following notation is used. The $n$-dimensional vectors $f_x(x)$, $\psi_x(x)$ and $\bar f_x(x,\rho)$ denote the gradients of $f$, $\psi$ and $\bar f$ w.r.t $x$. The scalar function $g(x,\rho)$ denotes the euclidian norm of $\bar f_x$ at $(x,\rho)$, namely: \begin{eqnarray} g(x,\rho):=\|\bar f_x(x,\rho)\| \end{eqnarray} For a given weighting coefficient $\rho>0$, the set of stationary points of $\bar f(\cdot,\rho)$ is denoted by $\mathcal S_\rho$, namely: \begin{eqnarray} \mathcal S_\rho:=\bigl\{x\in \mathbb{R}^{n}\ \vert\ g(x,\rho)=0\bigr\} \end{eqnarray} For any $x\in \mathbb{R}^{n}$, the notation $d(x,\rho)$ refer to the distance between $x$ and the set $\mathcal S_\rho$, namely: \begin{eqnarray} d(x,\rho):=\min_{z\in \mathcal S_\rho}\|x-z\| \end{eqnarray} The set of admissible values of $x$ is denoted by: \begin{eqnarray} \mathcal A:=\bigl\{x\in \mathbb{R}^{n}\ \vert\ \psi(x)=0\bigr\} \end{eqnarray} \section{Working Assumptions} \noindent The first assumption states that the optimization problem is well posed in the sense that either the original cost $f(x)$ is already lower bounded or the constraints are such that the weighted cost $\bar f$ is lower bounded:\\ \begin{assumption}{\bf [Well posedness]} For any $\rho>0$, there is a lower bound $\bar f_{\min}(\rho)$ such that $\bar f(x,\rho)\ge \bar f_{min}(\rho)$ for all $x\in \mathbb{R}^{n}$. \end{assumption} \ \\ Note that in the framework of the present paper, it is not assumed that the functions involved are convex. This means that the set $\mathcal S_\rho$ may not be a singleton $\{x^*(\rho)\}$, it is assumed that the norm of the gradient of the weighted cost $\bar f(\cdot,\rho)$ {\bf away} from $\mathcal S_\rho$ can be bounded below by the distance to the set $\mathcal S_\rho$ through some coefficient $k_c$. This leads to the following generalization of the strong convexity assumption:\\ \begin{assumption}{\bf [$\mathcal (S_\rho$)-Strong Convexity]} There is a constant $k_c>0$ such that the following inequality holds: \begin{eqnarray} g(x,\rho)\ge k_c\times d(x,\rho) \end{eqnarray} for all $x\in \mathbb{R}^{n}$. \end{assumption} \ \\ Note that contrary to the classical strong convexity assumption that involves two arbitrary points $x_1$ and $x_2$, the inequality () involves the distance from an arbitrary $x$ to those points lying inside the set of stationary points $\mathcal S_\rho$.\ \\ \ \\ The Next assumption describes a generalized Lypschitz-like assumption on the constraints and the way they are used to construct the exact penalty term $\psi(x)$. \\ \begin{assumption}{\bf [Growth rate of $\psi$]} There is a polynomial $P$ of degree $n_\psi\in \mathbb{N}$ with $P(0)=0$ that satisfies the following inequality \begin{eqnarray} \vert\psi(x_2)-\psi(x_1)\vert \le P(\|x_2-x_1\|)\bigr] \end{eqnarray} for all $x\in \mathbb{R}^{n}$. \end{assumption} \ \\ Note that the use of the polynomial $P$ of the form: \begin{eqnarray} P(d):=\sum_{i=1}^{n_\psi} \alpha_i d^i \end{eqnarray} accounts for the possibility to use different penalty exponents $m$ in the definition of the constraint penalty term in () and the fact that the bounding function may involve lower powers for small distances $d$ and higher powers far from the set $\mathcal A$. \ \\ \ \\ The following assumption is needed to guarantee the existence of solutions to the ODE built up with the functions $\bar f_x$ and $\psi$:\\ \begin{assumption}{\bf [Locally-Lypschitz maps]} For all finite $\rho>0$ the maps $f_x(\cdot,\rho)$, $\psi(\cdot)$ and $\psi_x(\cdot)$ are {\bf locally} Lypschitz. \end{assumption} \ \\ Note that this last assumption expresses {\em local} requirement while () and () are required to hold for any $x$.\ \\ \ \\ The last assumption concerns the relevance of the use of the penalty method to solve (). It states that when the penalty coefficient $\rho$ goes to infinity, the possible stationary points for the weighted cost converge towards the admissible set $\mathcal A$:\\ \begin{assumption}{\bf [Relevance of the penalty approach]} \begin{eqnarray} \lim_{\rho\rightarrow \infty}\Bigl[\ \sup_{x\in \mathcal S_\rho}\psi(x)\ \Bigr]=0 \end{eqnarray} \end{assumption} \ \\ This assumption is almost implicitly required in any penalty-based approach to solve the constrained optimization problem (). It can obviously be replaced by some more apparently trivial assumptions that can be used to prove (). The short form is preferred here for the sake of clarity. \section{Main Results} The main result of the paper can be stated in the following proposition:\\ \begin{proposition}{\bf [Main Result]} Assume that some $(\lambda,q)\in \mathbb{R}_+^*\times \mathbb{N}$ is chosen. Consider the following system of differential equations: \begin{eqnarray} \dot x&=&-\Bigl[\sum_{i=1}^{q}\dfrac{(\lambda\cdot g(x,\rho))^{i-1}}{(i-1)!}\Bigr]\times \bar f_x(x,\rho) \\ \dot \rho &=& \gamma\times \psi(x) \end{eqnarray} {\bf if} the following conditions hold: \begin{enumerate} \item Assumptions - are satisfied \item $q\ge n_\psi$ [see ()] \end{enumerate} {\bf then} for any $\lambda>0$, there is a sufficiently small $\gamma>0$ such that any asymptotic solution of ()-() satisfies the KKT necessary conditions of optimality for the constrained optimization problem (). $\hfill \heartsuit$ \end{proposition} \ \\ {\sc Proof}. Let us compute the derivative of the weighted cost $\bar f(x,\rho)$: \begin{eqnarray} \dfrac{d\bar f}{dt}&=&\gamma\psi^2(x)-\Bigl[\sum_{i=1}^{q}\dfrac{(\lambda\cdot g(x,\rho))^{i-1}}{(i-1)!}\Bigr]^2\times g^2(x,\rho) \nonumber \\ &=&\gamma\psi^2(x)-\Bigl[\sum_{i=1}^{q}\dfrac{(\lambda\cdot g(x,\rho))^{i}}{\lambda\cdot (i-1)!}\Bigr]^2 \end{eqnarray} Let $x_s(\rho)\in \mathcal S_\rho$ be the closest point to $x$ that lies inside the stationary set $\mathcal S_\rho$. According to () of Assumption , one can write: \begin{eqnarray} \psi(x)&\le& \psi(x_s(\rho))+P(\|x-x_s(\rho)\|)\\ &\le& \psi(x_s(\rho))+P(d(x,\rho)) \end{eqnarray} Now by virtue of Assumption , $\psi(x_s(\rho))$ satisfies the following asymptotic property: \begin{eqnarray} \psi(x_s(\rho))=O(1/\rho) \end{eqnarray} Therefore, () becomes [using ()] : \begin{eqnarray} \psi(x)&\le& P(d(x,\rho))+O(1/\rho)\\ &\le& \Bigl[\sum_{i=1}^{n_\psi}\alpha_id^i(x,\rho)\Bigr]+O(1/\rho) \end{eqnarray} On the other hand, the sum in the r.h.s of () satisfies [because of ()] the following inequality: \begin{eqnarray} \Bigl[\sum_{i=1}^{q}\dfrac{(\lambda\cdot g(x,\rho))^{i}}{\lambda\cdot(i-1)!}\Bigr]^2\ge \Bigl[\sum_{i=1}^{q}\dfrac{(\lambda\cdot k_c\cdot d(x,\rho))^{i}}{\lambda\cdot(i-1)!}\Bigr]^2 \end{eqnarray} Now using () and () in () enables to write [dropping all the terms with indices higher than $n_\psi\le q$ in the summing term of ()] and using the identity $y_1^2-y_2^2=(y_1+y_2)(y_2-y_1)$: \begin{eqnarray} \dfrac{d\bar f}{dt}\le \Bigl[\sum_{i=1}^{n_\psi}\beta_i^+d^i+O(1/\rho)\Bigr]\cdot \Bigl[\sum_{i=1}^{n_\psi}\beta_i^-d^i+O(1/\rho)\Bigr] \end{eqnarray} where $d^i:=d^i(x,\rho)$ while $\beta_i^+$ and $\beta_i^-$ are given by: \begin{eqnarray} \beta_i^+&=&\sqrt{\gamma}\alpha_i+ \dfrac{(\lambda k_c)^i}{\lambda\cdot(i-1)!} \\ \beta_i^-&=&\sqrt{\gamma}\alpha_i- \dfrac{(\lambda k_c)^i}{\lambda\cdot(i-1)!} \end{eqnarray} and taking $\gamma$ sufficiently small so as to satisfy the following inequality: \begin{eqnarray} \sqrt{\gamma}\le \min_{i=1}^{n_\psi}\left[\dfrac{(\lambda k_c)^i}{2\alpha_i(\lambda\cdot(i-1)!)}\right] \end{eqnarray} the following inequalities hold for $\beta_i^+$ and $\beta_i^-$: \begin{eqnarray} \beta_i^+\ge \dfrac{(\lambda k_c)^i}{\lambda\cdot(i-1)!}\quad;\quad \beta_i^-\le -\dfrac{(\lambda k_c)^i}{2\lambda\cdot\alpha_i((i-1)!)} \end{eqnarray} With these inequalities, inequality () implies: \begin{eqnarray} \dfrac{d\bar f}{dt}\le -\Bigl[\sum_{i=1}^{n_\psi}\dfrac{(\lambda k_c)^i}{(\lambda\cdot(i-1)!)}d^i(x,\rho)+{\relsize{-2}O(1/\rho)}\Bigr]\times\nonumber \\ \Bigl[\sum_{i=1}^{n_\psi}\dfrac{(\lambda k_c)^i}{2(\lambda\cdot(i-1)!)}d^i(x,\rho)+{\relsize{-2}O(1/\rho)}\Bigr] \end{eqnarray} Let us now show that inequality () implies that $\lim_{t\rightarrow \infty}\psi(x(t))=0$. Indeed, if this was not the case, then by the very definition of the dynamic on $\rho$ [see ()] it comes that $\rho$ goes to infinity. This together with () and the lower boundedness of $\bar f$ [Assumption ] implies that $\lim_{t\rightarrow \infty} d(x,\rho)$=0 ($x$ converges to the set $\mathcal S_\rho$). But this implies by () of assumption that $\psi(x)$ converges to $0$ which contradicts the assumption. Now since $\psi(x)$ converges to $0$, the inequality () together with the lower boundedness of $\bar f$ implies also that $g(x,\rho)$ converges to $0$.\ \\ \ \\ By now it has been shown that provided that $\gamma$ is sufficiently small to satisfy (), the trajectory of $(x,\rho)$ converges to the following set \begin{eqnarray} \Bigl\{(x,\rho)\quad \vert\quad g(x,\rho)=0 \ \mbox{\rm and} \ \psi(x)=0\Bigr\} \end{eqnarray} It remains to prove that if $(x,\rho)$ belongs to the set defined by (), then $x$ satisfies the KKT necessary conditions of optimality. Remember that these conditions require the existence of a vector $\mu\in \mathbb{R}^{n_c}$ such that the following conditions hold: \begin{eqnarray} \begin{array}{ll} f_x(x)+\sum_{i=1}^{n_c}\mu_i\dfrac{\partial c_i}{\partial x}(x)=0 \cr c_i(x)\le 0 &(\forall i\in \{1,\dots,n_c\})\cr \mu_i\ge 0 &(\forall i\in \{1,\dots,n_c\})\cr \mu_i\times c_i(x)=0&(\forall i\in \{1,\dots,n_c\}) \end{array} \end{eqnarray} But $g(x,\rho)=0$ can be explicitly written as follows: \begin{eqnarray} f_x(x)+\rho\times m\sum_{i=1}^{n_c}\left[\max\{0,c_i(x)\}\right]^{m-1}\times \dfrac{\partial c_i}{\partial x}(x) \end{eqnarray} which obviously shows that by taking $\mu$ such that: \begin{eqnarray} \mu_i:= \left\{ \begin{array}{ll} 0& \mbox{\rm if $c_i(x)<0$}\\ \rho\times m\times [c_i(x)]^{m-1}& \mbox{if $c_i(x)\ge 0$} \end{array} \right. \end{eqnarray} the first KKT condition is satisfied by construction. The second condition ($c_i(x)\le 0$) results from $\psi(x)=0$. The third and the fifth conditions ($\mu_i\ge 0$ and $\mu_i\cdot c_i(x)=0$) result from (). This ends the proof. $\hfill \Box$ \ \\ \ \\ Note that if the summation in () is performed with an infinite number of terms, the following corollary can be obtained:\\ \begin{corollary} Assume that some $\lambda>0$ is chosen. Consider the following system of differential equation: \begin{eqnarray} \dot x&=&-\exp\left[{\lambda\cdot g(x,\rho)}\right]\times \bar f_x(x,\rho) \\ \dot\rho&=&\gamma\times \psi(x) \end{eqnarray} {\bf If} Assumptions - hold then for sufficiently small $\gamma>0$, any asymptotic solution of ()-() satisfies the necessary conditions of optimality for the constrained optimization problem (). $\hfill \heartsuit$ \end{corollary} \ \\ Note that the result of Proposition holds for any initial condition that can be used to initialize the trajectory of ()-(). The price to obtain such a global result lies in the use of the $q$-term summation that premultiplies the gradient term $-\bar f_x(x,\rho)$ in (). The next proposition gives a weaker result that can nevertheless be preferable in some circumstances. In this weaker result, the convenient sufficiently small $\gamma$ would depend on the initial values of $x$ and $\rho$. \ \\ \begin{proposition}{\bf [A Simpler Weaker Result]} Assume that some $\lambda>0$ is chosen. Consider the following system of differential equations: \begin{eqnarray} \dot x&=&-\bar f_x(x,\rho) \\ \dot \rho&=&\gamma \times \psi(x) \end{eqnarray} {\bf If} the following conditions hold: \begin{enumerate} \item Assumptions - are satisfied, \item $f(\cdot)$ is proper (that is $\lim_{\|x\|\rightarrow \infty}f(x)=\infty$) \end{enumerate} then for any initialization $(x_0,\rho_0)$, there is sufficiently small $\gamma>0$ such that the resulting asymptotic solution of ()-() satisfies the KKT necessary conditions of optimality for the optimization problem (). $\hfill \heartsuit$ \end{proposition} \ \\ {\sc Proof}. Note that the result can be obtained if one can show that everything behaves as if $n_\psi=1$ holds in (). Indeed, in this case $q=n_\psi=1$ can be used and () is equivalent to (). This can be done using classical arguments that are typically used to derive semi-global results. More precisely, given the initial state $(x_0,\rho_0)$, define the following level set in $\mathbb{R}^{n}$: \begin{eqnarray} \mathcal V(x_0,\rho_0):=\Bigl\{x\ \vert \ f(x)\le 2\bar f(x_0,\rho_0)\Bigr\} \end{eqnarray} to which the initial value $x_0$ obviously belongs [because $f(x_0)\le \bar f(x_0,\rho)$ for all $\rho$]. Note that since $f$ is proper by assumption, the set $\mathcal V(x_0,\rho_0)$ is a compact set. Consequently, there is some sufficiently high $\bar\alpha_1$ such that the following inequality holds for all $(x_1,x_2)\in \mathcal V(x_0,\rho_0)$: \begin{eqnarray} \|\psi(x_2)-\psi(x_1)\|\le P(\|x_2-x_1\|)\le \bar\alpha_1\|x_2-x_1\| \end{eqnarray} this means that as far as the trajectory remains in $\mathcal V(x_0,\rho_0)$, the result of Proposition can be used with $n_\psi=1$ therefore, there exists sufficiently small $\gamma$ such that the dynamics defined by () with $q=1$ [which is the same as ()] decreases the value of $\bar f(x,\rho)$. But this guarantees that the trajectory of $x$ remains in $\mathcal V(x_0,\rho_0)$. This implies that the inequality () remains true and the result obviously follows. $\hfill \Box$ \ \\ \ \\ Note that such finite $\bar\alpha_1>0$ exists as long as the last inequality in () is required only on the compact set $\mathcal V(x_0,\rho_0)$. The latter is defined in terms of the initial paire $(x_0,\rho_0)$. This is why the value of $\bar\alpha_1$ does depend on the initialization and may not exit globally. \subsection{General Comments} \noindent Before getting to the examples section, it is worth mentioning that the results of the present section build a theoretical bridge between NLP and ODE algorithms in a rather systematic way and for a large class of problems. However, it must be underlined that although the following examples show rather efficient computational results, the integration of the resulting ODE may not be the more efficient way to solve the underlying optimization problems. This is because integration schemes try to reproduce high precision solution over the whole trajectories while from the NLP solution point of view, only the asymptotic trajectory matters. \ \\ \ \\ To this respect, the results of the present section can be used to derive gradient-based algorithms (fast gradient for instance ) using the r.h.s of the ODE as extended gradient in the extended space of $(x,\rho)$ with $\bar f$ as cost function. By doing so, even certification results similar to the one proposed in can be extended from the case where only saturations on the control input is used to the more general case of affine constraints on the state. This being said, no such efficiency-oriented development is done here focusing on the main theoretical contribution of the paper. \ \\ \ \\ On the other hand, another consequence of the theoretical result of the present section is the possibility to built electronic circuits that realize analogic ultra-fast integration of the ODE for a class of NLPs. This is briefly discussed in section . \section{Illustrative Examples} \subsection{Example 1: QP problems} \noindent As a first examples let us consider the use of the ODE framework described in Proposition to solve Quadratic Programming (QP) problem with inequality constraints. This leads to the following instantiation of the cost function $f(x)$ and the constraints $c_i(x)$: \begin{eqnarray} f(x)&=&\dfrac{1}{2}x^THx+F^Tx \\ c_i(x)&=&A_ix-B_i \quad ;\quad i=1,\dots,n_c \end{eqnarray} where $H\in \mathbb{R}^{n\times n}$, $F\in \mathbb{R}^{n\times 1}$, $A_i\in \mathbb{R}^{1\times n}$ and $B_i\in \mathbb{R}$. Now using $m=2$ to define the constraints-related weighting term: \begin{eqnarray} \psi(x)=\sum_{i=1}^{n_c}\bigl[\max\{0,A_ix-B_i\}\bigr]^2 \end{eqnarray} gives $n_\psi=2$ [see () and ()]. Consequently, following Proposition , the following ODE is defined (taking $q=2$): \begin{eqnarray} \dot x&=&-\Bigl[1+\lambda\cdot \|\bar f_x(x,\rho)\|\Bigr]\times \bar f_x(x,\rho) \\ \dot \rho&=&\gamma \times \sum_{i=1}^{n_c}\bigl[\max\{0,A_ix-B_i\}\bigr]^2 \end{eqnarray} where \begin{eqnarray*} \bar f_x(x,\rho)= Hx+F+\rho\sum_{i=1}^{n_c}\max\{0,A_ix-B_i\}]\times A_i^T \end{eqnarray*} This ODE is then integrated using the Matlab ODE15s stiff solver to get the solution of the original QP problem defined by ()-(). Fifty randomly generated sets of matrices $\{H,F,A,B\}$ are generated leading to $50$ feasible QPs with $n=15$ unknown and $n_c=20$ constraints. The resulting ODEs ()-() are defined with the parameters $\lambda=10^{-4}$ and $\gamma=10^{-6}$. Figure shows the resulting trajectories of $\psi$ and the cost function normalized by the optimal cost value (computed using the standard Matlab QuadProg solver). All the trajectories are started from $x_0=0$ and $\rho=0$. The Figure clearly shows that the trajectories converge to the solutions of the problems as the constraints are satisfied and the cost function values converge toward the optimal values for all the generated problems. Figure shows a typical behavior of the system's trajectory starting from $(0,0)$ and converging towards the optimal values. The computation times shows a mean of $49\ ms$ with a variance of $3\ ms$ (Using Matlab on Mac PowerBook OSX, 2.8 GHz Intel Core i7 processor). \subsection{Example 2: Analogic MPC Solvers} \noindent Recall that Model Predictive Control (MPC) for linear time invariant systems of the form: \begin{eqnarray} \dot \xi=A\xi+Bu \end{eqnarray} is based on the repetitive solution of a quadratic programming problem of the form: \begin{eqnarray} \min_{x\in \mathbb{R}^{n}} \Bigl[\dfrac{1}{2}x^THx+\bigl[f_0+F_1\xi\Bigr]^Tx\Bigr] \end{eqnarray} under the constraint: \begin{eqnarray} A_ix-\left[b_i^0+B_i\xi\right]\le 0\quad i\in \{1,\dots,n_c\} \end{eqnarray} where $x\in \mathbb{R}^{n}$ is the parameter vector that defines the control trajectory over the prediction horizon, namely: \begin{eqnarray} \begin{pmatrix} u(k)\cr \vdots\cr u(k+N-1) \end{pmatrix}=\Pi\cdot x \end{eqnarray} for some appropriately chosen parametrization matrix $\Pi\in \mathbb{R}^{(Nn_u)\times n}$ where $n_u$ is the dimension of the control input $u$. Note that the only difference between ()-() and ()-() is that the affine term in () and the r.h.s of the inequalities () depends on the state of the controlled system $\xi$. For more details on MPC design, the reader can refer to \ \\ \ \\ Now applying Corollary to the QP defined by ()-() with a sufficiently small $\gamma$ for all initial conditions of interest, it comes that the QP solution (for a given $\xi$) can be obtained by integrating the following set of ODEs: \begin{eqnarray} \dot x&=&-\Bigl[Hx+f_0+F\xi\Bigr]+\nonumber \\ &+&2\rho\sum_{i=1}^{n_c}\Bigl[\max\bigl\{0,A_ix-b_i^0-B_i\xi\bigr\}\Bigr]\cdot A_i^T \\ \dot \rho&=&\gamma \sum_{i=1}^{n_c}\Bigl[\max\bigl\{0,A_ix-b_i^0-B_i\xi\bigr\}\Bigr]^2 \end{eqnarray} The idea is then to perform the integration through analogic circuits. The literature is very rich regarding the way transfer functions and more generally nonlinear differential relationships can be realized by analogic circuits (see and the references therein). Let us concentrate on the operations involved in ()-() to check that analogic realizations can be derived considering that $x$ is represented by a vector of currents (voltage options is also possible although it is not discussed here): \begin{itemize} \item {\em Constant current sources} Note first of all that the constant terms $f_0$ and $b_i^0$ corresponds to tunable source of currents. \item {\em State dependent current sources} The state dependent terms $F\xi$ and $B_i\xi$ are computed numerically and the results is assigned to another vector of current sources that remain constant during the integration step. This is the only numerical operation which determines the sampling rate of the resulting MPC controller. \item {\em Linear combination of currents}. This concerns the terms $Hx$ and $A_ix$ and can be realized for instance using unity gain cells as shown for instance in . \item {\em Current summation and substraction}. This concerns the realization of the sums $Hx+f_0+F\xi$ and $A_ix-b_i^0-B_i\xi$ and can be viewed as a particular instantiation of the previous item and can therefore be realized using unity gain cells. \item {\em Squaring signals}. This is necessary to compute the summated terms in () and can be achieved for instance using the circuits proposed in or any later work containing more recent devices and architectures. \item {\em Multiplication by $\rho$}. This operation can be realized through tunable gain or by using standard signal multipliers as the on proposed in . \end{itemize} \ \\ Note that the time needed to analogically integrate ()-() is the time necessary to fill the corresponding circuit's capacitors. This time can be made extremely short (nano or even pico-seconds) if the problem is appropriately normalized so as to have its normalized solution components $\bar x_i$ scaled down so that they correspond to tiny capacitor voltages. \ \\ \ \\ Note that in the above presentation, the linear character of the controlled system plays no determinant role. Indeed, thanks to the possibility of signal multiplication, squaring and even the possibility to implement the square rooting of signals , a wide class of ODE's that would be associated to the solution of a wide class of non quadratic constrained NLP can be analogically solved in extremely fast way. Moreover, the potential use of massively integrated circuit makes it possible to solve large scale problems in this way.\ \\ \ \\ Note finally that many of the above mentioned circuits can be realized using on-line assignable gains which makes it potentially possible to use the same circuits for many different problems. It remains however necessary to analyse the cost of such circuit design and realization which is beyond the scope of the present paper that studied the conceptual opportunities that are made possible by the ODE-related formulation of constrained optimization problems. \subsection{Example 3: Solving Nonlinear Mixed-Integer Optimization Problems} \noindent In this section, presentation is done for the special case where all the decision variables are binary. The case where some decision variable can be continuous can be obtained easily with extra notational complexity. Consider the optimization problem given by: \begin{eqnarray} \min_{x\in \mathbb{R}^n} f(x)\quad \mbox{\rm under $\bar c_i(x)\le 0$ and $x_i\in \{0,1\}$} \end{eqnarray} for all $i\in \{1,\dots,\bar n_c\}$. \ \\ \ \\ It is well known that this problem can be put in the standard form () by transforming the binary constraints $x_i\in \{0,1\}$ into standard constraints of the form \begin{eqnarray} x_i-x_i^2\le 0\quad -x_i\le 0\quad \mbox{\rm and}\quad x_i-1\le 0 \end{eqnarray} which yields a number of inequality constraints $n_c=\bar n_c+3n$. Moreover, the integer $n_\psi$ that characterizes the growth of $\psi$ [see () and () is given by $n_\psi=2m$ where $m$ is the exponent used in the definition () of $\psi$.\ \\ \ \\ Using $m=2$ leads to the following definition of $\psi(x)$: \begin{eqnarray*} \psi(x)&=&\sum_{i=1}^{n_c}\Bigl[\max\{0,c_i(x)\}\Bigr]^2+\sum_{i=1}^{n}\Bigl[\max\{0,x_i-x_i^2\}\Bigr]^2\\ &+&\sum_{i=1}^{n}\Bigl[\max\{0,-x_i\}\Bigr]^2+\sum_{i=1}^{n}\Bigl[\max\{0,x_i-1\}\Bigr]^2 \end{eqnarray*} Now applying the result of Proposition suggests the solution of the combinatoric optimization problem can be done by integrating the following ODE: \begin{eqnarray} \dot x&=&-\Bigl[\sum_{i=1}^{4}\dfrac{(\lambda\cdot g(x,\rho))^{i-1}}{(i-1)!}\Bigr]\times \bar f_x(x,\rho) \\ \dot \rho &=& \gamma\times \psi(x) \end{eqnarray} Now obviously the admissible set is not a convex set and the presence of local minima is very likely. The following algorithm can be used to visit such local minima successively. In this algorithm a successively modified cost function $f^{(s)}(\cdot)$ where $f^{(0)}\equiv f$ is initialized to the original cost and where $s$ denotes the number of already visited local minima. The local minimum $x^{(s)}$ is found by integrating the ODE defined by the weighted function $\bar f^{(s)}(x,\rho)$ and its corresponding norm of the gradient $g^{(s)}(x,\rho):=\bar f_x^{(s)}(x,\rho)$, namely: \begin{eqnarray} \dot x&=&-\Bigl[\sum_{i=1}^{4}\dfrac{(\lambda\cdot g^{(s)}(x,\rho))^{i-1}}{(i-1)!}\Bigr]\times \bar f_x^{(s)}(x,\rho) \\ \dot \rho &=& \gamma\times \psi(x) \end{eqnarray} This is done starting from the initial condition $(x^{(s-1)},0)$ and integrating the ODE until some stopping conditions on both $\psi$ and $g$ are satisfied. Then a term is added to the cost function which makes the current solution $x^{(s)}$ inappropriate. This can be done by first defining a neighbor vector $z^{(s)}$ to $x^{(s)}$ such that: \begin{eqnarray} \|z^{(s)}-x^{(s)}\|_\infty=1\quad \mbox{\rm and}\quad c(z^{(i)})\le 0 \end{eqnarray} The new cost function $f^{(s+1)}$ is now defined by: \begin{eqnarray*} f^{(s+1)}(x):=f^{(s)}(x)+(1+2f^{(s)}(z^{(s)}))\cdot \exp(\dfrac{\mu}{4}\|x-x^{(s)}\|^2) \end{eqnarray*} Now for sufficiently high $\mu$, this new cost function is such that $x^{(s)}$ is no more a local minimum since $$\bar f^{(s+1)}(x^{(s)})=f^{(s)}(x^{(s)})+f^{(s)}(z^{(s)})+1>f^{(s)}(z^{(s)})$$ therefore, incrementing $s$ and firing the integration of the new resulting ODE ()-() starting from the initial condition $(x^{(s)},0)$ leads to a necessarily different minimum and so on. \ \\ \ \\ The only assumption that is implicitly assumed is that there always exists a neighbor vector $z^{(i)}$ that is admissible in the sense of (). If this is not satisfied, less close $z^{(s)}$ can be searched provided that a deterministic generation process is defined. \section{Conclusion and Future Work} \noindent In this paper, it is shown that the solution of optimization problems with inequality constraints can be obtained by solving appropriately defined ODEs. In these ODEs, simultaneous dynamics are given to the decision variable as well as to the weight associated to the exact penalty term on the constraint violation. One of the major impacts of this result lies in the possibility to design analogic circuits that can quickly and physically integrate the corresponding ODEs. Pushing this latter idea towards a concrete realization is the obvious follow up of the present work. Another direction is to use the result to derive a fast gradient algorithm together with its associated certification bounds regarding the number of iterations that would be necessary to achieve a prescribed level of precision following the steps of while including affine constraints that are not considered in . This was not possible precisely because when standard fast gradient is used, only projection on the box-like set can be done while guaranteeing the decrease of the cost function. The formulation proposed in the present paper provide generalization of this property to an extended monotonically decreasing cost function provided that the r.h.s of the ODE is used as an extended gradient. \bibliographystyle{plain} \bibliography{mybibfile} |
1501.04057 | Title: Measured geodesic laminations in Flatland
Abstract: Since their introduction by Thurston, measured geodesic laminations on
hyperbolic surfaces occur in many contexts. In this survey, we give a
generalization of geodesic laminations on surfaces endowed with a
half-translation structure, called flat laminations, and we define transverse
measures on flat laminations similar to transverse measures on hyperbolic
laminations, taking into account that the images of the leaves of a flat
lamination are in general not pairwise disjoint. One aim is to construct a tool
that could allow a fine description of the space of degenerations of
half-translation structures on a surface. We define a topology on the set of
measured flat laminations and a natural continuous projection of the space of
measured flat laminations onto the space of measured hyperbolic laminations,
for any arbitrary half-translation structure and hyperbolic metric on a
surface. We prove in particular that the space of measured flat laminations is
projectively compact. The main result of this survey is a classification
theorem of (measured) flat laminations on a compact surface endowed with a
half-translation structure. We also give an exposition of that every finite
metric fat graph, outside four homeomorphisms classes, is the support of
uncountably many measured flat laminations with uncountably many leaves none of
which is eventually periodic, and that the space of measured flat laminations
is separable and projectively compact.
Body: \usepackage{amssymb,epsfig,amsmath} \usepackage[dvips]{color} \usepackage[T1]{fontenc} \usepackage[english]{babel} \usepackage[utf8]{inputenc} \usepackage{comment} \usepackage{hyperref} \usepackage{graphicx} \usepackage{multicol} \usepackage{caption} \captionsetup{ font=small, labelfont=bf, tableposition=bottom } \renewcommand{\labelenumi}{{\rm (\arabic{enumi})}} \pagestyle{plain} \textwidth 15cm \textheight 22.5 cm \oddsidemargin 0.5cm \evensidemargin 0.5cm \topmargin 30pt \headheight 0pt \headsep 0pt \footskip 20pt \newtheorem{defi}{Definition} \newtheorem{prop}[defi]{Proposition} \newtheorem{theo}[defi]{Theorem} \newtheorem{conj}[defi]{Conjecture} \newtheorem{lemm}[defi]{Lemma} \newtheorem{coro}[defi]{Corollary} \newtheorem{rema}[defi]{Remark} \newtheorem{exem}[defi]{Example} \newtheorem{exems}[defi]{Examples} \newcommand{\bdefi}{\begin{defi}} \newcommand{\edefi}{\end{defi}} \newcommand{\bprop}{\begin{prop}} \newcommand{\eprop}{\end{prop}} \newcommand{\btheo}{\begin{theo}} \newcommand{\etheo}{\end{theo}} \newcommand{\blemm}{\begin{lemm}} \newcommand{\brema}{\begin{rema}} \newcommand{\erema}{\end{rema}} \newcommand{\bexer}{\begin{exem}} \newcommand{\eexer}{\end{exem}} \newcommand{\bexems}{\begin{exems}} \newcommand{\eexems}{\end{exems}} \newcommand{\bconj}{\begin{conj}} \newcommand{\econj}{\end{conj}} \newcommand{\elemm}{\end{lemm}} \newcommand{\bcoro}{\begin{coro}} \newcommand{\ecoro}{\end{coro}} \newcommand{\dem}{\noindent{\bf Proof. }} \newcommand{\rem}{\noindent{\bf Remark. }} \newcommand{\exemp}{\medskip\noindent{\bf Example. }} \usepackage{mathrsfs} \renewcommand\mathcal{\mathscr} \newcommand{\A}{{\cal A}} \newcommand{\T}{{\cal T}} \newcommand{\B}{{\cal B}} \renewcommand{\L}{{\cal L}} \newcommand{\M}{{\cal M}} \newcommand{\N}{{\cal N}} \newcommand{\G}{{\cal G}} \newcommand{\D}{{\cal D}} \newcommand{\E}{{\cal E}} \newcommand{\F}{{\cal F}} \newcommand{\V}{{\cal V}} \newcommand{\W}{{\cal W}} \newcommand{\K}{{\cal K}} \renewcommand{\H}{{\cal H}} \newcommand{\OOO}{{\cal O}} \newcommand{\C}{{\cal C}} \newcommand{\I}{{\cal I}} \newcommand{\SSSS}{{\cal S}} \newcommand{\Scal}{{\cal S}} \newcommand{\Q}{{\cal Q}} \renewcommand{\P}{{\cal P}} \newcommand{\maths}[1]{{\mathbb #1}} \renewcommand{\AA}{\maths{A}} \newcommand{\OO}{\maths{O}} \newcommand{\RR}{\maths{R}} \newcommand{\NN}{\maths{N}} \newcommand{\CC}{\maths{C}} \newcommand{\QQ}{\maths{Q}} \newcommand{\BB}{\maths{B}} \newcommand{\SSS}{\maths{S}} \newcommand{\DD}{\maths{D}} \newcommand{\HHHH}{\cal{H}} \newcommand{\FF}{\maths{F}} \newcommand{\KK}{\maths{K}} \newcommand{\ZZ}{\maths{Z}} \newcommand{\PPPP}{\maths{P}} \newcommand{\XX}{\maths{X}} \newcommand{\LLLL}{\cal{L}} \newcommand{\TT}{\maths{T}} \newcommand{\UU}{\maths{U}} \newcommand{\aaa}{{\mathfrak a}} \renewcommand{\ggg}{{\mathfrak g}} \newcommand{\sss}{{\mathfrak s}} \renewcommand{\lll}{{\mathfrak l}} \newcommand{\ooo}{{\mathfrak o}} \newcommand{\uuu}{{\mathfrak u}} \newcommand{\vvv}{{\mathfrak v}} \newcommand{\ppp}{{\mathfrak p}} \newcommand{\hhh}{{\mathfrak h}} \newcommand{\fff}{{\mathfrak f}} \newcommand{\zzz}{{\mathfrak z}} \newcommand{\CCC}{{\mathfrak C}} \newcommand{\mmm}{{\mathfrak m}} \renewcommand{\uuu}{{\mathfrak u}} \newcommand{\noted}[1]{\marginpar{\scriptsize #1}} \newcommand{\weakstar}{\overset{*}\rightharpoonup} \newcommand{\ra}{\rightarrow} \newcommand{\bs}{\backslash} \newcommand{\ov}[1]{{\overline #1}} \newcommand{\wt}[1]{{\widetilde{#1}}} \newcommand{\wh}[1]{{\widehat{#1}}} \newcommand{\ga}{\gamma} \newcommand{\Ga}{\Gamma} \newcommand{\semidirect}{{\Bbb n}} \newcommand{\cqfd}{\hfill$\Box$} \newcommand{\card}{{\operatorname{Card}}} \renewcommand{\Re}{{\operatorname{Re}}} \renewcommand{\Im}{{\operatorname{Im}}} \newcommand{\Vol}{\operatorname{Vol}} \newcommand{\Isom}{\operatorname{Isom}} \newcommand{\codim}{\operatorname{codim}} \newcommand{\vol}{\operatorname{vol}} \newcommand{\covol}{\operatorname{Covol}} \newcommand{\id}{\operatorname{id}} \newcommand{\Leb}{\operatorname{Leb}} \newcommand{\PSL}{\operatorname{PSL}} \newcommand{\SL}{\operatorname{SL}} \newcommand{\Res}{\operatorname{Res}} \newcommand{\dvol}{\;d\operatorname{vol}} \newcommand{\Det}{\operatorname{Det}} \newcommand{\arcosh}{\operatorname{argcosh}} \newcommand{\arsinh}{\operatorname{argsinh}} \newcommand{\axis}{\operatorname{Axis}} \newcommand{\bigO}{\operatorname{O}} \newcommand{\smallo}{\operatorname{o}} \newcommand{\Heis}{\operatorname{Heis}} \newcommand{\qquadr}{\QQ_{\rm quad}} \newcommand{\haarheis}{\operatorname{Haar}_{\operatorname{Heis}_3}} \newcommand{\hdr}{{\HH}^2_\RR} \newcommand{\htr}{{\HH}^3_\RR} \newcommand{\hcr}{{\HH}^5_\RR} \newcommand{\hnr}{{\HH}^n_\RR} \newcommand{\hnc}{{\HH}^n_\CC} \newcommand{\hdc}{{\HH}^2_\CC} \newcommand{\htc}{{\HH}^3_\CC} \newcommand{\phcr}{\partial_\infty\hcr} \newcommand{\SLO}{\operatorname{SL}_{2}(\OOO)} \newcommand{\GLO}{\operatorname{GL}_{2}(\OOO)} \newcommand{\PSLO}{\operatorname{PSL}_{2}(\OOO)} \newcommand{\SLOK}{\operatorname{SL}_{2}(\OOO_K)} \newcommand{\PSLOK}{\operatorname{PSL}_{2}(\OOO_K)} \newcommand{\SLH}{\operatorname{SL}_{2}(\HH)} \newcommand{\PSLH}{\operatorname{PSL}_{2}(\HH)} \newcommand{\SLC}{\operatorname{SL}_{2}(\CC)} \newcommand{\PSLC}{\operatorname{PSL}_{2}(\CC)} \newcommand{\SLZ}{\operatorname{SL}_{2}(\ZZ)} \newcommand{\PSLZ}{\operatorname{PSL}_{2}(\ZZ)} \newcommand{\SU}{\operatorname{SU}} \newcommand{\PSU}{\operatorname{PSU}} \newcommand{\autom}{\operatorname{SU}_f(\OOO_K)} \newcommand{\Pautom}{\operatorname{PSU}_f(\OOO_K)} \newcommand{\automH}{\operatorname{SU}_f(\OOO)} \newcommand{\PautomH}{\operatorname{PSU}_f(\OOO)} \newcommand{\const}{\iota} \newcommand{\discr}{\Delta} \newcommand{\cinfty}{\operatorname{\C_\infty}} \newcommand{\plane}{\operatorname{\C}} \newcommand{\SpO}{\operatorname{Sp}_1(\OOO)} \newcommand{\diam}{\operatorname{diam}} \newcommand{\tr}{\operatorname{tr}} \newcommand{\n}{\operatorname{n}} \newcommand{\Tr}{\operatorname{Tr}} \newcommand{\Norm}{\operatorname{N}} \newcommand{\Gal}{\operatorname{Gal}} \newcommand\Rep{\mathcal R} \newcommand{\stab}{\operatorname{Stab}} \newcommand{\flow}[1]{{g^{#1}}} \newcommand{\muss}[1]{{\mu^-_{W^{\rm ss}(#1)}}} \newcommand{\mus}[1]{{\mu^{\rm s}_{#1}}} \newcommand{\musu}[1]{{\mu^+_{W^{\rm su}(#1)}}} \newcommand{\muu}[1]{{\mu^{\rm u}_{#1}}} \newcommand{\normal}[1]{\partial^1_{+}{#1}} \newcommand\oPerp{\operatorname{\wt{Perp}}} \newcommand\Perp{\operatorname{Perp}} \newcommand{\curv}{\operatorname{curv}_{\rm S}} \newcommand\normalout{\partial^1_{+}} \newcommand\normalin{\partial^1_{-}} \newcommand\normalpm{\partial^1_{\pm}} \newcommand\normalmp{\partial^1_{\mp}} \newcommand\viiva{\medskip \begin{center}{---------} \end{center} \medskip } \newcounter{const} \newcounter{fig} \def\myfigure#1#2#3{ \addtocounter{fig}{1} \[ \begin{array}{c} \mbox{\psfig{figure=#1.ps,height=#2}}\\ \\ \hbox{\rm Figure \arabic{fig} ~: #3.} \end{array} \] } \def\myfigpstex#1#2{ \addtocounter{fig}{1} \[ \begin{array}{c} \mbox{\input{#1.pstex_t }}\\ \\ \hbox{\rm Figure \arabic{fig} ~: #2.} \end{array} \] } \maketitle \textbf{Abstract:} Since their introduction by Thurston, measured geodesic laminations on hyperbolic surfaces occur in many contexts. In this survey (see for a complete exposition and proofs), we give a generalization of geodesic laminations on surfaces endowed with a half-translation structure (that is a singular flat surface with holonomy $\{\pm\Id\}$), called {\it flat laminations}, and we define transverse measures on flat laminations similar to transverse measures on hyperbolic laminations, taking into account that the images of the leaves of a flat lamination are in general not pairwise disjoint. One aim is to construct a tool that could allow a fine description of the space of degenerations of half-translation structures on a surface. We define a topology on the set of measured flat laminations and a natural continuous projection of the space of measured flat laminations onto the space of measured hyperbolic laminations, for any arbitrary half-translation structure and hyperbolic metric on a surface. We prove in particular that the space of measured flat laminations is projectively compact. The main result of this survey is a classification theorem of (measured) flat laminations on a compact surface endowed with a half-translation structure. We also give an exposition of that every finite metric fat graph, outside four homeomorphisms classes, is the support of uncountably many measured flat laminations with uncountably many leaves none of which is eventually periodic, and that the space of measured flat laminations is separable and projectively compact. \footnote{ Keywords : Measured geodesic lamination, surface, half-translation structure, holomorphic quadratic differential, measured foliation, hyperbolic surface, dual tree. AMS codes 30F30, 53C12, 53C22.} \section{Introduction.} The main aim of this paper is to survey a generalization of geodesic laminations on hyperbolic surfaces (see for instance ) to the surfaces endowed with a half-translation structure, which is a flat metric with conical singular points and with holonomies in $\{\pm\Id\}$, that we will call {\it flat laminations}, as well as a definition of transverse measures on flat laminations. We refer to for complete proofs. Although the definitions are inspired of measured geodesic laminations on hyperbolic surfaces, the extension is non trivial because of several surprising phenomena, notably since the images of the leaves of a flat lamination are in general not pairwise disjoint. We will call {\it measured flat lamination} a flat lamination endowed with a transverse measure. We will define a sufficiently fine topology on the set of measured flat laminations. We will construct a (non injective) natural continuous projection of the space of measured flat laminations onto the space of measured hyperbolic laminations, for any choice of half-translation structure and (complete) hyperbolic metric on a surface. We will completely describe the fibers not reduced to a point: they are the set of simple closed hyperbolic geodesics which are homotopic to a non trivial flat cylinder of the surface (endowed with the half-tranlation structure). \medskip The space of measured flat laminations allows to consider the measured flat laminations that are the limits of some sequences of periodic local geodesics, in the projectivized space of measured flat laminations. This in turn could yield a better understanding of the degenerations of half-translation structures on a surface, as initiated in . In particular, as spaces of measures are suitable for analysis tools (distributions as in ), this could allow a finer study of the boundary of the space of half-translation structures that we will develop in a subsequent work. Let $\Sigma$ be a compact, connected, orientable surface. In this survey, we assume that $\Sigma$ is without boundary, and we refer to for the extension. A {\it half-translation structure} (or flat structure with conical singularities and holonomies in $\{\pm\Id\}$) on $\Sigma$ is the data consisting in a (possibly empty) discrete set of points $Z$ of $\Sigma$ and of a Euclidean metric on $\Sigma-Z$ with conical {\it singular points} of angles of the form $k\pi$, with $k\in\NN$ and $k\geqslant 3$, at each point of $Z$, such that the holonomy of every piecewise $\C^1$ loop of $\Sigma-Z$ is contained in $\{\pm\Id\}$. The surface $\Sigma$ endowed with a half-translation structure is a complete and locally $\CAT(0)$ metric space $(\Sigma,d)$. Let $p:(\widetilde{\Sigma},\widetilde{d})\to (\Sigma,d)$ be a locally isometric universal cover. Two local geodesics ${\ell},{\ell}'$ of $(\Sigma,d)$, defined up to changing the origins, are said to be {\it interlaced} if they have some lifts $\widetilde{\ell},\widetilde{\ell}'$ in $\widetilde{\Sigma}$ such that the image of $\widetilde{\ell}$ intersects both complementary components of $\widetilde{\ell}'(\RR)$ in $\widetilde{\Sigma}$, and conversely. A local geodesic is said to be {\it self-interlaced} if it is interlaced with itself. We endow the set of oriented, but non parametrized, local geodesics of $(\Sigma,d)$ with the quotient topology of the compact-open topology for the action by translations on the parametrizations, of $\RR$ on the parametrized local geodesics, which is called the {\it geodesic topology}. \bdefi A (geodesic) flat lamination on $(\Sigma,d)$ is a non empty set $\Lambda$ of complete local geodesics of $(\Sigma,d)$, defined up to changing origin, whose elements are called {\it leaves}, such that: \begin{itemize} \item[$\bullet$]the leaves of $\Lambda$ are non self-interlaced and pairwise non interlaced; \item[$\bullet$] $\Lambda$ is invariant by changing the orientations of the leaves; \item[$\bullet$] $\Lambda$ is closed for the geodesic topology. \end{itemize} We will call {\it support} of $\Lambda$ the union of the images of the leaves of $\Lambda$. \edefi Here are some examples of flat laminations on $\Sigma$. A {\it cylinder lamination} is a closed set of parallel leaves (and their opposites) whose images are contained in a non degenerated flat cylinder (hence, these leaves are periodic). A {\it minimal flat lamination} is a flat lamination which is the closure, for the geodesic topology, of the union of a leaf $\ell$ and its opposite $\ell^-$. It is of {\it recurrent type} if $\ell$ is regular (i.e. does not meet any singular point) and $\ell$ is not periodic. All the images of its regular leaves are then the images of some regular leaves of the vertical foliation of a quadratic differential, and thus are dense in a {\it domain} of $\Sigma$, i.e. the closure of a connected open subset bounded by periodic local geodesics. And it is of {\it finite graph type} if the image of $\ell$ is a finite graph, and if neither $\ell$ nor its opposite are periodic after a certain time. All the images of its leaves are then equal, and no leaf is periodic after a certain time. For example, a minimal vertical foliation of a holomorphic quadratic differential $q$ on a compact, connected Riemann surface is a minimal flat lamination of recurrent type, for the half-tranlation structure defined by $q$ (with a little work for the non regular leaves, see \cite[Lem.~4.11]{Morzy1}). We will construct minimal flat laminations of finite graph type in Theorem . We say that an end of a leaf (see Section for the definition) {\it terminates} in a minimal lamination or in a cylinder lamination if a representative ray is equal to a ray of a leaf of the lamination or of a boundary component of the cylinder containing the support of the cylinder lamination, up to changing the origin. The main results of this paper are the following two theorems. \btheo Every flat lamination on $\Sigma$ is a finite union of cylinder components, of minimal components (of recurrent type, finite graph type and periodic leaf travelled in both orientations) and of isolated leaves (for the geodesic topology) both of whose ends terminate in a minimal component or a cylinder component. \etheo New phenomena appear in flat laminations compared with hyperbolic ones: the images of two leaves are generally not disjoint, the flat laminations are not determined by their supports (uncountably many flat laminations can have the same support, and contrarily to a hyperbolic lamination, a flat lamination may not be minimal whereas the image of each of its leaves is dense in the support of the lamination), the cylinder components may have uncountably many leaves. Finally, there are three types, and no longer two, of minimal components of a flat lamination on a compact surface (periodic leaf travelled in both orientations, minimal component of recurrent type or of finite graph type, see Theorem for a complete statement). Compared with hyperbolic laminations, the main difficulty to define transverse measures on flat laminations is that the images of the leaves are not necessarly disjoint and that the support does not determine the lamination. Hence, we no longer define the transverse measure as a family of measures on the images of the arcs transverse to the lamination, but as a family of measures on the sets of local geodesics that intersect them transversally, and we have to refine the notion of invariance by holonomy of these families of measures. A {\it cyclic orientation} on a finite metric graph $X$ is the data of a cyclic order (see \cite[§~2.3.1]{Wolf11} for the definition) on the set of germs of edges starting at each vertex of the graph. \btheo Every cyclically oriented, connected, finite, metric graph $X$, without terminal point , is the support of uncountably many uncountable minimal flat laminations without eventually periodic leaf, on a compact, connected, orientable surface endowed with a half-translation structure, except if $X$ is homeomorphic to a circle, a dumbbell pair, a flat height or a flat theta, by a homeomorphism preserving the cyclic orientations (i.e \input{figure22.pdf_t}, where the orientations are given by the plan). \etheo \medskip Since $\Sigma$ is compact, if $m$ is a hyperbolic metric on $\Sigma$ and $\widetilde{m}$ is the pull back of $m$ on $\widetilde{\Sigma}$, there exists a unique $\grperevet$-equivariant homeomorphism between the boundaries at infinity of $\revet$ and of $\revetm$, that allows to identify them. Let $\ddp$ denote the boundary at infinity of $\widetilde{\Sigma}$ and $\dddp=\ddp\times\ddp-\{(x,x),\,x\in\ddp\}$. If $g$ is a geodesic of $\revet$ or of $\revetm$, let $E(g)=(g(-\infty),g(+\infty))$ be its ordered pair of points at infinity. Then, for every $(x,y)\in\dddp$, there exists a unique geodesic $\widetilde{\lambda}$ of $\revetm$ such that $E(\widetilde{\lambda})=(x,y)$. However, it always exists a geodesic $\widetilde{\ell}$ of $\revet$ such that $E(\widetilde{\ell})=(x,y)$, but it may happen that this geodesic is not unique, and then the set of such geodesics of $\revet$ is a set of parallel geodesics foliating a flat strip, and they project to periodic local geodesics foliating a flat cylinder. This point allows to define a non bijective correspondance between flat and hyperbolic measured laminations, from which we deduce many results of this survey and notably the two theorems above. We will also give an exposition of that the space of measured flat laminations is separable and projectively compact. \medskip In Section , we define the flat laminations and the transverse measures on flat laminations, and we endow the set of measured flat laminations with a topology. In Section , we define a proper, surjective, continuous map from the space of measured flat laminations to the space of measured hyperbolic laminations, for a complete hyperbolic metric, and we characterize its lack of injectivity. From this, we give a sketch of proof of Theorem . In Section , we define the tree associated to a measured flat lamination on a compact surface, and we construct the covering group action on it. \medskip\noindent {\small{\it Acknoledgement: I want to thank Frederic Paulin for many advice and corrections that have deeply improved the redaction of this survey.} \section{Definitions.} In this section , we give the definitions of half-translation structures on surfaces, of flat laminations and of transverse measures on flat laminations. \subsection{Half-translation structures on a surface.} In the whole paper, we will use the definitions and notation of for a surface endowed with a distance $(\Sigma,d)$: (locally) $\CAT(0)$, $\delta$-hyperbolic,... Notably, a {\it geodesic} (resp. a {\it local geodesic}) of $(\Sigma,d)$ is an isometric (resp. locally isometric) map $\ell:I\to \Sigma$, where $I$ is an interval of $\RR$. It will be called a {\it segment}, a {\it ray} or a {\it geodesic line} of $(\Sigma,d)$ if $I$ is respectively a compact interval, a closed half line (generally $[0,+\infty[$) or $\RR$. If there is no precision, a {\it geodesic} is a geodesic line. A {\it germ of geodesic ray}, or simply a {\it germ}, is an equivalence class of locally geodesic rays for the equivalence relation $r_1\sim_0 r_2$ if $r_1$ and $r_2$ coïncide on a non empty initial segment that is not reduced to a point. Similarly, the relation $r\sim_\infty r'$ if there exist $T,T'>0$ such that $r(t+T)=r'(t+T')$ for all $t\geqslant 0$, is an equivalence relation on the set of subrays of a local geodesic. An equivalence class for this equivalence relation is called an {\it end} (in the sense of Freudhental) of a local geodesic. A local geodesic has two ends. \medskip Let $\Sigma$ be a connected, orientable surface. In the whole article, we will assume that the boundary of $\Sigma$ is empty, for simplicity, but the results extend to surfaces with non empty boundary (see ). \bdefi A half-translation structure (or flat structure with conical singularities and holonomies in $\{\pm\Id\}$) on a surface $\Sigma$ is the data of a (possibly empty) discrete subset $Z$ of $\Sigma$ and a Euclidean metric on $\Sigma-Z$ with conical singularity of angle $k_z\pi$ at each $z\in Z$, with $k_z\in\NN$ and $k_z\geqslant 3$, such that the holonomy of every piecewise $\C^1$ loop in $\Sigma-Z$ is contained in $\{\pm\Id\}$. \edefi Two vectors $v_1$ and $v_2$ tangent to $\Sigma$ have the {\it same direction} if $v_2$ is the image of $\pm v_1$ by holonomy along a piecewise $\C^1$ path of $\Sigma-Z$ between the basepoints of $v_1$ and $v_2$. This definition does not depend on the choice of a path, since the holonomy of every loop is contained in $\{\pm\Id\}$. Hence, there is a notion of direction on a surface endowed with a half-translation structure, but there is no "vertical direction", as defined by a quadratic differential. A piecewise $\C^1$ path or union of paths is said to have {\it constant direction}, if all its tangent vectors, at the points in $\Sigma-Z$, have the same direction. \medskip We will denote by $[q]$ a half-translation structure on $\Sigma$, with $q$ a holomorphic quadratic differential (see \cite[§~2.5]{Morzy1} for an explanation of the notation). A half-translation structure defines a geodesic distance $d$ that is locally $\CAT(0)$. We will call {\it local flat geodesics} the local geodesics of a half-translation structure. A continuous map $\ell:\RR\to\Sigma$ is a local flat geodesic if and only if it satisfies (see \cite[Th.~5.4~p.24]{Strebel84} and \cite[Th.~8.1~p.~35]{Strebel84}): for every $t\in\RR$, \medskip \noindent $\bullet$~ if $\ell(t)$ does not belong to $Z$, there exists a neighborhoud $V$ of $t$ in $\RR$ such that $\ell_{|V}$ is an Euclidean segment (hence, $\ell_{|V}$ has constant direction); \noindent $\bullet$~ if $\ell(t)$ belongs to $Z$, then the two angles defined by the germs of $\ell([t,t+\varepsilon[)$ and $\ell(]t-\varepsilon,t])$, with $\varepsilon>0$ small enough, measured in both connected components of $U-\ell(]t-\varepsilon,t+\varepsilon[)$, with $U$ a small enough neighborhoud of $\ell(t)$, are at least $\pi$. \begin{center} \input{figure0.pdf_t} \end{center} \subsection{Geodesic laminations on surfaces endowed with a half-translation structure and with a (complete) hyperbolic metric.} Let $\Sigma$ be a connected, orientable surface (without boundary). Let $[q]$ be a half-translation structure and let $m$ be a complete hyperbolic metric on $\Sigma$. Let $p:\widetilde{\Sigma}\to\Sigma$ be a universal cover of covering group $\Gamma_{\widetilde{\Sigma}}$, let $[\widetilde{q}]$ be the unique half-translation structure and let $\widetilde{m}$ be the unique hyperbolic metric on $\widetilde{\Sigma}$ such that $p:\revet\to\srfce$ and $p:\revetm\to\srfcem$ are locally isometric. \medskip We call {\it geodesic topology} the compact-open topology on the set $\G_d$ of parametrized local geodesics for the distance $d$ defined by $[q]$ or by $m$, or the quotient topology of the compact-open topology by the action by translations on the parametrizations, of $\RR$, on the set $[\G_d]$ of local geodesics defined up to changing origin. They will be the only topologies taken into account on the spaces of local geodesics. The quotient projection from $\G_d$ to $[\G_d]$ will be denoted by $g\mapsto[g]$, and if $F\subset\G_d$, we will denote by $[F]$ its image in $[\G_d]$. If $g:\RR\to (\Sigma,d)$ is an element of $\G_d$ or of $[\G_d]$, we denote its {\it opposite} by $g^-(t):t\mapsto g(-t)$. Finally, if $F$ is a set of elements of $\G_d$ or of $[\G_d]$, we call {\it support of $F$} the union of the images of the elements of $F$, denoted by $\Supp(F)$. \blemm\cite[Coro.~2.4]{Morzy1} If $F$ is a closed set of $\G_d$ or of $[\G_d]$ (for the geodesic topology), then $\Supp(F)$ is closed in $\Sigma$. \elemm The opposite is true on a hyperbolic surface, but it may be false on a surface endowed with a half-translation structure. We will see (Theorem ) that it exists some surfaces endowed with a half-tranlation structure, having a flat local geodesic $\ell$ whose image is a finite graph (and hence is closed), such that the closure of $\ell$ for the geodesic topology contains uncountably many geodesics. \medskip In \cite[§~2.3]{Morzy1}, we have given a very global definition of {\it interlaced local geodesics}, in a locally $\CAT(0)$, complete, connected metric space, whose boundary at infinity of a universal cover is endowed with a (total) cyclic order. Here, we only recall the specific definition in the case of connected, orientable surfaces without boundary, endowed with a complete locally $\CAT(0)$ metric $d$. Since the universal cover $(\widetilde{\Sigma},\widetilde{d})$ is $\CAT(0)$, the intersection of the images of two geodesics of $(\widetilde{\Sigma},\widetilde{d})$ is connected, possibly empty. If $\widetilde{\ell}$ is a geodesic of $(\widetilde{\Sigma},\widetilde{d})$, defined up to changing the origin, then $\widetilde{\Sigma}-\widetilde{\ell}(\RR)$ has two connected components. Two complete geodesics $\widetilde{\ell},\widetilde{\ell'}$ of $(\widetilde{\Sigma},\widetilde{d})$, defined up to changing the origins, are {\it interlaced} if $\widetilde{\ell}$ meets both connected components of $\widetilde{\Sigma}-\widetilde{\ell}'(\RR)$, or the opposite (which is equivalent). Two local geodesics of $(\Sigma,d)$ are {\it interlaced} if they admit some lifts in $(\widetilde{\Sigma},\widetilde{d})$ which are interlaced, and a local geodesic is {\it self-interlaced} if it is interlaced with itself. If $d$ is the distance defined by $m$, two local geodesics are non interlaced if and only if their images are disjoint and a local geodesic is non self-interlaced if and only if it is simple, but this equivalence is not true in the case of $[q]$. We refer to \cite[§~3.1]{Morzy1} for a characterization of the local geodesics for $[q]$ that are non interlaced. \bdefi A geodesic lamination (or simply a lamination) of $(\Sigma,d)$, with $d$ the distance defined by $m$ or $[q]$, is a non empty set $\Lambda$ of (complete) local geodesics of $(\Sigma,d)$, defined up to changing origin, whose elements are called leaves, such that: \medskip \noindent $\bullet$~ leaves are non self-interlaced; \noindent $\bullet$~ leaves are pairwise non interlaced; \noindent $\bullet$~ if $\ell$ belongs to $\Lambda$ then so do $\ell^-$, with $\ell^{-}(t):t\mapsto\ell(-t)$; \noindent $\bullet$~ $\Lambda$ is closed for the geodesic topology. \edefi We say that $\Lambda$ is a {\it flat lamination} if $d$ is defined by $[q]$ and that $\Lambda$ is a {\it hyperbolic lamination} if $d$ is defined by $m$. Usually, a hyperbolic lamination is defined as a non empty closed subset of $\Sigma$, which is a disjoint union of images of simple hyperbolic local geodesics. The definitions are equivalent in the case of hyperbolic laminations but not in the case of flat laminations. For example, it may happen that a flat local geodesic is not simple whereas it is non self-interlaced (see \cite[§~4.1]{Morzy1}). \subsection{Measured flat laminations.} Let $\srfce$ be a connected, orientable surface (without boundary), endowed with a half-translation structure. An {\it arc} is a piecewise $\C^1$ map $\alpha:[0,1]\to\Sigma$ which is a homeomorphism onto its image. Let $\Lambda$ be a flat lamination of $\srfce$. An arc $\alpha$ is {\it transverse} to a leaf or to a segment of leaf $\ell$ of $\Lambda$ if \noindent $\bullet$~$\alpha$ is transverse to $\ell$ in the complementary of the singular points of $[q]$ and of the singular points of $\alpha$; \medskip \noindent $\bullet$~ for every singular point $x$ of $[q]$ or of $\alpha$ in $\Image(\ell)\cap\alpha(]0,1[)$, there exists a neighborhoud $U$ of $x$ that is a topological disk, and a segment $S$ of $\ell$ such that $U-\Image(S)\cap U$ has two connected components and the connected components of $U\cap(\alpha([0,1])-\{x\})$ are contained in different components of $U-\Image(S)\cap U$; \begin{center} \input{dessin1mesure.pdf_t} \end{center} \noindent $\bullet$~ $\alpha$ is tangent to $\ell$ neither in $0$ nor in $1$. \medskip An arc $\alpha$ is {\it transverse to a set $F$ of leaves} or of segments of leaves of $\Lambda$ if it is transverse to every element of $F$. In particular, an arc is transverse to $\Lambda$ if it is transverse to every leaf of $\Lambda$. If $\alpha:[0,1]\to\Sigma$ is an arc of $\Sigma$, we denote by $G(\alpha)$ the compact set of elements of $\Gq$ (the set of parametrized local geodesics of $\srfce$) which are transverse to $\alpha$ and whose origins belong to $\alpha([0,1])$. By definition, if $\alpha'([0,1])\subseteq\alpha([0,1])$, then $G(\alpha')\subseteq G(\alpha)$. Let $F_1\subseteq\Gq$ be such that $[F_1]\subseteq\Lambda$ and let $\alpha_1$ and $\alpha_2$ be two disjoint arcs transverse to $F_1$, such that $F_1\subseteq G(\alpha_1)$ and every element of $F_1$ intersects $\alpha_2([0,1])$ at a positive time. For every $g_1\in F_1$, we define $t_{g_1}= \min\{t>0\;:\;g_1(t)\in \alpha_2([0,1])\}$, and $F_2$ the subset of the elements $g_2\in G(\alpha_2)$ such that there exists $g_1\in F_1$ with $g_2(t)=g_1(t+t_{g_1})$ for all $t\in\RR$. A {\it holonomy} $h:F_1\to F_2$ of $\Lambda$ is a homeomorphism between $F_1$ and $F_2$ defined by $h(g_1)=g_2:t\mapsto g_1(t+t_{g_1})$ such that there exists a homotopy $H:[0,1]\times[0,1]\to\Sigma$ between $\alpha_1$ and $\alpha_2$ such that: \noindent $\bullet$~ for every $t\in\,[0,1]$, the map $s\mapsto H(s,t)$ is an arc transverse to every segment of leaf $g_{1|[0,\,t_{g_1}]}$, with $g_1\in F_1$; \noindent $\bullet$~ for every $\ell\in F_1$, there exists $s_\ell\in[0,1]$ such that $t\mapsto H(s_\ell,t)$ is a segment of $\ell$ (up to changing the parametrization); \noindent $\bullet$~ the intersections $H([0,1]\times]0,1[)\cap \alpha_i([0,1])$ with $i=1,2$ are empty. Contrarily to the case of measured foliations, if the images of the geodesics are not pairwise disjoint, the map $H$ may not be injective. \bdefi A transverse measure on $\Lambda$ is a family $\mu=(\mu_\alpha)_\alpha$ of Radon measures $\mu_\alpha$ defined on $G(\alpha)$, for every arc $\alpha$ transverse to $\Lambda$, such that: \noindent $(1)$~ the support of $\mu_\alpha$ is the set $\{\ell\in G(\alpha)\;:\;[\ell]\in\Lambda\}$; \noindent $(2)$~ if $h:F_1\to F_2$ is a holonomy of $\Lambda$, where $\alpha_1,\alpha_2$ are two disjoint arcs transverse to $F_1$ and $F_1\subset G(\alpha_1)$ and $F_2\subset G(\alpha_2)$ are some Borel sets, then $h_*(\mu_{\alpha_1|F_1})=\mu_{\alpha_2|F_2}$; \noindent $(3)$~ $\mu_\alpha$ is $\iota$-invariant, with $\iota(\ell)=\ell^-:t\mapsto\ell(-t)$; \noindent $(4)$~ if $\alpha'([0,1])\subseteq\alpha([0,1])$, then $\mu_{\alpha|G(\alpha')}=\mu_{\alpha'}$. \edefi We will denote by $(\Lambda,\mu)$ a flat lamination endowed with a transverse measure, that we will call a {\it measured flat lamination}, and we will denote by $\mathcal{M}\mathcal{L}_p(\Sigma)$ the set of measured flat laminations on $\Sigma$. We endow $\mathcal{M}\mathcal{L}_p(\Sigma)$ with the topology such that a sequence $(\Lambda_n,\mu_n)_{n\in\NN}$ converges to $(\Lambda,\mu)$ if and only if for every arc $\alpha$, if $\alpha$ is transverse to $\Lambda$, then $\alpha$ is transverse to $\Lambda_n$ for $n$ large enough and $\mu_{n,\alpha}\overset{*}{\rightharpoonup}\mu_\alpha$ in the space of Radon measures on $G(\alpha)$. A leaf $\ell$ of $\Lambda$ is {\it positively recurrent} if there exists an arc $\alpha$ transverse to ${\ell}$ such that $\ell$ intersects $\alpha([0,1])$ at an infinite number of positive times. \blemm\cite[Lem.~7]{Morzy2} If $\Lambda$ is endowed with a transverse measure $\mu$, then the only leaves of $\Lambda$ which are isolated and positively recurrent are the periodic leaves. \elemm \section{Link between measured flat laminations and measured hyperbolic laminations.} In this section , we denote by $\srfce$ a compact, connected, orientable surface (without boundary), endowed with a half-translation structure, and by $p:\revet\to\srfce$ a locally isometric universal cover whith covering group $\Gamma_{\widetilde{\Sigma}}$. We assume that $\chi(\Sigma)<0$, and we denote by $m$ a hyperbolic metric on $\Sigma$ and by $\widetilde{m}$ the pull back of $m$ on $\widetilde{\Sigma}$. Since $\Sigma$ is compact, there exists a unique $\grperevet$-equivariant homeomorphism between the boundaries at infinity of $\revet$ and of $\revetm$, that allows to identify them. Let $\ddp$ denote the boundary at infinity of $\widetilde{\Sigma}$ and $\dddp=\ddp\times\ddp-\{(x,x),\,x\in\ddp\}$. If $g$ is a geodesic of $\revet$ or of $\revetm$, lets $E(g)=(g(-\infty),g(+\infty))$ be its ordered pair of points at infinity. If $\Lambda$ is a geodesic lamination of $\srfce$ (resp. $\srfcem$), and if $\widetilde{\Lambda}$ is the set of lifts of the leaves of $\Lambda$ in $\widetilde{\Sigma}$, then $\widetilde{\Lambda}$ is a geodesic lamination of $\revet$ (resp. $\revetm$) which is $\Gamma_{\widetilde{\Sigma}}$-equivariant. Conversely, if $\widetilde{\Lambda}$ is a $\Gamma_{\widetilde{\Sigma}}$-equivariant lamination of $\revet$ (resp. $\revetm$), then it projects to a lamination of $\srfce$ (resp. $\srfcem$), and the set $\mathcal{L}_{[q]}(\Sigma)$ (resp. $\mathcal{L}_h(\Sigma)$) of geodesic laminations of $\srfce$ (resp. $\srfcem$) is homeomorphic to the set $\mathcal{L}_{\grperevet}([\Gqr])$ (resp. $\mathcal{L}_{\grperevet}([\Gmr])$) of geodesic laminations of $\revet$ (resp. $\revetm$) which are $\grperevet$-equivariant. Similarly, the spaces $\mathcal{M}\mathcal{L}_{[q]}(\Sigma)$ (resp. $\mathcal{M}\mathcal{L}_h(\Sigma)$) of measured flat (resp. hyperbolic) laminations on $\Sigma$, endowed respectivelly with the topology defined above and with the topology defined in \cite[p.~19]{Bonahon97}, are homeomorphic to the spaces of Radon measures on $[\Gqr]$ (resp. $[\Gmr]$) which are $\Gamma_{\widetilde{\Sigma}}$ et $\iota$-invariant (with $\iota(\widetilde{g})=\widetilde{g}^-:t\mapsto\widetilde{g}(-t)$), whose supports are flat (resp. hyperbolic) laminations, denoted by $\M_{\grperevet}([\Gqr])$ (resp. $\M_{\grperevet}([\Gmr])$) (see \cite[Prop.~17~p.~154]{Bonahon88} and \cite[§4]{Morzy2}). If $(x,y)\in\dddp$ there exists a unique geodesic $\widetilde{\lambda}$ of $\revetm$ such that $E(\widetilde{\lambda})=(x,y)$. It is not the same in $\revet$. There always exists a geodesic $\widetilde{\ell}$ of $\revet$ such that $E(\widetilde{\ell})=(x,y)$, but it may not be unique. We recall that two geodesics $c$ and $c'$ are at \textit{finite Hausdorff distance} if there exists $K>0$ such that $d(c(t),c'(t))\leqslant K$ for all $t\in\RR$. \blemm\cite[Th.~2.(c)]{MS85} We recall that $\Sigma$ is compact. Let $\widetilde{\ell}_1$ and $\widetilde{\ell}_2$ be two geodesics of $\revet$, such that $\widetilde{\ell}_1$ and $\widetilde{\ell}_2$ are at finite Hausdorff distance and such that there exists $\delta>0$ with $d(\widetilde{\ell}_1(t),\widetilde{\ell}_2(\RR))\geqslant\delta$ for all $t\in\RR$. Then, the convex hull of $\widetilde{\ell}_1(\RR)\cup \widetilde{\ell}_2(\RR)$ is isometric to a flat strip $\RR\times[0,D]\subset \RR^2$, with $D\geqslant 0$, and the projections $p\circ\widetilde{\ell}_1$ and $p\circ\widetilde{\ell}_2$ of $\widetilde{\ell}_1$ and $\widetilde{\ell}_2$ on $\Sigma$ are periodic local geodesics which are freely homotopic to the boundary of a flat cylinder. \elemm Hence, if $\varphi:[\Gqr]\to[\Gmr]$ is the map associating to the geodesic $\widetilde{\ell}\in[\Gqr]$ the unique geodesic $\varphi(\widetilde{\ell})\in[\Gmr]$ such that $E(\varphi(\widetilde{\ell}))=E(\widetilde{\ell})$ (see \cite[§4.2]{Morzy1}), then $\varphi$ is surjective and continuous, and a closed subset $F$ of $[\Gqr]$ is a flat lamination if and only if $\varphi(F)$ is a hyperbolic lamination. However, the map $\varphi$ is not injective. By definition, two different geodesics $\widetilde{\ell},\widetilde{\ell}'$ of $\revet$ have the same image by $\varphi$ if and only if they have the same ordered pair of points at infinity. According to Lemma , their images are thus parallel and contained in a maximal flat strip of $\revet$, and their projections in $\srfce$ are freely homotopic periodic local geodesics, hence their images are contained in a maximal flat cylinder. The points at infinity of $\widetilde{\ell}$, $\widetilde{\ell}'$ and $\varphi(\widetilde{\ell})=\varphi(\widetilde{\ell}')$ are hence the attractive and repulsive fixed points of an element of the covering group, and the projection of $\varphi(\widetilde{\ell})$ in $\srfcem$ is a simple closed local geodesic. Hence, the restriction of $\varphi$ to the set of geodesics of $\revet$ whose projections are not periodic is injective and its image is the set of geodesics of $\revetm$ whose projections are not closed local geodesics, and the preimage of a geodesic $\widetilde{\lambda}$ of $\revetm$ whose projection is a simple closed local geodesic, is the set of geodesics of $\revet$ having the same ordered pair of points at infinity than $\widetilde{\lambda}$, which may be a unique geodesic, or may be a set of parallel geodesics whose images are contained in a flat strip, and whose projections are some periodic local geodesics freely homotopic to $p\circ\widetilde{\lambda}$. Since the sets $\mathcal{L}\srfce$ and $\mathcal{L}\srfcem$ of flat and hyperbolic laminations on $\Sigma$ are respectivelly homeomorphic to the sets $\mathcal{L}_{\grperevet}\revet$ and $\mathcal{L}_{\grperevet}\revetm$ of $\grperevet$-invariant flat and hyperbolic laminations on $\widetilde{\Sigma}$, the map $\varphi$ defines a continous, surjective map $\psi:\mathcal{L}\srfce\to\mathcal{L}\srfcem$. It implies the structure theorem of flat laminations. We recall the two main results of about flat laminations. In Theorem , $\Lambda$ is a flat lamination. A {\it cylinder component} is a maximal set of leaves of $\Lambda$ whose images are contained in a non degenerated flat cylinder (hence, these leaves are periodic). A {\it minimal component} is a sublamination which is the closure, for the geodesic topology, of a leaf $\ell$ and its opposite $\ell^-$. The minimal component is of {\it recurrent type} if $\ell$ is regular (i.e. does not meet any singular point) and is not periodic. All the images of its regular leaves are then the images of some regular leaves of the vertical foliation of a quadratic differential, and thus are dense in a {\it domain} of $\Sigma$, i.e. the closure of a connected open subset bounded by periodic local geodesics. The minimal component is of {\it finite graph type} if the image of $\ell$ is a finite graph, and if neither $\ell$ nor its opposite are eventually periodic. All the images of its leaves are then equal, and no leaf is eventually periodic. An end of a local geodesic {\it terminates} in a minimal component or in a cylinder component if there exists a ray in the equivalence class of this end which is the ray of a leaf of the minimal component or a ray of a boundary component of the corresponding flat cylinder. \input{feuilletage.pdf_t}\;\input{graph.pdf_t}\;\input{cylindre.pdf_t} \btheo\cite[§~6]{Morzy1} Let $\Lambda$ be a flat lamination on a compact, connected, orientable surface endowed with a half-translation structure. Then $\Lambda$ is a finite union of cylinder components, of minimal components (of recurrent type, finite graph type and periodic leaf travelled in both orientations) and of isolated leaves (for the geodesic topology) both of whose ends terminate in a minimal component or a cylinder component. \etheo This theorem comes from the structure theorem of hyperbolic laminations and from a few observations (see \cite[Coro.~4.6, Lem.~4.10, Lem.~4.11]{Morzy1}): the preimage of a minimal lamination of $\srfcem$, which is not a closed leaf, is a minimal lamination of $\srfce$ which either contains a leaf whose image is not compact and then the lamination is of recurrent type, or contains a leaf whose image is compact, and then the lamination is of finite graph type. The preimage by $\psi$ of a closed leaf of $\srfcem$ is either a periodic leaf or a cylinder lamination. Finally, the preimage of an isolated local geodesic of $\srfcem$ whose an end spirales to a minimal hyperbolic lamination is an isolated local geodesic of $\srfce$ whose the corresponding end terminates in the preimage of the hyperbolic lamination, since the flat geodesics can change direction at most at singular points (see \cite[§4.4]{Morzy1}). \begin{center} \input{figure21.pdf_t} \end{center} \rem If $\Lambda$ is a minimal flat lamination, then the image of every ray of every leaf of $\Lambda$ is dense in $\Supp(\Lambda)$. However, contrary to the case of hyperbolic lamination, the opposite is false. For example, let consider the surface endowed with a half-translation structure above, whose singular points are of angle $3\,\pi$ and located at the vertices of the graph. Then, the union of both local flat geodesics in the free homotopy classes of the two closed curves drawn on the picture (and their opposites) is a flat lamination such that the image of every ray of every leaf is equal to the graph, but the lamination has two minimal components. \medskip We now show that every finite, connected graph, without terminal point , outside four homeomorphism classes, is the support of uncountably many geodesic laminations with uncountably many leaves none of which is eventually periodic. We use the definitions and notation about graphs of . Every graph is identified with its topological realization, and it is endowed with a geodesic distance such that each half-edge is isometric to a compact interval of $\RR$. In this paper, we only consider finite, connected graph, without terminal point. A {\it cyclic orientation} on a finite metric graph $X$ is the data of cyclic orders on the sets of half-edges issued from each vertex of $X$ (a cyclic order on the set of half-edges issued from a vertex is globally an embedding of this set in the oriented unit circle $\SSS^1$, see \cite[§~2.3.1]{Wolf11} for the exact definition). A {\it cyclically oriented} graph is a graph endowed with a cyclic orientation. In \cite[§5]{Morzy1}, we have recalled the classical result that every cyclically oriented, finite, connected graph $X$, without terminal point, can be isometrically embedded in a compact, connected surface $\Sigma(X)$ endowed with a half-translation structure, made of flat cylinders locally isometrically glued on the graph $X$, and that retracts by strong defomation on $X$ (that is why such a graph is often called a {\it fat graph}). \begin{center} \input{figure17.pdf_t} \end{center} The Euler characteristic of this surface is negative, except if $X$ is a circle (which embbeds in a flat cylinder), hence it can be endowed with a hyperbolic metric. If $\Sigma(X)$ is not a pair of pant, we have seen (see \cite[Lem.~4.2]{Morzy1}) that there exists a minimal hyperbolic lamination $\Lambda$ on $\Sigma(X)$ (endowed with a complete hyperbolic metric with totally geodesic boundary) such that the completions of the connected components of $\Sigma(X)- \Supp(\Lambda)$ are ideal triangles or ideal monogons minus a disk (actually, the proof of the lemma implies that there exists uncountably many such hyperbolic lamination, with a fixed hyperbolic metric). Then, we show that the preimage by $\psi$ of such a lamination is a minimal flat lamination, which is not a periodic leaf travelled in both orientations, whose support is the graph $X$. Since $\Sigma(X)$ is a pair of pant if and only if $X$ is homeomorphic to a circle, a dumbbell pair, a flat height or a theta by a homeomorphism preserving the cyclic orientations, we get the following theorem. \begin{center} \input{figure18.pdf_t} \end{center} \btheo\cite[Th.~5.2]{Morzy1} Every cyclically oriented, connected, finite, metric graph $X$, without extremal point, may be the support of uncountably many uncountable minimal flat laminations with no eventually periodic leaf, on a compact and connected surface endowed with a half-translation structure, except if $X$ is homeomorphic to a circle, a dumbbell pair, a flat height or a flat theta, by a homeomorphism preserving the cyclic orientations (i.e \input{figure22.pdf_t}, where the orientations are given by the plan). \etheo We now focus on the transverse measures on the laminations. The map $\varphi$ is proper. Hence we have the following lemma. \blemm\cite[Lem.~10]{Morzy2} The map $\varphi$ defines a continuous, surjective and proper map $\varphi_*$ from the space of Radon measures on $[\Gqr]$ to the space of Radon measures on $[\Gmr]$, and then a continuous map $\psi:\mathcal{M}\mathcal{L}_p(\Sigma)\to\mathcal{M}\mathcal{L}_h(\Sigma)$ which is proper and surjective. \elemm A hyperbolic lamination on a compact surface can be endowed with a transverse measure if and only if it has no isolated leaf spiraling on a minimal sublamination. Similarly, a measured flat lamination has no isolated leaf terminating in a minimal sublamination or a cylinder sublamination. Moreover, if a flat lamination $\Lq$ has no isolated leaf which is not periodic, then $\psi(\Lq)$ has no isolated leaf spiraling on a minimal sublamination, then there exists a transverse measure on $\psi(\Lq)$ and hence there exists a transverse measure on $\Lq$. In conclusion, a flat lamination on a compact surface can be endowed with a transverse measure if and only if it has no isolated leaf which is not periodic. Notably, every minimal flat lamination of finite graph type $\Lambda$ can be endowed with a transverse measure, and the space of transverse measures on $\Lambda$ is homeomorphic to the space of transverse measures on the hyperbolic minimal lamination $\psi(\Lambda)$. \medskip The group $\RR^{+*}$ acts on $\mathcal{M}\mathcal{L}_p(\Sigma)$ and on $\mathcal{M}\mathcal{L}_h(\Sigma)$ by multiplication of the measures. We denote by $\mathcal{P}\mathcal{M}\mathcal{L}_p(\Sigma)$ and $\mathcal{P}\mathcal{M}\mathcal{L}_h(\Sigma)$ the quotient spaces for these actions. Since $\psi$ is equivariant by these actions, it defines a continuous map $\overline{\psi}: \mathcal{P}\mathcal{M}\mathcal{L}_p(\Sigma)\to \mathcal{P}\mathcal{M}\mathcal{L}_h(\Sigma)$ that is surjective and proper. We deduce from this the following lemmas. \blemm\cite[Lem.~11]{Morzy2} The space $\mathcal{P}\mathcal{M}\mathcal{L}_p(\Sigma)$ is compact. \elemm A {\it measured cylinder lamination} is a measured flat lamination having a unique component which is a cylinder component or a pair of (periodic) leaves. \blemm\cite[Lem.~12]{Morzy2} Since $\Sigma$ is compact, the set of measured cylinder laminations having finitely many leaves is dense in $\mathcal{M}\mathcal{L}_p(\Sigma)$. In particular, $\mathcal{M}\mathcal{L}_p(\Sigma)$ is separable. \elemm As $\varphi$, the maps $\psi$ and $\overline{\psi}$ are not injective. Assume that $\Sigma$ is compact with genus $g\in\NN$. Considering the lack of injectivity of $\varphi_*$, we see that the preimage of a measured hyperbolic lamination having no closed leaf consists in a unique measured flat lamination having no periodic leaf. However, if $(\Lambda_m,\mu_m)$ is a closed leaf $\lambda$ endowed with a transverse measure which, for every arc $\alpha$ such that $G(\alpha)$ contains $\lambda$, is a Dirac measure at $\lambda$ of mass $\delta>0$, the preimage of $(\Lambda_m,\mu_m)$ by $\psi$ is the set of measured cylinder laminations whose supports are closed sets $F$ of leaves that are freely homotopic to $\lambda$. If the set of local geodesics of $\srfce$ that are freely homotopic to $\lambda$ contains at least two elements, it foliates a maximal flat cylinder. Then, this set is homeomorphic to $[0,L]$, with $L$ is the height of the flat cylinder, so $F$ is homeomorphic to a closed subset of $[0,L]$. Hence, the preimage of $(\Lambda_m,\mu_m)$ by $\psi$ is homeomorphic to the set of Borel measures on $[0,L]$ of mass $\delta$. Since $\varphi_*$ is equivariant for the addition of measures, we see that if a measured hyperbolic lamination has some closed leaves $\lambda_1,\dots,\lambda_p$, of respective masses $\delta_1,\dots,\delta_p$ ($p$ is always at most $3g-3$), then the preimage of $(\Lambda_m,\mu_m)$ by $\psi$ is homeomorphic to the Cartesian product of the sets of Borel measures on $[0,L_i]$, $1\leqslant i\leqslant p$, where $L_i$ is the height of the maximal flat cylinder, union of the local geodesics of $\srfce$ freely homotopic to $\lambda_i$, whose total mass is $\delta_i$. Since $\Sigma$ is compact, the projectified space $\mathcal{P}\mathcal{M}\mathcal{L}_h(\Sigma)$ is homeomorphic to the sphere $\SSS^{6g-6}$ (see \cite[Th.~17]{Bonahon97}). If the support of the measured hyperbolic laminations in the equivalence class of $x\in\mathcal{P}\mathcal{M}\mathcal{L}_h(\Sigma)$ has no closed leaf, then the preimage of $\{x\}$ by $\overline{\psi}$ is a single point. However, if the closed leaves $\lambda_1,\dots\lambda_p$ belong to the support of the measured hyperbolic laminations in the equivalence class of $x$, and if $L_1,\dots,L_p$ are the heights of the maximal flat cylinders, unions of the images of the periodic local geodesics of $\srfce$ that are freely homotopic to $\lambda_1,\dots,\lambda_p$ ($L_i=0$ if there is only one periodic local geodesic that is freely homotopic to $\lambda_i$), then the preimage of $\{x\}$ by $\overline{\psi}$ is homeomorphic to the Cartesian product with $p$ terms of the sets of Borel measures on $[0,L_i]$ of total masses at most $1$. In particular, since the set of measured hyperbolic laminations whose support contains a closed leaf is projectively dense in the space of measured hyperbolic laminations, the subset of $\mathcal{P}\mathcal{M}\mathcal{L}_h(\Sigma)$ of points whose preimages by $\overline{\psi}$ are not single points, is dense in $\mathcal{P}\mathcal{M}\mathcal{L}_h(\Sigma)$. \section{Tree associated to a measured flat lamination.} Let $\srfce$ be a compact, connected, orientable surface (without boundary), endowed with a half-translation structure, such that $\chi(\Sigma)<0$, let $p:\revet\to\srfce$ be a locally isometric universal cover of covering group $\Gamma_{\widetilde{\Sigma}}$ and let $(\Lambda,\mu)$ be a measured flat lamination on $\srfce$. We denote by $(\widetilde{\Lambda},\widetilde{\mu})$ its preimage in $\revet$ and by $\nu_{\widetilde{\mu}}$ the element of $\M_{\grperevet}([\Gqr])$ that it defines (see the beginning of Section or \cite[§4]{Morzy2}). We assume that $\nu_{\widetilde{\mu}}$ has no atom. If it had some, we would blow up them (see the end of \cite[§6]{Morzy2}). If $\widetilde{\ell}$ is a leaf of $\widetilde{\Lambda}$, then $\widetilde{\Sigma}-\widetilde{\ell}(\RR)$ has two connected components and the image of a leaf of $\widetilde{\Lambda}$ is contained in the closure of one of them, since the leaf is not interlaced with $\widetilde{\ell}$. Let $\widetilde{\ell}_0$ and $\widetilde{\ell}_1$ be two leaves of $\widetilde{\Lambda}$. We denote by $C_i$ the connected component of $\widetilde{\Sigma}-\widetilde{\ell}_i(\RR)$ that contains $\widetilde{\ell}_{i+1}(\RR)$ (with $i\in\ZZ/2\ZZ$). We define $C(\widetilde{\ell}_0,\widetilde{\ell}_1)=C_0\cap C_1$. Let $c$ be a geodesic segment joining their images (if the images are not disjoint, the segment $c$ may be a point). A leaf of $\widetilde{\Lambda}$ {\it intersects $c$ non trivially } if it is contained in $C(\widetilde{\ell}_0,\widetilde{\ell}_1)$ and if it intersects both complementary components of the image of $c$ in $C(\widetilde{\ell}_0,\widetilde{\ell}_1)$, and we denote by $B(\widetilde{\ell}_0,\widetilde{\ell}_1)= B_{\widetilde{\Lambda}}(\widetilde{\ell}_0,\widetilde{\ell}_1)$ the set of leaves of $\widetilde{\Lambda}$ intersecting $c$ non trivially. \blemm The compact set $B(\widetilde{\ell}_0,\widetilde{\ell}_1)$ does not depend on the choice of $c$. \elemm If $\{\widetilde{\ell}_1,\widetilde{\ell}_2\}$ is a pair of leaves of $\widetilde{\Lambda}$, we set $\widetilde{d}_{\widetilde{\Lambda}}(\widetilde{\ell}_1,\widetilde{\ell}_2)=\frac{1}{2}\nu_{\widetilde{\mu}}(B(\widetilde{\ell}_1,\widetilde{\ell}_2))$. Then $\widetilde{d}_{\widetilde{\Lambda}}$ is a pseudo-distance on $\widetilde{\Lambda}$. We denote by $(T,d_T)$ the quotient metric space $(\widetilde{\Lambda},\widetilde{d}_{\widetilde{\Lambda}})/\sim$, where $\widetilde{\ell}\sim\widetilde{\ell}'$ if and only if $\widetilde{d}_{\widetilde{\Lambda}}(\widetilde{\ell},\widetilde{\ell}')=0$ (or equivalently if $B(\widetilde{\ell},\widetilde{\ell}')=\{\widetilde{\ell},\widetilde{\ell}'\}$). \blemm\cite[Lem.~16]{Morzy2} The metric space $(T,d_T)$ is an $\RR$-tree. \elemm Let $m$ be a hyperbolic metric on $\Sigma$. We recall that $\mathcal{M}\mathcal{L}_h(\Sigma)$ is the space of measured hyperbolic laminations on ${\Sigma}$. The set of free homotopy classes of simple closed curves on $\Sigma$ endowed with transverse measures which are Dirac measures of positive masses, embeds into $\mathcal{M}\mathcal{L}_h(\Sigma)$, and the intersection number on this set can be extended, in a unique way, to a continuous map $i:\mathcal{M}\mathcal{L}_h(\Sigma)\times\mathcal{M}\mathcal{L}_h(\Sigma)\to\RR^+$ (see \cite[Prop.~3]{Bonahon88}). According to Lemma , the map $\varphi_*$ defines a map $\psi:\mathcal{M}\mathcal{L}_p(\Sigma)\to\mathcal{M}\mathcal{L}_h(\Sigma)$. Let $\alpha$ be a non trivial free homotopy class of closed curves, let $(\Lq,\mu_{[q]})$ be a measured flat lamination and let $\nu_{\mu_{[q]}}\in\M_{\grperevet}([\Gqr])$ be the measure defined by $(\Lq,\mu_{[q]})$ on $[G_{[\widetilde{q}]}]$. We define the intersection number between $(\Lq,\mu_{[q]})$ and $\alpha$ by $$i_{[q]}(\mu_{[q]},\alpha)= i(\psi(\Lq,\mu_{[q]}),\alpha).$$ If $k\in\NN$, we have $i_{[q]}(\mu_{[q]},\alpha^k)=k\, i_{[q]}(\mu_{[q]},\alpha)$. Hence, we assume that $\alpha$ is primitive (i.e. if there exists a free homotopy class $\alpha_0$ such that $\alpha=\alpha_0^k$, then $k=\pm1$). We denote by $\alpha_{[q]}$ a periodic local flat geodesic in the class of $\alpha$, and by $\widetilde{\alpha}_{[q]}$ a lift of $\alpha_{[q]}$ in $\widetilde{\Sigma}$. Let $\gamma\in\grperevet-\{e\}$ be one of the two primitive hyperbolic elements of $\grperevet$ whose translation axis is $\widetilde{\alpha}_{[q]}(\RR)$, and let $(\widetilde{\Lambda}_{[q]},\mutilde_{[q]})$ be the preimage of $(\Lambda_{[q]},\mu_{[q]})$ in $\widetilde{\Sigma}$. \blemm\cite[Lem.~15]{Morzy2} Let $\widetilde{\ell}$ be a leaf of $\Lqr$ which is interlaced with $\widetilde{\alpha}_{[q]}$. The number $i_{[q]}(\mu_{[q]},\alpha)$ is equal to $\frac{1}{2}\nu_{\mu_{[q]}}(B_{\widetilde{\Lambda}} (\widetilde{\ell},\gamma\widetilde{\ell})-\gamma\widetilde{\ell})$. If there exists no such leaf, then $i_{[q]}(\mu_{[q]},\alpha)=0$. \elemm \rem We could define the intersection number between a free homotopy class of closed curves with $(\Lambda_{[q]},\mu_{[q]})$ by the infimum of the masses given by the measured flat lamination to the closed curves that are piecewise transverse to the lamination, similarly to the intersection number with a measured foliation, but this infimum would not be necessarly attained since the periodic local geodesics are generally not piecewise transverse to the lamination. Furthermore, in the case of compact surfaces endowed with a half-translation structure, contrarily to the case of measured hyperbolic lamination (see \cite[Th.~2]{Ota90}), the intersection numbers with the free homotopy classes of closed curves of $\Sigma$ do not separate the measured flat laminations, but only their images in $\mathcal{M}\mathcal{L}_h(\Sigma)$. In particular, the topology defined after Definition is not equivalent to the one induced by the product topology on $\RR^{{\cal{H}}}$, with ${{\cal{H}}}$ the set of free homotopy classes of closed curves, on the image of $\mathcal{M}\mathcal{L}_p(\Sigma)$ by the map $(\Lambda_{[q]},\mu_{[q]})\mapsto (i(\mu_{[q]},\alpha))_{\alpha\in{{\cal{H}}}}$. \medskip The covering group $\Gamma_{\widetilde{\Sigma}}$ acts on $\widetilde{\Sigma}$ by isometries. Hence, it defines an action on the set $[\G_{[\widetilde{q}]}]$ of geodesics of $\revet$ that are defined up to changing origin. Since $\widetilde{\Lambda}$ is $\grperevet$-invariant, this action defines an action on $\widetilde{\Lambda}$. Since for every $\gamma\in\grperevet$ we have $\gamma_*\nu_{\widetilde{\mu}}=\nu_{\widetilde{\mu}}$ and $\gamma B(\widetilde{\ell},\widetilde{\ell}')=B(\gamma\widetilde{\ell},\gamma\widetilde{\ell}')$, for every pair of leaves $\{\widetilde{\ell},\widetilde{\ell}'\}$ of $\widetilde{\Lambda}$, it defines an isometric action of $\grperevet$ on the tree $(T,d_T)$ associated to $(\widetilde{\Lambda},\widetilde{\mu})$ defined in Lemma . \blemm\cite[Lem.~17]{Morzy2} For every primitive element $\gamma\in\grperevet-\{e\}$, if $\widetilde{\alpha}_{\gamma}(\RR)$ is a tranlation axis of $\gamma$ in $\revet$ and $\alpha_\gamma$ is the projection of $\widetilde{\alpha}_\gamma$ in $\Sigma$, then the translation distance $\ell_T(\gamma)$ of $\gamma$ in $(T,d_T)$ is equal to $i_{[q]}(\mu,\alpha_\gamma)$. Moreover, if $\ell_T(\gamma)>0$, the translation axis of $\gamma$ is the image in $T$ of the set of leaves of $\widetilde{\Lambda}$ which are interlaced with $\widetilde{\alpha}_\gamma$. \elemm If $m$ is a hyperbolic metric on $\Sigma$, and if $(\Lm,\mu_m)$ is a measured hyperbolic lamination on $\srfcem$, then there exists a usual definition of dual tree to $(\Lm,\mu_m)$ (see for example \cite[§1]{MorSha91}). The dual tree to $(\Lm,\mu_m)$ can also be defined in a similar way to the tree associated to a measured flat lamination, and we easily see that the trees we get by the two procedures are isometric by an equivariant isometry (for the isometric actions of the covering group, see \cite[§.~8]{Morzy2}). \medskip Let $(\Lq,\mu_{[q]})$ be a measured flat lamination on $\srfce$, let $(\Lm,\mu_m)=\psi(\Lq,\mu_{[q]})$, and let $(\Lqr,\mutilde_{[\widetilde{q}]})$ and $(\Lmr,\mutilde_m)$ be their preimages in $\widetilde{\Sigma}$. We denote by $\nu_{\mutilde_{[{q}]}}$ and $\nu_{\mutilde_m}$ the elements of $\M_{\grperevet}([\Gqr])$ and $\M_{\grperevet}([\Gmr])$) they define respectivelly. We assume that $\nu_{\mutilde_{[{q}]}}$ and $\nu_{\mutilde_m}$ are atomless (see \cite[§8]{Morzy2} for the general case). The map $\varphi$ defined at section , defines an $\grperevet$-equivariant map from $\Lmr$ to $\Lqr$, still denoted by $\varphi$. By construction, for every pair of leaves $\{\widetilde{\ell}_0,\widetilde{\ell}_1\}$ of $\Lqr$, we have $B_{\Lmr}(\varphi(\widetilde{\ell}_0), \varphi(\widetilde{\ell}_1))= \varphi(B_{\Lqr}(\widetilde{\ell}_0,\widetilde{\ell}_1))$. Hence the map $\varphi$ defines a map $\varphi_T:(T_{[q]},d_{T_{[q]}})\to(T_{m},d_{T_{m}})$, where $(T_{[q]},d_{T_{[q]}})$ and $(T_{m},d_{T_{m}})$ are respectively the tree associated to $(\Lqr,\mutilde_{[q]})$ and the dual tree to $(\Lmr,\mutilde_m)$. \blemm\cite[Lem.~19]{Morzy2} The map $\varphi_T:(T_{[q]},d_{T_{[q]}})\to(T_{m},d_{T_{m}})$ is a $\grperevet$-equivariant isometry. \elemm \bibliographystyle{alphanum} \bibliography{biblio}{} Département de mathématique, UMR 8628 CNRS, Université Paris-Sud, Bât. 430, F-91405 Orsay Cedex, FRANCE. Bureau : 16. {\it thomas.morzadec@math.u-psud.fr} |
1501.04058 | Title: Power Control and Soft Topology Adaptations in Multihop Cellular
Networks with Multi-Point Connectivity
Abstract: The LTE standards account for the use of relays to enhance coverage near the
cell edge. In a traditional topology, a mobile can either establish a direct
link to the base station (BS) or a link to the relay, but not both. In this
paper, we consider the benefit of multipoint connectivity in allowing User
Equipment (UEs) to split their transmit power over simultaneous links to the BS
and the relay, in effect transmitting two parallel flows. We model decisions by
the UEs as to: (i) which point of access to attach to (either a relay or a
relay and the BS or only the BS); and (ii) how to allocate transmit power over
these links so as to maximize their total rate. We show that this flexibility
in the selection of points of access leads to substantial network capacity
increase against when nodes operate in a fixed network topology. Individual
adaptations by UEs, in terms of both point of access and transmit power, are
interdependent due to interference and to the possibility of over-loading of
the backhaul links. We show that these decisions can converge without any
explicit cooperation and derive a closed-form expression for the transmit power
levels.
Body: \title{{\fontsize{17}{17}\selectfont Power Control and Soft Topology Adaptations in Multihop Cellular Networks with {Multi-Point Connectivity}}} \author{Syed~Amaar~Ahmad and~Luiz~A.~DaSilva,~\IEEEmembership{Senior Member,~IEEE} \thanks{Syed A. Ahmad is currently working with Datawiz Corporation. Luiz A. DaSilva is with the Department of Electrical and Computer Engineering, Virginia Tech and also with CTVR, Trinity College, Dublin, Ireland. Email: $\{$saahmad, ldasilva$\}$@vt.edu.} \thanks{Manuscript received xxxx; revised xxxx}} \markboth{IEEE Transactions on Communications,~Vol.~, No.~, ~} {Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for Journals} \maketitle \begin{abstract} The LTE standards account for the use of relays to enhance coverage near the cell edge. In a traditional topology, a mobile can either establish a direct link to the base station (BS) or a link to the relay, but not both. In this paper, we consider the benefit of {\emph{multipoint connectivity}} in allowing User Equipment (UEs) to split their transmit power over simultaneous links to the BS and the relay, in effect transmitting two parallel flows. We model decisions by the UEs as to: (i) which point of access to attach to (either a relay or a relay and the BS or only the BS); and (ii) how to allocate transmit power over these links so as to maximize their total rate. We show that this flexibility in the selection of points of access leads to substantial network capacity increase against when nodes operate in a fixed network topology. Individual adaptations by UEs, in terms of both point of access and transmit power, are interdependent due to interference and to the possibility of over-loading of the backhaul links. We show that these decisions can converge without any explicit cooperation and derive a closed-form expression for the transmit power levels. \end{abstract} \begin{IEEEkeywords} Resource allocation, hetereogenous networks, coordinated multipoint transmission, dual connectivity \end{IEEEkeywords} \IEEEpeerreviewmaketitle \section{Introduction} \IEEEPARstart{L}{TE} networks anticipate the use of low-cost relays to increase the coverage region of a base station (e{N}ode{B}). In the currently deployed LTE architecture, each User Equipment (UE) can only connect to a single access point at a given time, despite the availability in the cellular system of multiple access points such as relays and e{N}ode{B}s . {In this paper, we consider a flexible two-hop uplink network where UEs have \emph{multi-point connectivity} and choose between transmitting their data via either a single or multiple access points simultaneously, with transmit power adaptations.} In a multihop cellular network, the flexibility to dynamically choose between different access points enables nodes to better overcome the resource limitations introduced by the wireless channel and the relay backhaul link. If not adequately balanced, the link of a relay to the eNodeB (wireless backhaul) may limit the end-to-end data rate of UEs connected to the relay, especially if there is a large number of users. To support a higher load via the relay, a network operator may allocate a larger bandwidth to the backhaul to increase the uplink's capacity. However, in LTE-Advanced, the channels used for the relay backhaul and for the UEs come from the same pool . Hence, such a re-allocation to the relay-to-eNodeB link would also reduce the bandwidth available to the UEs and thus this approach may not lead to an increase in the network-wide performance. Therefore, in lieu of increasing the bandwidth of the relay backhaul, we instead propose to allow a UE to maintain parallel connections to a relay and to the eNodeB, potentially transmitting two independent data streams on two separate channels. \subsection{Contributions} We devise a link-adaptive scheme where UEs are capable of having parallel links to the base station and to the relays. Depending on the load on the relay backhaul, the UE can choose to split its data into two streams, where one is sent directly to the base station and the other one to the relay, or it can send both data streams to a single access point (i.e. either a relay or the base station). Under this scheme, each UE attempts to maximize its own achievable data rate by appropriately selecting its access points (topology adaptation) and by performing distributed power allocation. The power allocation essentially works in two modes: (1) waterfilling or (2) the UE iteratively re-allocates more power on its link to the base station whenever there is over-loading of the backhaul between the relay and the base station. We show significant performance improvement owing to this flexibility to adapt the topology and use power allocation in response to the load on the backhaul links. \subsection{Background} The approach taken in our work overlaps with aspects of cellular topology control and {parallel relay channels} . In a conventional cellular topology, nodes maintain a connection to a single access point, which is often the one with the highest received SNR, and aim to either maintain a target SINR with minimum power or maximize some related utility function . Moreover, several kinds of relaying techniques have been studied for cellular networks to improve coverage and provide better end-to-end performance . Our work incorporates elements of {parallel relay channels} which have been studied in the literature in the context of: (i) a single source-destination node pair with transmit power allocation over the sub-channels without interference from other source nodes , (ii) a Gaussian interference channel for the two-user case (i.e. two source-destination pairs) where the capacity region is explored with a relay backhaul operating on an {out-of-band} channel from the users and (iii) a general multiple access channel from multiple source nodes to a single shared relay and common destination . Our approach differs from each of the above categories in the following ways. Unlike in , nodes maintain parallel links to multiple access points. Also unlike , we assume there can be an arbitrary number of source nodes (i.e. UEs) in the network which may create mutual co-channel interference. The capacity bounds on interference channels with relaying are provided by for certain cases (e.g. 2-user). In contrast, the motivation for our work is how UEs can autonomously make power allocations and select access points by taking into account network-wide conditions to maximize achievable data rate while treating interference as noise. In our work, UE transmissions to the same access point (e.g. base station or relay) are orthogonal in frequency, as we do not assume any underlying multi-user detection (MUD) or interference cancellation capability at the receivers. Moreover, in contrast to , we study a generalized network where there can be any number of noisy and interference channels of unequal bandwidth and where the number of relays may be more than one. {In LTE, a UE may transmit to two access points through either a \emph{coordinated multipoint transmission} approach or a \emph{dual connectivity} approach . In the former, a given bandwidth resource is shared for transmission to the different access points, whereas the latter requires two different bandwidth resources. Our \emph{multipoint connectivity} approach is a hybrid of the two: a UE has two bandwidth resources and it can transmit data to two access points.} In Section II, we present the system model, followed by the problem formulation in Section III. In Section IV, we propose the adaptation scheme for nodes to select their access points and allocate transmit power accordingly. In Section V, we show convergence of the link adaptations under certain constraints and derive a closed-form expression for the converged transmit power. In Section VI, we provide preliminary analysis of how our scheme would compare against cooperative MIMO, interference cancellation and centralized cross-layer optimization. In Section VII, we provide simulation results for our scheme. Finally, Section VIII provides the conclusions of the paper. \section{System Model} We consider a single-cell {uplink} network with a set ${\mathcal{N}}=\{1,2,\cdots, n\}$ of user equipment (UEs) that want to send their data to a base station (BS). Located towards the edges of the cell are $N_r$ fixed relays (RS) that provide coverage for UEs located farther away from the base station (see Fig. 1 with $N_r=3$). {All nodes are equipped with a single antenna}. All Points of Access (PoA) are denoted by a set ${\mathcal{R}}=\{1,2,\cdots,N_r, N_r+1\}$, where elements $1$ to $N_r$ indicate an RS and the BS is represented by element $N_r+1$. The RS decodes the data it receives from the UE and forwards it to the BS (i.e. a decode-and-forward scheme). Each UE can have at most two parallel links. A {node $i$ maintains its connection to a PoA $r_i \in {\mathcal{R}}$ on its first link, which is referred to as the \emph{dedicated-PoA} link. The PoA chosen by each UE on its dedicated-PoA link is the one with the highest received pilot signal strength on a downlink control channel. On its second link (referred to as the \emph{adaptive-PoA} link), node $i$ may choose to send data to the same PoA or to an alternative PoA. The UE can transmit either to the BS or to an RS on its adaptive-PoA link (i.e., $r_i^A \in {\mathcal{R}}$). All links operate on a set of channels ${\mathcal{F}}=\{1,2,\cdots,F\}$. The assignment of each channel $f$ to nodes is represented as a set ${\mathcal{A}}(f) \subseteq {\mathcal{N}}$. To better manage intra-cell interference, some channels are re-used whereas some are exclusively assigned to UEs. We assume that the dedicated-PoA link of a node may experience intra-cell interference, due to frequency re-use, but its adaptive-PoA link operates on a channel which is orthogonal with respect to all other links, so that it is only noise-limited. All links in the network (i.e., UE-to-RS and UE-to-BS) can adapt their transmit power and, consequently, their data rate. Given the UEs' available power vector as ${\bf{P}}_{\max}$ Watts, the transmit power of nodes on their first link are represented by the vector ${\bf{P}}_d=\left[P^{(d)}_{1}, P^{(d)}_{2}, \cdots, P^{(d)}_{n}\right]$ and those on their adaptive-PoA links as ${\bf{P}}_a=\left[P^{(a)}_{1}, P^{(a)}_{2}, \cdots, P^{(a)}_{n}\right]$ where $P^{(d)}_{i}+P^{(a)}_{i}\leq P_{\max, i}$. The channel gain between the UE node $j\in {\mathcal{A}}(f)$ and the intended receiver (PoA) of node $i \in {\mathcal{A}}(f)$ is represented as $h^{(f)}_{j, r_i}$, which depends on several factors such as shadowing, path loss and fading on channel $f\in {\mathcal{F}}$. For the adaptive-PoA link of UE $i$, we denote as $h^{(q)}_{i, r}$ and $h^{(q)}_{i, b}$ its gain to the RS and to the BS, respectively, on channel $q \in \mathcal{F}$. As an example, consider Fig. , where each node has a dedicated-PoA link on channel $1$. {For UE $i$, given ${{x}}^{(d)}_i$ and ${{x}}^{(a)}_i$ as unit-power complex-valued input symbols sent on dedicated-PoA and adaptive-PoA links respectively, the output symbols ${{y}}^{(d)}_i$ and ${{y}}^{(a)}_i$ are: \begin{equation} \small \begin{split} {{y}}^{(d)}_i &= h^{(f)}_{j, r_i}\sqrt{P^{(d)}_i}x^{(d)}_i + \sum_{\forall j\neq i} \sqrt{P^{(d)}_j}h^{(f)}_{j, r_i}x^{(d)}_j+\nu^{(d)}_i \\ &\mbox{, dedicated-PoA link}\\ {{y}}^{(a)}_i &= \sqrt{P^{(a)}_i}h^{(q)}_{i, r}x^{(a)}_i +\nu^{(r)}_i \mbox{, if adaptive-PoA link to RS}\\ {{y}}^{(a)}_i &= \sqrt{P^{(a)}_i}h^{(q)}_{i, b}x^{(a)}_i +\nu^{(b)}_i \mbox{, if adaptive-PoA link to BS} \end{split} \end{equation} where $\nu^{(d)}_{i}$, $\nu^{(b)}_i$ and $\nu^{(r)}_i$ denote the zero-mean complex Gaussian noise on the dedicated-PoA and the adaptive-PoA links respectively.} Corresponding to each channel gain, we define the power gain as $g^{(f)}_{j, r}= |h^{(f)}_{j, r}|^2$. For the dedicated-PoA links, let ${\bf{F}}$ be a normalized cross-link gain matrix of dimensions $n \times n$ \begin{equation} {\bf{F}}(i,j) = \begin{cases} 0 \mbox{ if $i= j$ or $i,j$ orthogonal} \\ \frac{g^{(f)}_{j, r_i}}{g^{(f)}_{i, r_i}} \mbox{ otherwise}, \end{cases} \end{equation} where $i,j$ are orthogonal if their links operate on different channels in ${\mathcal{F}}$.} We also define ${\bf{D}}_d$, ${\bf{D}}_{r}$ and ${\bf{D}}_b$ as $n \times 1$ vectors, which represent normalized noise powers on the dedicated-PoA links and the adaptive-PoA links (to RS and BS) respectively. {The $i^{th}$ elements of these vectors are such that ${\bf{D}}_d(i)= n_{d}/g^{(f)}_{i, r_i}$, ${\bf{D}}_{r}(i)= n_{a}/g^{(q)}_{i, r}$ and ${\bf{D}}_b(i)= n_{a}/g^{(q)}_{i, b}$ where $n_{d} = \mathrm{Var}[{\nu^{d}}_i]$ and $n_{a}= \mathrm{Var}[{\nu^{r}}_i]=\mathrm{Var}[{\nu^{b}}_i]$ are the thermal noise powers respectively}. \newcounter{tempequationcounter} \begin{figure*}[!b] \hrulefill \small \setcounter{tempequationcounter}{\value{equation}} \begin{eqnarray} \setcounter{equation}{5} \begin{split} \eta_N = \sum^{N_r}_{r=1} \min\left(\eta_r, \sum_{\forall i:r_i=r}W_d\eta\left(\gamma^{(d)}_{i}\right)+ \sum_{\forall i: r^A_i=r}\left(1-a_i\right)W_a\eta\left(\gamma^{(r)}_{i}\right)\right) +\sum_{\forall i} a_i W_a\eta\left(\gamma^{(b)}_{i}\right)+\sum_{\forall i: r_i=N_r+1} W_d\eta\left(\gamma^{(d)}_{i}\right) \end{split} \end{eqnarray} \setcounter{equation}{\value{tempequationcounter}} \end{figure*} \newcounter{tempequationcounter2} \begin{figure*}[!b] \small \setcounter{tempequationcounter2}{\value{equation}} \setcounter{equation}{7} \begin{equation} \begin{split} f_i(a_i, P^{(d)}_{i}, P^{(a)}_{i})=\hspace*{7cm}\\ \begin{cases} \left(1-a_i\right)\min\left((V_{r_i}+q_i)^+, W_d\eta\left(\gamma^{(d)}_{i}\right)+W_a\eta\left(\gamma^{(r)}_{i}\right) \right)+a_i \left(W_a\eta\left(\gamma^{(b)}_{i}\right)+\min \left((V_{r_i}+q_i)^+, W_d\eta\left(\gamma^{(d)}_{i}\right) \right)\right)\mbox{, $r_i\leq N_r$}\\ \left(1-a_i\right)\min\left((V_{r^A_i}+q_i)^+, W_a\eta\left(\gamma^{(r)}_{i}\right)\right) + a_iW_a\eta\left(\gamma^{(b)}_{i}\right)+W_d\eta\left(\gamma^{(d)}_{i}\right)\mbox{, $r_i= N_r+1$} \end{cases} \end{split} \end{equation} \setcounter{equation}{\value{tempequationcounter2}} \hrulefill \end{figure*} \newcounter{tempequationcounter3} The effective interference for node $i$ on its dedicated-PoA link and on its adaptive-PoA link is then defined as : \begin{equation} \small E^{(d)}_{i} = \frac{n_d + \sum_{j\neq i, j \in {\mathcal{A}}(f) }g^{(f)}_{j,r_i}P^{(d)}_{j}}{g^{(f)}_{i, r_i}}, \end{equation} \begin{displaymath} \small E^{(b)}_{i} = \frac{n_a}{g^{(q)}_{i, b}} \end{displaymath} \begin{displaymath} \small E^{(r)}_{i} = \frac{n_a}{g^{(q)}_{i, r}}. \end{displaymath} With transmitter-side channel state information (CSIT), each UE knows these effective interferences. The corresponding SINRs are: \begin{eqnarray} \small \begin{split} \gamma^{(d)}_{i} = \frac{P^{(d)}_{i}}{E^{(d)}_{i}}=\frac{P^{(d)}_i}{\frac{n_d + \sum_{j\neq i, j \in {\mathcal{A}}(f)}g^{(f)}_{j, r_i}P^{(d)}_i}{g^{(f)}_{i, r_i}}}\\ \end{split} \end{eqnarray} \begin{displaymath} \small \gamma^{(b)}_{i} = \frac{P^{(a)}_i}{E^{(b)}_{i}}=\frac{P^{(a)}_i}{\frac{n_a}{g^{(q)}_{i, b}}} \\ \end{displaymath} \begin{displaymath} \small \gamma^{(r)}_{i} = \frac{P^{(a)}_i}{E^{(r)}_{i}}=\frac{P^{(a)}_i}{\frac{n_a}{g^{(q)}_{i, r}}} \end{displaymath} where either $\gamma^{(b)}_{i}$ or $\gamma^{(r)}_{i}$ will be the achieved SINR depending on the choice of PoA by the UE on its adaptive-PoA link, and $\gamma^{(d)}_i$ will be the SINR on the dedicated-PoA link. Note that for a single-cell system, the dedicated-PoA links are interference-limited while the adaptive-PoA links are noise-limited. {The total spectrum resource for the UEs in ${\mathcal{F}}$ is fixed at $W_s$ Hz. From this pool, the channel bandwidth $W_d$ Hz of each dedicated-PoA link and that of each adaptive-PoA link $W_a$ Hz depend on the total number of UEs; when $n$ is large, the channel bandwidths $W_d$ and $W_a$ are smaller.} We define the achievable rate on either link as $W\cdot\eta(\gamma)=W\log_2(1+\gamma)$ where $W\in\{W_d, W_a\}$. Finally, the RS-to-BS links operate on \emph{out-of-band} channels from the UEs, and are assumed to have a capacity $\eta_r$ bps. \section{Problem Formulation} If UE $i$ has a dedicated-PoA link to an RS then it has the choice of (a) either transmitting both its data streams to the same RS or (b) sending one stream to the RS and the other one to the BS. Conversely, if UE $i$ has a dedicated-PoA link to the BS, then it can also send one stream to an RS and the other one to the BS or (c) it can transmit both its data streams to the BS. Henceforth, we denote (a),(b) and (c) as RS-RS, BS-RS and BS-BS transmissions, respectively. \begin{table}[t]\caption{Symbol definitions: Subscript $i$ denotes UE $i$.} \begin{center} \begin{tabular}{r c p{0.01cm} } \toprule $a_{i}$& adaptive-PoA link to the BS\\ $P^{(d)}_{i}$& dedicated-PoA link transmit power\\ $P^{(a)}_{i}$& adaptive-PoA transmit power\\ $E^{(d)}_{i}$& effective interference dedicated-PoA link\\ $E^{(b)}_{i}$& effective interference adaptive-PoA (BS)\\ $E^{(r)}_{i}$& effective interference adaptive-PoA (RS)\\ $\gamma^{(d)}_{i}$& SINR dedicated-PoA link\\ $\gamma^{(b)}_{i}$& SNR adaptive-PoA link to BS\\ $\gamma^{(r)}_{i}$& SNR adaptive-PoA link to RS\\ $\tau^{(br)}_{i}$& peak data rate BS-RS transmission\\ $\tau^{(rr)}_{i}$& peak data rate RS-RS transmission\\ $\tau^{(bb)}_{i}$& peak data rate BS-BS transmission\\ $V_r$& rate differential at RS $r$\\ $\widehat{V}$ & rate differential threshold\\ $r_i$& dedicated-PoA link receiver\\ $\eta(\gamma)$& $\log_2(1+\gamma)$\\ $\eta_r$& relay backhaul capacity\\ $q_i$& Data rate sent via relay\\ $W_d$ & Channel bandwidth dedicated-PoA link\\ $W_a$ & Channel bandwidth adaptive-PoA link\\ \bottomrule \end{tabular} \end{center} \end{table} We next define a Boolean variable $a_i \in \{0,1\}$ for a UE $i$. If UE $i$ transmits to the BS on the adaptive-PoA link, then $a_i=1$; otherwise, if the PoA is a relay, $a_i=0$. Using the Max-Flow Min-Cut Theorem , the aggregate end-to-end data rate (network capacity), denoted as $\eta_N$, is formally defined in equation ()\addtocounter{equation}{1}. This expression represents the achievable sum rate in the network when interference is treated as noise. Next, we define the \emph{rate differential} across the RS $r \in \{1,2,\cdots N_r\}$ as \begin{equation} {\smallX \begin{split} V_r = \eta_r-\sum_{\forall i:r_i=r} W_d\eta\left(\gamma^{(d)}_{i}\right)- \sum_{\forall i:r^A_i=r} \left(1-a_i\right)W_a\eta\left(\gamma^{(r)}_{i}\right) \end{split}} \end{equation} which denotes the difference between the outgoing RS-to-BS link capacity and the aggregate incoming data rate of associated UEs at relay $r$. When $V_r < 0$, the RS-to-BS link represents a \emph{bottleneck} link that limits the end-to-end data rate for the UEs sending data to the BS via RS $r$. Conversely, $V_r \geq 0$ indicates that the capacity of the RS-to-BS link is high enough to support the data rate load on the relay. Note that the UE-to-BS links (i.e. when nodes transmit directly to the BS) are also deemed to be bottleneck links. The aggregate end-to-end data rate $\eta_N$ in equation () is the sum of the data rates of all bottleneck links in the network. When UEs adapt transmit power or change PoAs, the data rates at which they can transmit to their RS will vary, thereby causing a change in the rate differential. When $V_r \geq 0$, there is room for UE $i$ to send at a higher data rate to its RS. Conversely, when $V_r < 0$, the UE cannot increase its end-to-end rate by forwarding its data exclusively through the relay. In this case, it may choose to transmit some of its data to the BS directly and allocate its transmit power accordingly. We denote the data rate sent to the relay associated with a UE $i$ as \setcounter{equation}{6} \begin{equation} q_i = \begin{cases} W_d\eta(\gamma^{(d)}_i)+(1-a_i)W_a\eta(\gamma^{(r)}_i) \mbox{, $r_i\leq N_r$}\\ a_iW_a\eta(\gamma^{(r)}_i) \mbox{, $r_i= N_r+1$.} \end{cases} \end{equation} By adjusting the rate differential for $q_i$, we denote $(V_r+q_i)^+$ as the available backhaul capacity at relay $r$ that UE $i$ can achieve given the current load caused by other nodes. We therefore set the adaptation objective of each node $i$ as given by ()\addtocounter{equation}{1}, which reduces in the following manner: \newcounter{tempequationcounterPeakDATA} \begin{figure*}[!b] \normalsize \hrulefill \setcounter{tempequationcounterPeakDATA}{\value{equation}} \setcounter{equation}{9} \begin{equation}{\small \begin{split} \tau^{(rr)}_i=\underbrace{\max}_{P^{(d)}_{i}, P^{(a)}_{i}} W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right)+W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(r)}_{i}}\right) \mbox{, RS-RS transmission $r_i\leq N_r$} \hspace*{6.3cm} \\ \tau^{(bb)}_i=\underbrace{\max}_{P^{(d)}_{i}, P^{(a)}_{i}} W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right)+W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(b)}_{i}}\right)\mbox{, BS-BS transmission $r_i= N_r+1$} \hspace*{5.7cm} \\ \tau^{(br)}_i=\underbrace{\max}_{P^{(d)}_{i}, P^{(a)}_{i}} W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right)+W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(x)}_{i}}\right) \mbox{, BS-RS transmission $x\in \{b, r\}$}\hspace*{6cm} \\ \mbox{$x=b$ if $r_i\leq N_r$, $x=r$ if $r_i=N_r+1$} \hspace*{10cm} \end{split}} \end{equation} \setcounter{equation}{\value{tempequationcounterPeakDATA}} \hrulefill \end{figure*} \addtocounter{equation}{1} \subsubsection{$r_i \leq N_r$} UE $i$ has a dedicated-PoA link to an RS $r_i$ with rate differential $V_{r_i}$. In an RS-RS transmission, the UE transmits on both links to RS $r_i$ only. We thus have $a_i=0$ and () equals $\min\bigg((V_{r_i}+q_i)^+, W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right)+$$W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(r)}_{i}}\right)\bigg)$. Conversely, when node $i$ transmits to the BS on the adaptive-PoA link (i.e. BS-RS transmission), we have $a_i=1$ and () becomes $W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(b)}_{i}}\right) + \min\left((V_{r_i}+q_i)^+, W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right) \right)$. \subsubsection{$r_i =N_r+1$} UE $i$ has a dedicated-PoA link to the BS. When the UE transmits on both links to the BS (i.e. BS-BS transmission), we have $a_i=1$ and () equals $W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right) + W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(b)}_{i}}\right)$. Conversely, when the UE transmits to an RS on its adaptive-PoA link, we have $a_i=0$. This RS is one from which the UE receives the highest pilot signals. Thus, in a BS-RS transmission, equation () becomes $\min\left((V_{r^A_i}+q_i)^+, W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(r)}_{i}}\right) \right)+$\\ $W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right)$, where $V_{r^A_i}$ is the rate differential at the RS. We assume that the relays can broadcast the current values of the rate differentials on a downlink control channel, as in ; the UEs then use the rate differential information to make adaptations. Each UE's decision can then be expressed as the following optimization problem: \setcounter{equation}{8} \begin{equation} \begin{split} \underbrace{\max}_{a_i, P^{(d)}_{i}, P^{(a)}_{i}} f_i(a_i, P^{(d)}_{i}, P^{(a)}_{i}) \hspace*{2cm} \\ P^{(d)}_{i}+P^{(a)}_{i} \leq {P_{\max,i}} \hspace*{2cm} \\ P^{(d)}_{i},P^{(a)}_{i} \geq 0 \hspace*{2.7cm} \\ a_i \in \{0,1\} \hspace*{2.1cm}\\ \end{split} \end{equation} The total transmit power of node $i$ is constrained by $P_{\max,i}$ over both of its links. {We next define expressions for the transmit power allocations that result in the peak (maximum) data rates in \eqref{peakData} for any transmission configuration: BS-RS (base station and relay), RS-RS (relay only) or BS-BS (base station only).} \begin{figure*}[!b] \hrulefill \end{figure*} \newcounter{tempequationcounter5} \begin{figure*}[!b] \normalsize \setcounter{tempequationcounter5}{\value{equation}} \begin{equation*} {P^{(d)}_i(k+1)=}\setcounter{equation}{11} \end{equation*} \addtocounter{equation}{0} \begin{subequations} \small{ \begin{empheq}[left={}\empheqlbrace]{align} &\min\left(P_{\max,i}, \left(\frac{W_dP_{\max,i}-W_aE^{(d)}_i(k)+W_dE^{(r)}_i(k)}{W_a+W_d}\right)^{+}\right) \mbox{, if $a_i(k+1)=0$}\\ &\min\left(P_{\max,i}, \left(\frac{W_dP_{\max,i}-W_aE^{(d)}_i(k)+W_dE^{(b)}_i(k)}{W_a+W_d}\right)^{+}\right) \mbox{, if $a_i(k+1)=1$ and ${V_{r_i}(k)}\geq 0$} \\ &P^{(d)}_{i}(k) \mbox{, if $a_i(k+1)=1$ and ${-\widehat{V}\leq V_{r_i}(k)}< 0$} \\ &z\cdot P^{(d)}_{i}(k) \mbox{, if $a_i(k+1)=1$ and ${V_{r_i}(k)}< -\widehat{V}$} \end{empheq}} \end{subequations} \hrulefill \setcounter{equation}{\value{tempequationcounter5}} \end{figure*} \addtocounter{equation}{1} \begin{proposition} Given $E^{(d)}_{i}$, $E^{(b)}_{i}$ and $ E^{(r)}_{i}$, the peak data rate $\tau^{(rr)}_i$, $\tau^{(br)}_i$ or $\tau^{(bb)}_i$ for UE $i$ is achieved with a water-filling power allocation such that $P^{(d)}_{i} = \min\left(P_{\max,i},\left(\frac{W_d P_{\max,i}+W_dE^{(x)}_{i}-W_aE^{(d)}_{i}}{W_a+W_d}\right)^{+}\right)$ and $P^{(a)}_i=P_{\max,i}-P^{(d)}_i$ where $x \in \{b, r\}$ denotes the adaptive-PoA link to either the BS or RS. \end{proposition} The proof of this is in the Appendix. Note that the above may also represent the power allocation strategy of a UE when it does not have any information about the rate differential and maximizes its data rate on its two links. \begin{corollary} The peak data rate $\tau^{(br)}_i > \tau^{(rr)}_i$ if $E^{(b)}_{i}< E^{(r)}_{i}$, and $\tau^{(br)}_i > \tau^{(bb)}_i$ if $E^{(r)}_{i}< E^{(b)}_{i}$. \end{corollary} The peak data rate achievable for UE $i: r_i\leq N_r$ through a BS-RS transmission is higher than that via RS-RS transmission if the effective interference to the BS is lower than that to the RS on the adaptive-PoA link. The same result applies for UE $i: r_i= N_r+1$ in relation to the peak data rate via BS-BS transmission. \section{Network state-based Distributed Transmission (NDT)} We assume that adaptations occur in time intervals denoted as $k \in \{1,2,..\}$. In interval $k$, the effective interferences $E^{(b)}_{i}(k)$, $E^{(d)}_{i}(k)$, $E^{(r)}_{i}(k)$ and the rate differential $V_r(k)$ at its associated relay are available to each UE $i$. {If the UEs were to apply a greedy algorithm to maximize \eqref{opt_func2}, each UE would make a locally optimal choice of transmit power levels on its two links and PoA selection. Due to the inter-dependence of interference and load on backhaul links, use of a greedy strategy by all UEs may not generally produce a globally optimal solution or converge}. Instead we propose an alternative algorithm, which we call Network state-based Distributed Transmission (NDT), that approaches the objective in \eqref{opt_func2} and which outperforms the greedy algorithm. A positive-valued constant $\widehat{V}$, known to all UEs, is a rate differential threshold that represents some tolerable load at each relay. Under our strategy, a node selects its receiver on the adaptive-PoA link by comparing the available backhaul capacity against the peak data rates in \eqref{peakData} and the threshold $\widehat{V}$. This approach ensures that the achieved data rates do not become erratic as the UE autonomously adapt. \subsection{UE $i: r_i\leq N_r$} We propose that UE $i$, which has a dedicated-PoA link to an RS, update its adaptive PoA as follows: \begin{displaymath} {a_i}(k+1){=} \end{displaymath} \begin{subequations} \small{ \begin{empheq}[left={}\empheqlbrace]{align} & 0, \parbox{15em}{ if ${V_{r_i}(k)}+q_i(k)\geq\tau^{(br)}_i(k)$ \\ and $\tau^{(rr)}_i(k)>\tau^{(br)}_i(k)$}\\[0.1cm] & 1, \parbox{15em}{ if $\tau^{(rr)}_i(k)\leq\tau^{(br)}_i(k)$ \\ or ${V_{r_i}(k)}+q_i(k)\leq\tau^{(br)}_i(k)-\widehat{V}$ } \\[0.1cm] & a_i(k) \mbox{, otherwise.} \end{empheq}} \end{subequations} When ${V_{r_i}(k)}$ is sufficiently high, the choice of PoA between () and () reduces to which transmission (RS-BS or RS-RS) can achieve a higher peak data rate. In \eqref{localadapta}, the UE choses an RS-RS transmission if the current backhaul capacity is large enough and if the achievable rate is more than that via a BS-RS transmission. In (), the UE chooses a BS-RS transmission whenever $\tau^{(rr)}_i(k)\leq \tau^{(br)}_i(k)$ or if the current backhaul capacity is low. The load thresholds of each node are chosen such that the relay backhaul link remains a bottleneck link even after accounting for its own data rate. For all intermediate cases, in (), the PoA on the adaptive-PoA link remains unchanged. Corresponding to PoA choice in iteration $k+1$ in \eqref{aplus}, UE $i$ allocates transmit power to the dedicated-PoA link as shown in \eqref{power_alloc}. The first two cases \eqref{pa} and \eqref{pb} correspond to the waterfilling allocation for RS-RS and BS-RS transmissions, respectively, as indicated in Proposition~. Given $z: 0<z<1$, \eqref{pd} is the power allocation on the dedicated-PoA (UE-to-RS) link when the relay backhaul link is overloaded. The node iteratively reduces its transmit power on its UE-to-RS link and re-allocates this power to its link to the BS. This continues until the backhaul load drops to an acceptable level such that $-\widehat{V}\leq V_{r_i}(k)<0$. As per \eqref{pc} then the UE can maintain its transmit power. Given its transmit power on the dedicated-PoA link, the remaining power is allocated by node $i$ to the adaptive-PoA link such that: \addtocounter{equation}{1} \begin{align} \begin{split} P^{(a)}_{i}(k+1)&=P_{\max,i} - P^{(d)}_{i}(k+1) \\ \end{split} \end{align} \subsection{UE $i: r_i= N_r+1$} Next consider a UE $i$ that has a dedicated-PoA link to the BS. As before, its decision to select the RS on its adaptive-PoA link depends on whether the peak data rate via BS-RS transmission is higher than that via BS-BS transmission and on whether there is sufficient backhaul capacity (i.e. ${V_{r^A_i}(k)}+q_i(k)\geq -\widehat{V}$): \begin{displaymath} {a_i}(k+1){=} \end{displaymath} \begin{subequations} \begin{empheq}[left={{}}\empheqlbrace]{align} & 0, \parbox{15em}{if ${V_{r^A_i}(k)}+q_i(k)> \widehat{V}$ and \\ $\tau^{(br)}_i(k)> \tau^{(bb)}_i(k)$}\\[0.1cm] & 1,\parbox{15em}{if $\tau^{(bb)}_i(k)\leq\tau^{(br)}_i(k)$ or \\ ${V_{r^A_i}(k)}+q_i(k)\leq-\widehat{V}$} \\[0.1cm] & a_i(k) \mbox{, otherwise.} \end{empheq} \end{subequations} If $a_i(k+1)=1$, then node $i$ performs BS-BS transmission with the corresponding transmit power allocation on the dedicated-PoA as \eqref{pb}. Conversely, if $a_i(k+1)=0$ then the node performs BS-RS transmission and allocates its power as per \eqref{pa}. As before, the remaining power is allocated to the adaptive-PoA link as per \eqref{localadapt3}. The key feature of the NDT algorithm is that it gives UEs the flexibility to switch between different transmission modes (i.e. RS-RS versus BS-RS transmission) based on both link-layer and backhaul load conditions. In contrast, in a fixed single connection (RS-RS and BS-BS) or dual connection (BS-RS) topology, the UEs cannot adapt to traffic load and link layer conditions. Moreover, when the backhaul links are over-loaded, conventional waterfilling on the links is not an optimal allocation. Under NDT the UEs re-allocate more transmit power (and data rate) on the direct link to the BS until the backhaul link is load-balanced. \section{Convergence} We analyze the conditions under which the distributed adaptations performed by the nodes under our NDT mechanism will converge to a solution of \eqref{opt_func2}. We define matrix ${\bf{A}}(k) = diag{({[a_1(k), a_2(k), \cdots, a_n(k)]})}$ which is an $n \times n$ diagonal matrix where element $a_i(k)$, as defined earlier, represents whether node $i$ is connected to the BS or to an RS on its adaptive-PoA link. We let $\overline{\bf{A}}={\bf{I}}-{\bf{A}}$ where ${\bf{I}}$ is the identity matrix. For example, ${\bf{A}}(0)={\bf{I}}$ implies that all UEs initially transmit to the BS on their adaptive-PoA links. The vector representation of the effective interference in () in iteration $k$ is then \begin{equation} \begin{split} {\bf{E}}_d(k) &= {\bf{D}}_d + {\bf{F}}{\bf{P}}_d(k) \mbox{, (dedicated-PoA links)},\\ {\bf{E}}_{a}(k) &= {\bf{A}}{\bf{D}}_{b} + \overline{\bf{A}}{\bf{D}}_{r} \mbox{, (adaptive-PoA links)} \end{split} \end{equation} where ${\bf{E}}_{a}(k)(i)={\bf{D}}_b(i)$ when $a_i(k)=1$ and ${\bf{E}}_{a}(k)(i)={\bf{D}}_{r}(i)$ when $a_i(k)=0$. \subsection{High $\eta_r$ regime} Note that under NDT when the capacity of the RS-to-BS backhaul links is sufficiently high, the choice of PoA of UEs depends only on the peak data rates achievable through the BS-RS, RS-RS or BS-BS transmissions. As per Corollary~, this in turn depends on the effective interferences on their adaptive-PoA links. \begin{corollary} Given a large enough RS-to-BS link capacity $\eta_r$, $a_i=1$ when ${\bf{D}}_{b}(i)<{\bf{D}}_{r}(i)$ for each UE $i$ and conversely $a_i=0$ when ${\bf{D}}_{b}(i)\geq {\bf{D}}_{r}(i)$, regardless of the initial PoA assignments. \end{corollary} \begin{proof} Suppose that there is some $\eta_r$ which is more than the maximum aggregate data rate the UEs can ever forward through the relay. Under NDT, the choice of a PoA by each UE depends on the peak data rate it can achieve via waterfilling allocation on the two links. Recall that as per Corollary~, $\tau^{(rr)}_i< \tau^{br}_i$ when $E^{(b)}_i< E^{(r)}_i$ for UE $i:r_i\leq N_r$. Likewise, $\tau^{(bb)}_i< \tau^{br}_i$ when $E^{(b)}_i< E^{(r)}_i$ for UE $i:r_i=N_r+1$. If ${\bf{D}}_b(i)< {\bf{D}}_{r}(i)$ for UE $i$ then $E^{(b)}_i(k) < E^{(r)}_i(k)$ always. Hence the UE chooses the BS as the PoA under NDT on the adaptive-PoA link. Conversely, $E^{(b)}_i(k) \geq E^{(r)}_i(k)$ when ${\bf{D}}_b(i)\geq {\bf{D}}_{r}(i)$ and the UE will choose the RS. This applies regardless of which PoA the UE initially attaches to. \end{proof} Given a high enough backhaul capacity, we now derive a solution of \eqref{opt_func2} for each UE in the network under NDT. Corresponding to Corollary~, the choice of PoAs of all UEs on their adaptive-PoA links can be represented by ${\bf{A}}^{*}= diag {[a_1, a_2, \cdots a_n ]}$. \begin{theorem} The transmit powers converge to ${\bf{P}}_d^*=\frac{1}{W_a+W_d}\left[{\bf{I}}+\frac{W_a{\bf{F}}}{W_a+W_d}\right]^{-1}{\bigg[}{W_d\bf{P}}_{\max}-W_a{\bf{D}}_{d}+W_d{\bf{A}}^*{\bf{D}}_{b}+W_d\overline{\bf{A}}^*{\bf{D}}_{r}{\bigg]}$ given a large enough RS-to-BS link capacity $\eta_r$ and if ${\bf{0}}\leq {\bf{P}}_d^*\leq {\bf{P}}_{\max}$. \end{theorem} \begin{proof} For some large $\eta_r$, the rate differential at each relay is larger than the thresholds set in () and in (). Hence, the choice of PoA on the adaptive-PoA link depends only on the peak data rates $\tau^{(rr)}_i$ and $\tau^{(br)}_i$ if $r_i \leq N_r$. Conversely, if $r_i=N_r+1$ then the choice is between $\tau^{(bb)}_i$ and $\tau^{(br)}_i$. Hence, regardless of the transmit powers, as per Corollary~, let ${\bf{A}}^*$ represent the choice of PoAs given the normalized noise powers on the channels for adaptive-PoA links: the nodes allocate transmit power according to either \eqref{pa} or \eqref{pb}. We represent the power updates for the dedicated-PoA links in matrix notation as \begin{displaymath} {\bf{P}}_d(k+1)=\frac{\left[W_d{\bf{P}}_{\max}-W_a{\bf{E}}_d(k)+W_d{\bf{E}}_{a}(k)\right]}{W_a+W_d} \end{displaymath} \begin{displaymath} \hspace*{0.9cm}=\frac{1}{W_a+W_d}{[}{W_d\bf{P}}_{\max}-W_a{\bf{D}}_d - W_a{\bf{F}}{\bf{P}}_d(k)+\\ \end{displaymath} \begin{displaymath} \hspace{0.12in}W_d{\bf{A}}^*{\bf{D}}_{b} + W_d \overline{{\bf{A}}^*}{\bf{D}}_{r}{]}\\ \end{displaymath} \begin{equation} \begin{split} &=\frac{{[}{W_d\bf{P}}_{\max}-W_a{\bf{D}}_{d}+W_d{\bf{A}}^*{\bf{D}}_{b}+W_d\overline{\bf{A}}^*{\bf{D}}_{r}{]}}{{W_a+W_d}}\\ &\hspace{0.15cm} - \frac{W_a}{W_a+W_d}{\bf{F}}{\bf{P}}_d(k) \\ &={\bf{N}}-{\bf{M}}{\bf{P}}_d(k) \end{split} \end{equation} where ${\bf{N}}= \frac{1}{W_a+W_d}{[}W_d{\bf{P}}_{\max}-W_a{\bf{D}}_{d}+W_d{\bf{A}}^*{\bf{D}}_{b}+W_d\overline{\bf{A}}^*{\bf{D}}_{r}{]}$ and ${\bf{M}}= \frac{W_a}{W_a+W_d}{\bf{F}}$. The above evolves to: \begin{equation} \begin{split} {\bf{P}}_d(k+1)&={\bf{N}}-{\bf{M}}\left({\bf{N}} - {\bf{M}}\left({\bf{N}}-{\bf{M}}\left(\cdots {\bf{P}}(0) \right) \right)\right) \\ &={[}{\bf{I}} - {\bf{M}}+{\bf{M}}^2-\cdots{]}{\bf{N}}+{\bf{M}}^{k}{\bf{P}}_d(0)\\ \lim\limits_{k \rightarrow \infty} {\bf{P}}_d(k+1)&={\bf{P}}^*_d=\left[{\bf{I+M}}\right]^{-1}{\bf{N}}\\ &=\frac{1}{W_a+W_d}\left[{\bf{I}}+\frac{W_a{\bf{F}}}{W_a+W_d}\right]^{-1}\\ {\bigg[}W_d{\bf{P}}_{\max} - &W_a{\bf{D}}_{d}+W_d{\bf{A}}^*{\bf{D}}_{b}+W_d\overline{\bf{A}}^*{\bf{D}}_{r}{\bigg]} \end{split} \end{equation} if the spectral radius (maximum absolute eigenvalue) of $\frac{W_a}{W_a+W_d}{\bf{F}}$ is less than one and where ${\bf{P}}_d(0)$ is the initial transmit power vector of the dedicated-PoA links. Likewise, the corresponding transmit powers on the adaptive-PoA channels are simply the difference between ${\bf{P}}_{\max}$ and (). They evolve to $\lim\limits_{k \rightarrow \infty} {\bf{P}}_a(k)={\bf{P}}_a^*={\bf{P}}_{\max}- {\bf{P}}^*_d$. \end{proof} \subsection{Limited $\eta_r$ regime} Next we consider the general case where the backhaul capacity is limited. \begin{theorem} Given a large enough tolerable $\widehat{V}$, NDT adaptations will converge if the spectral radius of ${\bf{F}}$ is less than $\frac{W_a+W_d}{W_a}$. \end{theorem} \begin{proof} According to Corollary~ the inequality between the peak data rates $\tau^{(rr)}_i$, $\tau^{(br)}_i$ or $\tau^{(bb)}_i$ always holds given the inequality relationship between ${\bf{D}}_b(i)$ and ${\bf{D}}_r(i)$ for each UE $i$. Thus, given a high enough $\widehat{V}$, the nodes will not cycle back and forth their links between different PoAs as per \eqref{aplus} and \eqref{aplus2}. The power allocation of UE $i: r_i= N_r+1$ will only follow the waterfilling allocation either in \eqref{pb} or \eqref{pa}. Likewise, if UE $i: r_i\leq N_r$, then this node will either perform waterfilling allocation as per \eqref{pa}\eqref{pb} or maintain its transmit power from the preceding iteration as per \eqref{pc}. Thus, the evolution of transmit power for the system is equivalent to that in \eqref{powerseries} except that some nodes may not update their transmit power every iteration (i.e. their transmit power level is based on an update from an earlier iteration). The power updates can therefore be described as an \emph{asynchronous iterative} system . It is known that such a system converges if the iterative matrix $\bf{M}$ in \eqref{powerseries} has a spectral radius less than one (corresponding to a spectral radius of less than $\frac{W_a+W_d}{W_a}$ for ${\bf{F}}$). \end{proof} \newcounter{optimalN} \begin{figure*}[!b] \hrulefill \setcounter{optimalN}{\value{equation}} \setcounter{equation}{17} \begin{equation}{\small \begin{split} \eta^{(A)}_N &= \underset{P^{(d)}_i P^{(a)}_i a_i}{\max}{\hspace{0.2cm}} {W_d\sum^{n=2}_{i=1}\log_2(1+P^{(d)}_i\frac{\sigma^2_i}{n_d})+W_a\left(\sum^{n=2}_{i=1} \eta\left( P^{(a)}_i\frac{g_{i,r}+g_{i,b}}{n_a}\right)\right)} \hspace*{0cm} \\ \eta^{(B)}_N &= \underset{P^{(d)}_i P^{(a)}_i a_i}{\max}\hspace{0.2cm} W_d\eta\left(\frac{g^{(1)}_{1,1}P^{d}_1}{n_d}\right) + W_d\eta\left(\frac{g^{(1)}_{2,2}P^{(d)}_2}{n_d+g^{(1)}_{1,2}P^{(d)}_1}\right) + W_a\left(\sum^{n=2}_{i=1} (1-a_i)\eta\left(\frac{g_{i,r} P^{(a)}_i}{n_a}\right)+\sum^{n=2}_{i=1} a_i\eta\left(\frac{g_{i,b} P^{(a)}_i}{n_a}\right)\right) \hspace*{0.4cm} \\ \eta^{(C)}_N &= \underset{P^{(d)}_i P^{(a)}_i a_i}{\max}{\hspace{0.1cm}}W_d \sum^2_{i=1} \eta\left(\frac{g^{(1)}_{i,i}P^{d}_i}{n_d+ \sum^2_{j=1;j\neq i} g^{(1)}_{j,i}P^{d}_j}\right) +W_a\left(\sum^{n=2}_{i=1} (1-a_i)\eta\left(\frac{g_{i,r}P^{(a)}_i}{n_o}\right)+\sum^{n=2}_{i=1} a_i\eta\left(\frac{g_{i,b} P^{(a)}_i}{n_o}\right)\right)\hspace*{1.2cm} \\ P^{(d)}_i+P^{(a)}_i\leq P_{max,i}, \hspace*{-8.8cm}\\ P^{(d)}_i,P^{(a)}_i\geq 0, \hspace*{-8cm}\\ a_i \in \{0,1\}, \hspace*{-8.5cm}\\ i\in\{1,2\} \hspace*{-8.6cm} \end{split}} \end{equation} \setcounter{optimalN}{\value{equation}} \hrulefill \end{figure*} \subsection{Practical Considerations} {In wireless systems, feedback delay and estimation error may render CSIT imperfect . The adaptations will thus be based on inaccurate CSIT that may render our scheme less effective in improving network capacity. Given the estimated gain $\widehat{g^{(f)}_{i,r}}$ for UE $i$, estimation error will project the effective interference vectors in \eqref{effintfvec} to: \begin{equation} \begin{split} \widehat{{\bf{E}}_d}(k) &= \widehat{{\bf{D}}_d} + \widehat{{\bf{F}}}{\bf{P}}_d(k),\\ \widehat{{\bf{E}}_{a}}(k) &= {\bf{A}}\widehat{{\bf{D}}_{b}} + \overline{\bf{A}}\widehat{{\bf{D}}_{r}} \end{split} \end{equation} Thus, corresponding to the imperfect channel knowledge, if the spectral radius of $\widehat{{\bf{F}}}$ is less than $\frac{W_a+W_d}{W_a}$ then the results in Corollary~, Theorem~ and still hold. Therefore, under NDT, adaptation would converge with the system vectors adjusted for estimation error. Moreover, UEs could switch between different transmission modes every iteration using the multipoint signaling feedback mechanism provided in the LTE architecture . Hence, implementing our proposed scheme will not require any additional overhead.} \section{Performance Optimality} We now discuss some relevant notions of optimality based on either interference cancellation or cooperative communication with centralized optimization, so as to provide achievable rate comparison with NDT. As in , we also consider a network with $n=2$ UEs, and $N_r=1$ relay. The first UE has a dedicated-PoA link to the BS and the other has a dedicated-PoA link to the RS. The two dedicated-PoA links operate on the same channel. The backhaul capacity $\eta_r$ is assumed high enough such that it does not become rate-limiting. \subsection{Cooperative MIMO (CO-MIMO)} As shown in , a multi-user cooperative network can be viewed as a virtual MIMO system. In our system, the two UEs collaborate as the transmitter set and the PoAs (BS and the RS) as the joint receivers in a $2\times 2$ MIMO. Given the channel matrix ${\bf{H}}=\begin{bmatrix} h^{(1)}_{1,1} & h^{(1)}_{1,2} \\[0.3em] h^{(1)}_{2,1} & h^{(1)}_{2,2} \\[0.3em] \end{bmatrix}$, {let its $i^{th}$ singular value obtained via Singular Value Decomposition (SVD) be $\sigma_i$.} The maximum sum rate $\eta^{(A)}_N$ is expressed in \eqref{Optimality}, where the left-most term denotes the {equivalent MIMO channel capacity using SVD of the dedicated-PoA links and the other summation term indicates the rate achieved through maximal ratio diversity combining on the adaptive-PoA links.} \subsection{Asymmetric Interference Cancellation (AIC)} In AIC, the BS can perform interference cancellation whereas the RS does not have any such capability. The maximum sum rate is bounded by $\eta^{(B)}_N$ in \eqref{Optimality} where the left-most term denotes the achievable rate for the dedicated-PoA UE-to-BS link when all interference received from the UE-to-RS dedicated link can be eliminated . \subsection{Cross-Layer Optimization (C-OPT)} In C-OPT, the optimal choice of PoAs and power allocations is determined to maximize the sum rate without any cooperation or interference cancellation and is expressed as $\eta^{(C)}_N$ in \eqref{Optimality}. \\ In the next section, through numerical results we compare the performance under CO-MIMO, AIC and C-OPT against NDT. \section{Simulation Results} In this section, for a multihop cellular network we use Matlab-based simulation to determine the aggregate end-to-end data rate $\eta_N$ over a range of control parameters. We assume that the cross-link power gains {of the UEs} ($g^{(f)}_{i, r_i}, g^{(q)}_{i, b}$, etc.) are of the form $\kappa^{(f)}_{j, i}\cdot d_{j, i}^{-\alpha}$ where $\kappa^{(f)}_{j, i}$ is an exponentially distributed random variable with unit variance (due to Rayleigh fading) on channel $f$ and $\alpha$ is the path loss exponent. The fading gains are assumed independent and identically distributed for all links. There are $50$ power control iterations in each simulation trial. The transmit powers of pilot signals from the BS and RS on the downlink control channel are assumed equal. The UEs are randomly and uniformly located in cluster regions of radius $R_L$ around each of the $N_r$ relays with $P_{\max,i}=1.0$ Watts. Unless otherwise stated, we assume the following parameters: for NDT $z=0.8$, $\widehat{V}=5$ Mbps, $R_L=250$ m, {noise power spectral density is $-200$ dBW/Hz (proportional to channel bandwidth ), $W_s = 10$ MHz and $\alpha=4.0$.} The backhaul links operate on reserved bands from an operator's pool that are non-overlapping with channels used by the UEs. {We use the backhaul channel model based on the 3GPPTR 36.814 specifications . The line-of-sight (LOS) and non-line-of-sight (NLOS) path losses of each wireless relay backhaul link are based on the urban macro model based on carrier frequency $2$ GHz, relay height $5$ m with all remaining parameters (e.g. transmit power, bandwidth or lognormal shadowing variance etc.) based on standard values specified in \cite[p.~72, p.~94]{3gpp2010}. The backhaul capacity $\eta_r$ is then determined based on Shannon capacity.} When $n$ is large, there is less bandwidth allocated to each UE. The division of the allocated bandwidth between the dedicated-PoA and adaptive-PoA links for each UE is set at $\frac{W_a}{W_d}=0.5$, unless stated otherwise. There are $N_r=4$ relays located randomly in a $2$ km by $2$ km square region around the BS with non-overlapping cluster regions. \subsection{Convergence} As an example of convergence, we first consider the network in Fig. with $n=3$ UEs located in the positions shown. We plot the evolution of transmit powers given that initially all UEs transmit on their adaptive-PoA link to their relays with a randomly chosen initial vector ${\bf{P}}_d(0)$. In Fig. (a), we observe that the transmit powers converge to the vector predicted in () where ${\bf{A}}^*=diag [0,0,0,1]$. In Fig. (b), we plot time evolution of the network-wide performance $\eta_N$ for two sample networks. We compare NDT against approaching \eqref{opt_func2} through a {greedy algorithm}, which represents an immediate solution for $a_i(k+1), P^{(d)}_i(k+1)$ and $P^{(a)}_i(k+1)$ for UE $i$. The greedy approach, in contrast to NDT, results in an erratic evolution of the transmit powers, which correspondingly causes an unstable and reduced network capacity over time. In Fig. , we plot the effects of the rate differential threshold $\widehat{V}$ on network capacity and the corresponding likelihood of convergence for those trials where the spectral radius $\rho(\bf{F})$ is less than $\frac{W_a+W_d}{W_a}$ (recall Theorem~). With a higher $\widehat{V}$, NDT will converge $100\ Note that in the intermediate case, when the backhaul capacity is limited and $\widehat{V}$ is small, then convergence cannot always be guaranteed as shown in Fig. a. For instance, to alleviate load on the backhaul links, UEs can send more data rate directly to the BS. Thus, if $\widehat{V}$ is set too small the data rates sent via relays decrease too fast and the backhaul links may become under-loaded. In turn this will induce the UEs to switch transmission mode and send all their data via the relays. Thus, with a small $\widehat{V}$, there may be oscillation between over-loading the backhaul links or under-utilizing them as UEs make local adaptations with instantaneous knowledge of the system. Nonetheless, we observe in Fig. b that even in these cases the likelihood of non-convergence is still relatively small. \subsection{Network Capacity} {We compare NDT against two schemes: \\ (1) \emph{Single-PoA Waterfilling (SP-WF)}: As in each node has a single PoA; if attached to an RS, it allocates transmit power as per \eqref{pa}\eqref{localadapt3} whereas if attached to the BS it uses \eqref{pb}\eqref{localadapt3}. \\ (2) \emph{BS-PoA Waterfilling (BS-WF)}: UEs at cell edge transmit to RS and BS both as in , whereas the remaining UEs only transmit to the BS . Either type of UE allocates its transmit power as per \eqref{pb} and \eqref{localadapt3}. } Fig. shows that, in contrast to BS-WF or SP-WF, NDT offers the best of both approaches over the entire range of the parameters $n$ and $\eta_r$. Under NDT, when the load is lighter (either $\eta_r$ is larger or $n$ is smaller), the UEs only transmit to a single PoA whereas, conversely, they switch to transmitting to the BS on their adaptive-PoA links when the load is heavier. Another observation is that, when the load is large, under NDT the network performance is better than for BS-WF. This is because under NDT, a UE uses power control in \eqref{pd} and iteratively re-allocates more power on the direct link to the BS, instead of simply waterfilling over its channels. {Note that in Fig. b, we obtain the mean $\eta_r$ by varying the bandwidth allocated to each relay backhaul link in the operator pool up to 30 MHz.} In Fig. , we plot the performance by varying the path loss exponent $\alpha$ and cluster radius $R_L$. We observe that, while performance for all schemes declines with an increase in either parameter, NDT still offers the best of any of the schemes. In Fig. (b), with increasing $R_L$, the inter-UE distance grows and the nodes become dispersed over a wider region, and thus more distant from their receivers. While the performance for NDT declines with increasing $R_L$ (due to greater path loss), it is still better than that of other schemes. In Fig. (a) we consider the case with a large number of relays. Since the RS-to-BS links operate on non-UE channels, the 2-hop network would resemble the situation when picocells coexist within a macro-cellular system. Moreover, the backhaul capacity of the picocells is constrained and is variable. Hence, in this scenario the relays of the 2-hop network could resemble picocell base stations. The UEs are evenly spread between the PoAs and where the maximum RS-BS separation is $R=1000$ m and $R_L=100$ m. An implication from the result is that when there are more small cells and the UEs can transmit to both the macro-cellular BS and the picocell BS then there are performance gains over when they are constrained to transmit only to picocell base stations. In Fig. (b), we consider the impact of the channel bandwidth size of the adaptive-PoA links relative to the dedicated-PoA links. When $W_a$ is large the network capacity of all schemes becomes better. Finally, in Fig. b, we compare network capacity under NDT against the mechanisms discussed in Section VII, for a 2-user 1-relay system. Each UE is located exactly $R_L$ m from its dedicated PoA perpendicular to the line representing BS-RS separation (of 2000 m). The CO-MIMO and AIC schemes essentially decode interference apart from the desired signal, which enables them much to achieve much higher rates than both NDT and cross-layer optimization. However, we observe that if the system becomes noise-limited or when the relative bandwidth of the adaptive-PoA links is larger, NDT approaches the optimal achievable performance. This is because more noise or a larger $W_a$ decreases the relative advantage of interference cancellation capability of AIC and CO-MIMO. \section{Conclusion} We have proposed, for a 2-hop cellular uplink network, a link adaptive scheme where nodes have the flexibility to transmit to multiple PoAs. Under our scheme, nodes either transmit to a single PoA or split their data streams between the base station and a relay, using CSIT and backhaul load condition. We demonstrate a significant improvement in the aggregate end-to-end data rate in the network due to the proposed adaptation scheme. Future work could consider coordination between the UEs to form cooperative MIMO links and the impact of an adaptive bandwidth allocation to the backhaul links. \section*{Acknowledgement} We would like to thank Dr Ravi Tandon for constructive feedback. This work is partially supported by the Science Foundation Ireland under Grant No. 10/IN.1/I3007. \appendix \subsubsection*{Proof of Proposition~} Consider the peak RS-RS transmission data rate $\tau^{(rr)}_i$ for UE $i$: \begin{eqnarray*} \begin{split} \underbrace{\max}_{P^{(a)}_{i}, P^{(d)}_{i}} W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(r)}_{i}}\right)+W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right)\\ P^{(a)}_{i}+P^{(d)}_{i}\leq P_{\max,i}\hspace*{1in}\\ P^{(a)}_{i}, P^{(d)}_{i} \geq 0 \hspace*{1.3in} \end{split} \end{eqnarray*} Using Lagrangian multipliers and KKT conditions we obtain: \begin{eqnarray*} \small \begin{split} &\underbrace{\max}_{P^{(d)}_{i}, P^{(a)}_{i}} W_d\log_2\left(1+\frac{P^{(d)}_{i}}{E^{(d)}_{i}}\right) + W_a\log_2\left(1+\frac{P^{(a)}_{i}}{E^{(a)}_{i}}\right)\\ &-\lambda_1(P^{(d)}_{i}+P^{(a)}_{i}- P_{\max,i})-\lambda_2 P^{(d)}_{i} -\lambda_3 P^{(d)}_{i}\\ &\frac{W_d}{\ln(2)}.\frac{1}{P^{(d)}_{i}+E^{(d)}_{i}}-\lambda_1-\lambda_2 =0\\ &\frac{W_a}{\ln(2)}.\frac{1}{P^{(a)}_{i}+E^{(a)}_{i}}-\lambda_1-\lambda_3 =0\\ &\lambda_1(P^{(d)}_{i}+P^{(a)}_{i}- P_{\max,i}) =0\\ &\lambda_2 P^{(d)}_{i}=0\\ &\lambda_3 P^{(a)}_{i}=0 \end{split} \end{eqnarray*} Case 1: With $\lambda_2=\lambda_3=0, \lambda_1 \neq 0$, we obtain: \begin{eqnarray*} \begin{split} &\frac{W_d}{\ln(2)}.\frac{1}{P^{(d)}_{i}+E^{(d)}_{i}}-\frac{W_a}{\ln(2)}.\frac{1}{P_{\max,i}-P^{(d)}_{i}+E^{(a)}_{i}}=0\\ & P^{(d)}_{i} = \frac{W_d P_{\max,i}+ W_d E^{(a)}_{i}- W_a E^{(d)}_{i}}{W_d+W_a}\\ & P^{(a)}_{i} ={P_{\max,i}-P^{(d)}_i}\\ & =\frac{W_d P_{\max,i}-W_d E^{(a)}_{i}+ W_a E^{(d)}_{i})}{W_d+W_a} \end{split} \end{eqnarray*} Case 2: With $\lambda_3 \neq 0, \lambda_2 = 0$ (i.e. $P^{(d)}_{i}=0$) we obtain: \begin{eqnarray*} \begin{split} \lambda_1=\frac{W_d}{\ln(2)}.\frac{1}{P^{(d)}_{i}+E^{(d)}_{i}}\\ P^{(d)}_{i}=0\\ P^{(a)}_{i}+0=P_{\max,i} \end{split} \end{eqnarray*} We thus have $P^{(d)}_{i} = \min\left(P_{\max,i},\frac{\left(W_dP_{\max,i}-W_a E^{(d)}_{i}+ W_d E^{(r)}_{i}\right)^+}{W_a+W_d}\right)$. The peak BS-RS or BS-BS transmission data rates $\tau^{(br)}_i$ and $\tau^{(bb)}_i$ have the same allocations as their maximization is identical. \bibliography{References} |
1501.04060 | Title: Analysis of Quantum Particle Automata for Solving the Density
Classification Problem
Abstract: To advance our understanding of Quantum Cellular Automata in problem solving
through parallel and distributed computing, this research quantized the density
classification problem and adopted the Quantum Particle Automata (QPA) to solve
the quantized problem. In order to solve this problem, the QPA needed a unitary
operator to carry out the QPA evolution and a boundary partition to make the
classification decisions. We designed a Genetic Algorithm (GA) to search for
the unitary operators and the boundary partitions to classify the density of
binary inputs with length 5. The GA was able to find more than one unitary
operator that can transform the QPA in ways such that when the particle was
measured, it was more likely to collapse to the basis states that were on the
correct side of the boundary partition for the QPA to decide if the binary
input had majority density 0 or majority density 1. We analyzed these solutions
and found that the QPA evolution dynamic was driven by a particular parameter
$\theta$ of the unitary operator: a small $\theta$ gave the particle small mass
hence fast evolution while large $\theta$ had the opposite effect. While these
results are encouraging, scaling these solutions for binary inputs of arbitrary
length of $n$ requires additional analysis, which we will investigate in our
future work.
Body: \begin{center} \textbf{\LARGE Analysis of Quantum Particle Automata for Solving the Density Classification Problem}\\ \vspace{0.5cm} \textbf{Tina Yu and Radel Ben-Av } \end{center} \begin{abstract} \noindent To advance our understanding of Quantum Cellular Automata in problem solving through parallel and distributed computing, this research quantized the density classification problem and adopted the Quantum Particle Automata (QPA) to solve the quantized problem. In order to solve this problem, the QPA needed a unitary operator to carry out the QPA evolution and a boundary partition to make the classification decisions. We designed a Genetic Algorithm (GA) to search for the unitary operators and the boundary partitions to classify the density of binary inputs with length 5. The GA was able to find more than one unitary operator that can transform the QPA in ways such that when the particle was measured, it was more likely to collapse to the basis states that were on the correct side of the boundary partition for the QPA to decide if the binary input had majority density 0 or majority density 1. We analyzed these solutions and found that the QPA evolution dynamic was driven by a particular parameter $\theta$ of the unitary operator: a small $\theta$ gave the particle small mass hence fast evolution while large $\theta$ had the opposite effect. While these results are encouraging, scaling these solutions for binary inputs of arbitrary length of $n$ requires additional analysis, which we will investigate in our future work. \end{abstract} \section{Introduction} The density classification problem is an important test case to measure the computational power of Cellular Automata (CA). In the CA computational model, there is a grid of cells where the states of the cells are updated synchronously according to a local rule. The density classification problem was proposed to study one-dimensional classical CA, where each cell contains a state of 0 or 1, and the task is to identify if the majority of the cells is 0 or 1. A solution to this problem is a local rule that converges the CA to a fixed point of all 1s if its initial configuration contains more 1s than 0s and to all 0s if its initial configuration contains more 0s than 1s. Such convergence normally takes place within \textit{M} time steps, where in general \textit{M} depends on the length of the lattice \textit{L} (which is assumed to have periodic boundary conditions, resulting in a circular grid). Over the years, various CA local rules have been proposed to demonstrate that classical CA can achieve global synchronization through parallel local interactions. These results indicate that the CA computational model is well suited for distributed and parallel computing. Meanwhile, extensions of classical CA to Quantum CA (QCA) for distributed and parallel computing have also been investigated . Unlike classical information of binary 0 \emph{or} 1, quantum information resides in superpositions of 0 \emph{and} 1 simultaneously. Computation takes place on each of the superpositions following a distinct path, called "quantum parallelism" . Therefore, there is a natural mapping between the parallel computation on \textit{quantum superpositions} and the parallel processing of classical \textit{CA cells}. In terms of distributed computing, Feynman and others have shown that quantum formalism of closed, locally interacting microscopic systems are able to perform universal computation. To advance our understanding of QCA in problem solving through parallel and distributed computing, this research quantized the density classification problem and adopted the Quantum Particle Automata (QPA) devised by Meyer to solve this problem. The QPA is a one-dimensional CA that consists of a \emph{single particle}. Similar to the classical CA, the state of each cell in the QPA at a given time step depends on the states of the cells in some local neighborhood at the previous time step. However, the evolution of the QPA is quantum mechanical. More precisely, unlike classical CA where the state of each cell $x_i$ is a binary value of 0 or 1, the state of a QPA cell is a (complex) probability amplitude $c_i$ for the particle being in state $|x_i\rangle$ when it is measured (being in state $|x_i\rangle$ means being in position $x_i$). The state of the QPA $|\psi\rangle$ is a linear combination of all $L$ possible states, where $L$ is the length of the lattice: \begin{align*} |\psi\rangle=c_0|x_0\rangle+c_1|x_1\rangle+\cdots+c_{L-1}|x_{L-1}\rangle \end{align*} Moreover, the QPA local rule that is used to update the probability amplitude in each cell is a unitary operator. Since the transition of the QPA cells is unitary, the total probability, i.e. the sum of the norm squared of the probability amplitude at each cell, is preserved. To solve the quantized density classification problem, the QPA needs a unitary operator to carry out the QPA evolution and a boundary partition to make the classification decisions. This research applied Genetic Algorithms (GA) to discover the unitary operators and the boundary partitions for the QPA to classify the density of binary inputs with length 5. GA is a population-based search algorithm that has been widely used in optimization and machine learning . Through the simultaneous search of a population of candidate solutions, the GA was able to discover more than one unitary operator that can transform the QPA in ways such that when the particle was measured, it was more likely to collapse to the basis states on the correct side of the boundary partition for the QPA to decide if the binary input had density 0 or density 1. We analyzed these solutions and found that the QPA evolution dynamic was driven by a particular parameter $\theta$ of the unitary operator: a small $\theta$ gave the particle small mass hence fast evolution while large $\theta$ had the opposite effect. While these results are encouraging, scaling these solutions for binary inputs of arbitrary length of $n$ requires additional analysis, which we will investigate in our future work. In addition to being the first to investigate QCA in solving the density classification problem, this research also made the following contributions: \begin{itemize} \item It devised a quantum version of the density classification problem that can be used to represent the problem for binary inputs of any length $n$. \item It demonstrated that for binary inputs of length 5, there are multiple solutions to the quantized density classification problem and the GA methodology can find many of these solutions. \item It analyzed these solutions and showed that the QPA evolution dynamic is driven by a particular parameter $\theta$ of the unitary operator: a small $\theta$ gives the particle small mass hence fast evolution while large $\theta$ has the opposite effect. \end{itemize} The rest of the paper is organized as follows. Section explains classical CA, quantum CA and the QPA that we used to conduct our research. In Section , we first review the local rules proposed to solve the classical density classification problem and then quantize the problem for our QPA to solve this problem for binary inputs of length 5. Section describes the GA system we designed to discover the unitary operator and the boundary partition solutions. The results are then presented in Section . In Section , we analyze the two most extreme case unitary operator solutions and discuss scaling these solutions to binary inputs of arbitrary length $n$. Finally, Section concludes this paper and outlines our future works. \section{Quantum Cellular Automata} In classical CA, there is a finite set of states $\Sigma$ and an infinite or finite lattice of $L$ cells, each of which is in one of the states in $\Sigma$. At each discrete time step $t$, the state of the lattice evolves according to some local rule, which transforms the state of each cell based on the state of some neighborhood cells at time step $t-1$. For example, and employed a classical CA with two possible states ( $\Sigma=\{0,1\}$) and $L=149$ to solve the density classification problem. We will discuss the proposed classical local rules in Section . CA updating is discrete in time and space. It is also space and time homogeneous, which means that at each time step the \emph{same local rule} is applied to update \emph{all cells} synchronously. When Gr$\ddot{o}$ssing and Zeilinger formulated the first QCA, they found that except for the trivial case, strictly local, unitary evolution of the whole QCA is impossible. In fact, Meyer later proved that ``there is no homogeneous one-dimensional QCA that is nontrivial, local and scalar." Gr$\ddot{o}$ssing and Zellinger therefore relaxed the unitary constraint in their QCA by allowing approximate unitarity during states updating. After the updating at each time step, the states of the cells are normalized to make their QCA evolution unitary. This extra step also caused non-local interaction, which makes their QCA evolution non-local and the quantum unitiary non-linear. An alternative QCA formulation approach is to relax the homogeneity constraint by partitioning CA . The main idea of partitioned CA is that the set of cells are partitioned in some periodic way where every cell belongs to exactly one block partition. At different time steps, the local rule acts on a different block partition of the lattice. Such a QCA is neither time homogeneous nor space homogeneous anymore, but periodic in time and space. However, the quantum unitarity property can be maintained by using a unitary operator to update each block partition. Since unitary operators preserve the norm squared of the probability amplitudes in each block, the evolution of the entire QCA is unitary. The QPA we used in this research is a partitioned QCA, where each block contains a pair of adjacent cells, with the pairing changing at alternating time step (see Figure ). A $2\times2$ unitary matrix is applied to each pair of the cells to update the states of the QPA. Hence, the QPA is 2-step translation invariant. The unitary operator used to update the QPA, $S \in U(2)$, is as follows: \begin{equation} S(\theta,\alpha,\beta,\gamma,\delta)=\begin{bmatrix} e^{i\alpha} sin \theta & e^{i\beta} cos \theta \\ e^{i\gamma} cos \theta & e^{i\delta} sin \theta \end{bmatrix} \end{equation} where $(\alpha-\beta-\gamma+\delta)$ \ Updating the QPA for a single time step is achieved by a series of matrix multiplications on each pair of the cells: \begin{equation} \begin{bmatrix} \phi_{t+1}(x-1) \\ \phi_{t+1}(x) \end{bmatrix}=\begin{bmatrix} e^{i\alpha} sin \theta & e^{i\beta} cos \theta \\ e^{i\gamma} cos \theta & e^{i\delta} sin \theta \end{bmatrix} \begin{bmatrix} \phi_{t}(x-1) \\ \phi_{t}(x) \end{bmatrix} \end{equation} where $\phi_{t}(x)$ is the state of cell $x$ at time step $t$. The cell-pairing $(x-1,x)$ in Equation changes at alternating time step with the initial $x=t$\ At each time step $t$, each cell $x$ of the lattice contains a complex number, $\phi_t(x)$, which is the probability amplitude for the particle being in position $x$ when it is measured. Since $S$ is unitary, the total probability i.e. the sum of the norm squared of the amplitude of each cell, is always 1: \begin{align*} \sum_{x=0}^{L-1}|\phi_t(x)|^2=1 \end{align*} where $L$ is the length of the lattice. \section{The Density Classification Problem} Since the density classification problem was first formulated, various classical CA local rules have been proposed to solve the problem. The first one was by Gace, Kurdymov and Levin (GKL) , which consists of 2 rules: \begin{itemize} \item if $\phi_{t}(x)\equiv0$, $\phi_{t+1}(x)=$ majority of $\phi_{t}(x),\phi_{t}(x+1),\phi_{t}(x+3)$; \item if $\phi_{t}(x)\equiv1$, $\phi_{t+1}(x)=$ majority of $\phi_{t}(x),\phi_{t}(x-1),\phi_{t}(x-3)$; \end{itemize} The GKL rule gives classification accuracy of approximately 80\ Later, Mitchell, Crutchfield and Hraber designed a GA while Andre, Bennett and Koza applied a Genetic Programming system to discover CA local rules that gave improved performance. However, under the current problem formulation, Land and Belew proved that for a one-dimensional grid of fixed size $L$ and fixed neighborhood size $r\ge1$, there is no perfect solution to correctly classify all possible inputs. Capcarrere, Sipper and Tomassini therefore modified the problem slightly with a different final output specification: instead of a fixed-point configuration of either all 0s (when 0 is the majority in the initial configuration) or 1s (when 1 is the majority in the initial configuration), the final CA configuration can be either a background of alternative 0 and 1, with one or more blocks of at least two consecutive 0s (when 0 is the majority) or 1s (when 1 is the majority). With such modification, this problem can be solved perfectly using either rule 184 or rule 226 defined under Wolfram's classification . To investigate the QPA's ability in solving this problem, we quantized the density classification problem in the following ways: \begin{itemize} \item The QPA lattice consists of $L$ cells, where each cell $x$ represents its equivalent binary input. In other words, it is the \emph{index}, not the \emph{content}, of the cell that corresponds to the binary input whose density the QPA is classifying. The binary inputs can be in binary code, for example, binary input 00011 is represented by cell 3 while binary input 01000 is represented by cell 8. Alternatively, the binary inputs can be in Gray code , for example, binary input 00011 is represented by cell 2 while binary input 01000 is represented by cell 7. To classify the density of binary inputs of length $n$, the QPA has $L=2^n$. \item The state of cell $x$ at time step $t$ ($\phi_t(x)$) is a complex number, which is the probability amplitude of the particle being in position $x$ when it is measured. For example, if $\phi_t(3) = 0.2+0.3i$, the probability of the particle being in position 3 is $|0.2+0.3i|^2=0.13$. Of course, $\sum_{x=0}^{L-1}|\phi_t(x)|^2=1$. \item Initially, only the cell which represents the binary input to be classified has probability amplitude 1, while the rest $L-1$ cells have probability amplitude 0. For example, to classify the binary input 1000 in binary code, the initial QPA has cell 8 with probability amplitude 1, while the rest of the cells have probability amplitude 0. \item After the QPA is evolved for $M$ time steps using a unitary operator $S$, the particle is measured. We assume that within the $L$ bases Hilbert space, there is a partition between the basis states that classify the density of the binary inputs as 0 ($Z_{cells}$) and those that classify the density of the binary inputs as 1 ($O_{cells}$). Hence, if the particle collapses to one of the zero-state cells ($Z_{cells}$) more often ($>$ 50 \ \end{itemize} This is very different from the classical version of the density classification problem, where a CA is expected to converge to all 0s or all 1s. Since the QPA is a reversible CA, different initial configurations can never converge to identical states. Moreover, the QPA evolution is unitary, hence preserves the sum of the lattice states at each time step. This is not the case in classical CA, where the sum of the lattice states may vary from one time step to another. Under this quantization, the density classification problem has a general classical solution for binary inputs of any length $n$: \begin{align*} S=I=\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \end{align*} \begin{align*} Z_{cells}=\{x: 0 \le x < n; majority(x_{binary})\equiv 0\} \\ O_{cells}=\{x: 0 \le x < n; majority(x_{binary})\equiv 1\} \end{align*} Since the identity operator (I) does not change the QPA initial configuration, by assigning all cells $x$ whose binary representation has majority 0 to $Z_{cells}$ and the others to $O_{cells}$, the particle always collapse to members of $Z_{cells}$ with certainty when the binary input has majority 0 and to members of $O_{cells}$ with certainty when the binary input has majority 1. But classical solutions are not interesting. We would like to know if there is a quantum solution that can manipulate the complex probability amplitudes of the QPA to distinguish binary inputs with majority 0 from those with majority 1. The next section presents a GA that we designed to search for quantum solutions. \section{Genetic Algorithms System Design} The GA we implemented searches for quantum solutions $S$ that solve the quantized density classification problem for binary inputs of length 5. As shown in Equation , $S$ has 5 parameters: $\theta, \alpha,\beta, \gamma, \delta$. Since $\alpha, \beta, \gamma, \delta$ are constrained, $(\alpha-\beta-\gamma+\delta)$ \ Initially, we specified the $Z_{cells}$ containing all odd cells of the lattice while the $O_{cells}$ containing all even cells of the lattice. Under this setup, the GA was not able to find a $S$ that can evolve the QPA to classify the density of all $2^5=32$ binary inputs correctly. We then made another attempt by specifying the left-half cells of the lattice as $Z_{cells}$ while the right-half cells of the lattice as $O_{cells}$ but the GA was still unable to find a correct $S$ solution. We therefore decided to let the GA to find the partition $Z$. Consequently, the total number of parameters that the GA optimized became 5: $\theta, \alpha,\beta, \gamma, Z$. The partition $Z$ is represented as an integer value between 0 and $2^{32}-1$. To divide the 32 basis states into $Z_{cells}$ and $O_{cells}$, the 10-based integer value $Z$ is first converted into a 2-based binary string $Z_{binary}$. For example, $Z=4279011723_{10}$, $Z_{binary}=11111111000011001000100110001011_{2}$. Next, the indices to all 0 bits in the $Z_{binary}$ become the members of the $Z_{cells}$ while the indices to all 1 bits in the $Z_{binary}$ become the members of the $O_{cells}$. In the above example, $Z_{cells}$=$\{2,4,5,6,9,10,12,13,14,16,17,20,21,22,23\}$. When the particle is measured at time step $t$, if $\sum |\phi_t(x_i)|^2 > 0.5, \forall i \in Z_{cells}$, the particle is more likely to collapse to one of the zero-state cells than to one of the one-state cells, hence the binary input is classified with majority 0. Otherwise, the binary input is classified with majority 1. In GA, the parameters undergoing optimization are called genes and they are packed into a linear chromosome, called individual. As the first attempt using GA to search for quantum solutions $S$, we only investigated binary inputs of length 5 in this study. Hence, the fitness ($f$) of an individual ($\theta, \alpha,\beta, \gamma, Z$) is its ability to correctly classify the density of $2^5=32$ binary inputs ($BI$). The higher the $f$ is, the better the $S$ and $Z$ are in solving the problem, i.e. this is a maximization problem. \begin{equation} \max f(\theta,\alpha,\beta,\gamma, Z) =\sum_{x=0}^{31} F(S(\theta,\alpha,\beta,\gamma), Z, BI_x) \end{equation} Here, $F(S(\theta,\alpha,\beta,\gamma), Z, BI_x$) is a function that returns 1 if $S$ classifies binary input $BI_x$ correctly based on the partition $Z$. Otherwise, F returns 0. An individual that correctly classifies the density of all 32 binary inputs would have $f=32$ and is a solution to the quantized density classification problem. Function $F$ operates in the following ways. First, partition $Z$ is used to generate $Z_{cells}$ and $O_{cells}$. Next, for each binary input $BI_x$, a QPA of lattice size $L=32$ is initialized with probability amplitude of 1 at cell $x$ and 0 at all other cells, i.e. $\phi_0(x)=1$ and $\forall_{y=0}^{31}, y \ne x, \phi_0(y)=0 $. At each time step $t$, $S(\theta,\alpha,\beta,\gamma)$ is applied to each cell block from time $t-1$ to produce new probability amplitudes for time step $t$. Once completed, the particle is measured. If $\sum |\phi_t(x_i)|^2 > 0.5, \forall i \in Z_{cells}$, the QPA classifies $BI_x$ with majority 0. Otherwise, it classifies $BI_x$ with majority 1. After that, the QPA classification is compared to the correct classification. If the two match, $F$ returns 1. Otherwise $F$ returns 0. The same process is repeated for all 32 binary inputs to obtain the fitness $f$. If $f$ is 32, indicating $S$ classifies all 32 binary inputs correctly, the QPA evolution stops. Otherwise, the QPA continues to time step t+1. In other words, the number of time step $M$ of the QPA is not fixed but varies depending on $S$ and $Z$. The maximum allowed time step $M$ is 2,048 in this study. In our implementation (see Appendix A), function F only evolved QPA once and the probability amplitudes in the 32 cells can be used to classify all 32 binary inputs. This is possible because the lattice is a grid with periodic boundary conditions. As a result, to classify each binary input, we only need to shift the grid one cell left and perform the same evaluation procedure to obtain the QPA classification. In other words, to classify binary input $BI_x$, cell number $(32-x)\ The GA uses the following operators to optimize $S$ and $Z$: \begin{itemize} \item Uniform crossover: each individual is paired with a randomly selected different individual (partner) in the population to produce one offspring. The gene values of the offspring are decided in the following ways. A random number is generated for each of the 5 genes; if the random number is less than the crossover rate ($cR$), the gene is selected from the original individual. Otherwise, it is selected from the partner individual. The smaller the $cR$ is, the higher portion of the original individual is disrupted to produce its offspring. \item Gaussian mutation: the gene values of an individual is mutated in the following ways. A random number is generated for each of the 5 genes; if the random number is less than the mutation rate ($mR$), the gene is mutated by adding a random value from the Gaussian distribution under the standard deviation ($std$) specified for that gene. Otherwise, the gene value remains the same. The higher the $mR$ is, the higher portion of the individual is disrupted with new gene values. \item Survival selection: if an individual is better than its offspring, i.e. with higher fitness $f$ according to Equation , the individual survives to the next generation. Otherwise, it is replaced by its offspring. \end{itemize} Figure gives the work flow of the GA optimization process. Initially, a population of individuals, each contains ($\theta, \alpha, \beta, \gamma, Z$), is randomly generated. After applying function F to evaluate its fitness, each individual in the population undergoes either uniform crossover or Gaussian mutation alone or both to produce one offspring. Under the current setup, $\sim80\ We experimented with different $cR$ and $mR$ values and found that imposing higher disruption on individuals to produce offspring worked better in optimizing $S$ and $Z$. This might be due to the fact that our GA uses a survival selection scheme which technically is performing hill-climbing in the parameters space. Hilling-climbing is a greedy search algorithm. When all individuals perform this greedy search, the population diversity is reduced quickly and the search suffers from premature convergence. By using high $mR$ and low $cR$, a large amount of diversity is introduced to the population to balance the ``greediness" of the survival selection, hence leads to more effective search. \begin{table}[!htp] \centering \caption{GA system parameter values used to run the optimization.} \begin{tabular}{|c|c|c|c|} \hline {\bf parameter}&{\bf value} & {\bf parameter}&{\bf value}\\ \hline\hline pop\_size&200& max\_gen& 1,000 \\\hline xover\_rate (cR)& 25\ $\theta$ range & $0\sim\pi/2$ & $\theta$ std & 0.05 \\\hline $\alpha,\beta,\gamma$ range & $-\pi\sim\pi$& $\alpha,\beta,\gamma$ std & 0.1\\\hline $Z$ range & $0\sim2^{32}-1$& $Z$ std & 20 \\\hline \end{tabular} \end{table} Table lists the GA parameter values used to run the optimization. Additionally, we list the value ranges of $\theta, \alpha, \beta, \gamma$ and $Z$. In function $F$, these values are validated prior to be used to simulate the QPA for density classification. If any of the values is out-of-range, a random value within the legal range is generated to replace the out-of-range value. \section{Results} We made multiple GA runs using both binary and Gray code representations of the binary inputs. It appeared that Gray code was more suitable for this problem and was easier for the GA to discover unitary operator solutions. Among the 50 Gray code runs, 16 found a solution while only 9 of the 50 binary code runs found a solution. This might be because under Gray code representation, neighboring cells always represent two binary inputs that are different in only 1 digit. As a result, the unitary operator $S$ always updates the complex probability amplitudes of 2 binary inputs that are different in 1 digit. Such consistency made it easier for the unitary operator to classify the density of the given binary input correctly. \begin{table}[!htp] \centering \begin{tabular}{c|cccc} \hline run&$\theta(\degree)$&$\alpha(\degree)$&$\beta(\degree)$&$\gamma(\degree)$\\ \hline\hline 1&11.050685& -55.57717& 113.034549& -140.809934\\ 2&8.0667& -24.173737& 44.174208& -129.677467\\ 3&10.880958& -19.904022& 104.97193& -176.437282\\ 4&75.337518& 5.369795& 43.854368 &-49.357459\\ 5&7.259019& -27.299086& -62.015774& 111.243832\\ 6&13.700568& 76.91783& 111.454715& 134.389406\\ 7&23.718054& -86.090944& 127.961266& 106.217157\\ 8&25.152128& 170.66293 &-49.58397& -5.4686\\ 9&57.076225& -111.908294& 88.120218& -106.898772\\ \hline \end{tabular} \caption{The 9 unitary operator solutions found from the binary code runs. } \end{table} \begin{table}[!htp] \centering \begin{tabular}{c|cccc} \hline run&$\theta(\degree)$&$\alpha(\degree)$&$\beta(\degree)$&$\gamma(\degree)$\\ \hline\hline 1& 4.407264& 119.105462& -56.65684& 155.808914\\ 2& 22.357464 &117.90735& -9.633698& 151.621098\\ 3& 10.431628& -129.653366& 35.876261& -37.38258\\ 4& 7.303212& -36.827341& 42.439786& 13.288965 \\ 5& 12.565333& -133.764371& -116.992579& 27.211073 \\ 6& 6.568701& 167.726702& -105.341636& 125.939087 \\ 7& 27.507306 & -72.340869& -127.871776& -23.952314 \\ 8& 20.083639 &93.743705& -44.820769& -109.822316 \\ 9& 27.284116& -29.442762& -49.124343& 32.082641 \\ 10& 10.472256& -63.163587& 11.258074& -110.796705 \\ 11& 28.257343& 41.031094 &-28.261072& -83.989587 \\ 12& 36.973794& 105.346366& 168.109796& 110.596441 \\ 13& 12.472202& -24.919773& 1.682202& -178.873208 \\ 14& 29.280147& -124.541844& 120.74784& 62.525769 \\ 15& 24.339945& 148.928022& 87.470762 &-106.736046 \\ 16& 37.054707& -22.12085& -48.986319& 14.976938 \\ \hline \end{tabular} \caption{The 16 unitary operator solutions found from the Gray code runs. } \end{table} Tables and give these unitary operator solutions ($\theta, \alpha, \beta, \gamma$) found in the binary and Gray code runs while Tables and present the partition $Z$ solutions and the number of time steps $M$. As shown, the unitary operators have a wide range of values and they evolved the QPA for a different number of time step $M$. They also used a different $Z$ to classify the density of the binary inputs. In Tables and , the row ``majority" gives the correct density classification for binary input $x$; they are listed from left to right based on cell order ($0\leq x\le 31$). Similarly, $Z_{binary}$ gives the binary string converted from the partition $Z$, which is also listed in cell order from 0 to 31. One similarity among the $Z_{binary}$ solutions is that the number of 0-bit is close to half of the total number of binary inputs. This makes sense as half of the 32 binary inputs has majority 0 and the other half has majority 1. By allocating half of the cells to members of $Z_{cells}$ and the other half to members of $O_{cells}$, it is easier for the unitary operator to balance probability amplitudes among members of $Z_{cells}$ and $O_{cells}$ to classify both majority-0 binary inputs and majority-1 binary inputs correctly. \begin{table}[!htp] \centering \begin{tabular}{|c||c||c||c|} \hline x&\scalebox{0.71}{0-1-2-3-4-5-6-7-8-9-10-11-12-13-14-15-16-17-18-19-20-21-22-23-24-25-26-27-28-29-30-31}&0-bits&\\ \hline\hline majority&0-0-0-0-0-0-0-1-0-0-0-1-0-1-1-1-0-0-0-1-0-1-1-1-0-1-1-1-1-1-1-1&16&\\ \hline run& $Z_{binary}$&&$M$\\ \hline \hline 1&1-1-0-1-0-0-0-1-1-0-0-1-0-0-0-1-0-0-1-1-0-0-0-0-1-1-1-1-1-1-1-1&15&2048\\ 2&0-0-0-1-0-0-0-0-1-0-0-1-1-1-1-1-0-1-1-1-1-1-0-0-0-0-0-1-1-1-0-1&16&1368\\ 3&0-1-1-1-0-1-0-0-0-1-1-0-0-0-0-0-0-1-1-1-1-0-1-1-0-1-1-0-0-1-0-0&17&1880\\ 4&1-0-1-0-1-1-1-0-0-1-1-1-1-1-0-0-0-1-1-0-0-0-0-0-0-1-1-0-1-1-0-0&16&2034\\ 5&1-1-1-1-0-0-0-1-0-0-0-1-1-0-0-1-1-1-1-1-1-1-0-0-1-1-0-0-0-0-0-0&16&2048\\ 6&1-1-0-1-1-0-0-1-0-0-0-0-1-1-0-1-0-0-1-1-0-0-0-1-1-0-0-0-1-1-1-1&16&1184\\ 7&0-1-0-0-0-1-1-0-0-0-0-0-0-1-1-0-1-0-1-1-0-1-1-1-1-1-0-0-0-1-1-1&16&1594\\ 8&1-1-0-1-0-0-0-1-0-0-0-0-0-0-1-1-1-1-0-0-0-1-1-0-1-0-0-1-1-1-1-1&16&1182\\ 9&0-1-0-1-1-0-0-0-0-0-0-1-0-0-0-0-1-1-1-1-1-1-0-1-1-1-0-1-1-0-0-1&16&768\\ \hline \end{tabular} \caption{The $Z_{binary}$ and the time step $M$ solutions from the binary code runs.} \end{table} \begin{table}[!htp] \centering \begin{tabular}{|c||c||c||c|} \hline x&\scalebox{0.71}{0-1-2-3-4-5-6-7-8-9-10-11-12-13-14-15-16-17-18-19-20-21-22-23-24-25-26-27-28-29-30-31}&0-bits&\\ \hline\hline majority&0-0-0-0-0-1-0-0-0-1-1-1-0-1-0-0-0-1-1-1-1-1-1-1-0-1-1-1-0-1-0-0&16&\\ \hline run& $Z_{binary}$&&$M$\\ \hline \hline 1&1-0-0-1-1-0-1-1-0-0-0-1-1-1-0-1-1-0-0-0-0-1-0-0-1-0-0-1-0-1-1-1&16&2048\\ 2&0-1-0-1-1-1-0-1-1-1-1-1-1-1-0-0-0-1-1-0-1-0-0-0-1-1-0-0-0-0-0-0&16&2012\\ 3&1-0-1-1-1-0-1-1-1-1-1-1-0-0-0-0-1-1-0-1-1-0-0-0-1-0-0-0-0-0-0-1&16&811\\ 4&0-1-0-0-1-1-0-0-0-0-1-0-0-0-0-0-0-1-0-1-1-1-1-1-1-1-0-1-1-1-0-0&17&2048\\ 5&0-1-0-1-1-1-0-0-1-0-1-1-1-1-0-0-0-1-0-1-1-1-1-0-1-1-0-0-0-0-0-0&16&624\\ 6&0-0-0-1-0-1-1-1-0-0-0-1-0-0-0-0-0-0-0-1-1-1-1-1-0-1-1-1-1-0-1-0&17&2046\\ 7&1-1-0-0-1-0-0-0-0-1-1-1-1-0-0-0-0-0-0-0-1-1-0-1-1-1-0-1-1-1-0-1&16&1904\\ 8&0-1-0-0-1-0-1-1-1-1-1-0-1-1-0-1-0-1-0-1-1-1-1-1-0-0-0-0-0-0-0-0&16&1274\\ 9&0-0-0-0-0-0-0-1-1-1-1-1-0-1-1-1-1-1-1-1-0-0-0-1-0-1-1-1-0-0-0-0&16&620\\ 10&0-1-1-1-1-0-1-1-0-1-0-1-0-0-0-1-1-0-1-0-0-0-0-1-0-0-1-1-0-1-1-1&15&1978\\ 11&0-0-1-1-0-0-0-1-0-0-0-1-1-1-1-1-0-0-1-1-0-1-1-1-0-0-0-1-0-1-1-0&16&1572\\ 12&1-1-1-0-0-1-0-1-1-1-0-0-0-1-0-0-1-0-0-0-0-1-0-1-0-1-0-1-1-0-1-1&16&632\\ 13&1-1-1-1-0-0-0-0-1-0-0-0-0-1-0-1-0-1-1-0-0-1-1-1-1-1-0-0-0-0-1-1&16&1750\\ 14&0-1-1-1-0-0-0-1-0-1-1-1-1-0-1-0-0-0-0-0-0-0-0-1-0-1-1-1-0-1-1-1&16&800\\ 15&1-1-1-0-0-0-0-0-0-0-0-0-0-1-1-1-0-0-1-1-0-0-1-1-1-1-1-1-0-1-0-1&16&2048\\ 16&0-1-1-0-0-1-0-1-1-1-1-1-1-0-0-1-1-1-0-0-1-1-0-0-1-1-0-0-0-0-0-0&16&684\\ \hline \end{tabular} \caption{The $Z_{binary}$ and the time step $M$ solutions from the Gray code runs.} \end{table} Since all these unitary operators classify the density of the 32 binary inputs of length 5 correctly, it is natural to ask ``do they produce the same QPA evolution dynamics, in terms of the probability amplitudes propagation?" We will analyze these unitary operator solutions through QPA simulations and answer this question in the following section. \section{Analysis and Discussion} We simulated QPA evolutions and found that their dynamics were different under different unitary operator solutions. In particular, some QPA propagated the probability amplitudes very fast while others transformed the states in a much slower pace. Further analysis shows that the major driving force of the propagation is the value of $\theta$: a small $\theta$ gives the particle small mass hence fast propagation while large $\theta$ has the opposite effect. When $\theta=0\degree$, $S$ simply interchanges the states of adjacent cells so the probability propagation speed is 1 in lattice units. By contrast, when $\theta=\pi/2$, $S$ is the identity with phase in diagonal, where the phase is unobservable, hence there is no flow. Meyer has made similar observations in his QPA simulations under one-parameter ($\theta$) unitary operators. Since we used a more general unitary operator with 5 parameters, the QPA evolution dynamics were more complicated, given that $\alpha, \beta, \gamma, \delta$ also influenced the probability amplitudes propagation. We examined the QPA dynamics that were evolved under unitary operators that handled binary code and that handled Gray code representations and found that they had similar patterns. Thereafter, we will use the QPA evolved by unitary operators that handle binary code to perform our analysis. Figure gives the QPA simulation that has the fastest propagation flow among all solutions. The unitary operator is $S=\{\theta=7.26\degree,\alpha=-27.3\degree,\beta=-62\degree,\gamma=111.24\degree\}$. The value shown at each cell $x$ at time step $t$ is the norm squared of the probability amplitudes, i.e. $|\phi_{t}(x)|^2$. Figure shows the evolution during the first 64 time steps. Measuring the locations of the peaks of the probability distribution at successive time steps indicates that the propagation speed is a little bit less than 1 in lattice units. Similarly, Figure gives the QPA simulation that has the slowest propagation flow among all solutions. The unitary operator is $S=\{\theta=75.3\degree,\alpha=-5.37\degree,\beta=43.84\degree,\gamma=-49.36\degree\}.$ Figure shows the probability propagation during the first 64 time steps. Measuring the locations of the peaks of the probability distribution at successive time steps indicates that the propagation speed is approximately 1/4 in lattice units. When $\theta=\pi/2$ or $\theta=0\degree$, the QPA evolution dynamics are not interesting; the probability ($|\phi_{t}(x)|^2$) in each cell at the current time step is either an identical copy or a swap of its neighboring cell from the previous time step. However, when $0<\theta<\pi/2$, the probabilities in the QPA are propagated in a wide range of speeds, depending mostly on the $\theta$ of the unitary operator. Moreover, there are many unitary operators that can propagate the probabilities in ways such that when the particle is measured, it is more likely to collapse to the basis states that are on the correct side of the boundary partition $Z$ for the QPA to decide if the binary input has majority 0 or majority 1. As mentioned in Section , the classical solution $S=I$ uses the $Z_{binary}$ that is the same as the \emph{majority binary string} to classify the density of the binary inputs. By contrast, quantum solutions use $Z_{binary}$ that are very different from the \emph{majority binary string} to classify the density of the binary inputs (see Table and ). In other words, within the 32 bases Hilbert space, there exists many partitions $Z$ between the basis states that classify the density of the binary inputs as 0 and those that classify the density of the binary inputs as 1. Many unitary operators can pair with these partitions to solve the density classification problem for binary inputs of length 5. However, in order for the unitary operator solutions to work for binary inputs of any length $n$, the paired partitions also have to be general solutions. In the future work, we plan to investigate the patterns in these partition solutions and abstract these patterns to be used to classify binary inputs of any length $n$. An alternative approach to investigate general solutions for the quantized density classification problem is by applying the GA to search $S$ and $Z$ for different binary input sizes $n$. It is possible that such solutions exist, but the GA may or may not be able to find them. This is because with the increase of the binary input length, the search space of $Z$ increases exponentially. However, it is also possible that the number of partition $Z$ solutions increases as the length of the binary inputs increases, hence the problem difficulty remains constant regardless of the binary inputs length. In either case, we can adjust GA parameters, such as $mR$ and $cR$, to improve the GA search efficiency. Another issue that we need to address using this approach is the hardware resources required to simulate the QPA evolution under large lattice sizes. These issues will be investigated in our future work. Compared to the classical CA, which requires a lattice of length $L=n$ to solve the density classification problem for binary inputs of length $n$, the QPA requires lattice length $L=2^n$ under this quantized version of the density classification problem. However, the trade-off we have gained is a larger CA evolution space (complex Hilbert space) which allows many unitary operators to manipulate and solve the problem. In this research, our GA has found some of these solutions. Moreover, it is possible that the solutions are within some areas of the parameters space and can be simplified as 2-parameter unitary operators. We will test this hypothesis in our future work. \section{Concluding Remarks} Our investigation of using a quantum CA to solve the density classification problem not only produced some interesting results but also posed some challenging questions. First of all, we found that the QPA devised by Meyer can be used to solve a quantized version of the density classification problem for binary inputs of length 5 successfully. Secondly, we found that there are more than one quantum solution to this quantized problem and the GA system we designed can find many of them. Thirdly, we found that these quantum solutions propagate probability amplitudes in different speeds, depending mostly on a particular parameter $\theta$ of the unitary operator. Lastly, we found that there exists many boundary partitions in the 32 bases Hilbert space that separate the basis states classifying the density of the binary inputs as 0 from those classifying the density of the binary inputs as 1. When the partitions are paired with suitable unitary operators, they are able to solve the quantized density classification problem for binary inputs of length 5. However, scaling the approach to find general solutions to the quantized density classification problem requires more work. We have identified a couple of avenues: one studies the partition $Z$ patterns and the other improves GA scaling performance. We plan to investigate both approaches in our future work. Moreover, given the number of variables in the solutions ($\theta,\alpha,\beta,\gamma,Z,M$), it is not clear to us if we can reduce the number of parameters by increasing/decreasing other parameter values. For example, can we remove $Z$ by increasing $M$? Or can we remove $\alpha$ by increasing $M$? These are open questions that we plan to answer in our futuer work. \begin{thebibliography}{} \bibitem {abk} D. Andre and F. H. Bennett III and J. R. Koza, \newblock Discovery by genetic programming of a cellular automata rule that is better than any known rule for the majority classification problem. \newblock In \textit{Genetic Programming 1996: Proceedings of the First Annual Conference}, pages 3-11, 1996. \bibitem {cst} M. S. Capcarrere and M. Sipper and M. Tomassini, \newblock Two-state, r = 1 cellular automaton that classifies density. \newblock \textit{Physical Review Letters}, 77(24):4969--4971, 1996. \bibitem{deutsch} D. Deutsch, \newblock Quantum theory, the Church-Turing principle and the universal quantum computer. \newblock \textit{Proceedings of the Royal Society of London A 400}, pages 97--117, 1985. \bibitem {feynman} R. P. Feynman, \newblock Quantum mechanical computers. \newblock \textit{Optics News}, pages 11-- 20, February 1985. \bibitem {gkl} P. Gach and G. L. Kurdyumov and L. A. Levin, \newblock One-dimensional uniform arrays that wash out finite islands. \newblock \textit{Problemy Peredachi Informatsii}, Vol. 14, No. 3, pages 92--98, 1978. \bibitem {goldberg} D. E. Goldberg, \newblock \textit{Genetic Algorithms in Search, Optimization and Machine Learning}, ISBN: 0-201-15767-5, 1989. \bibitem {gray} F. Gray, \newblock \textit{Pulse code communication}, U.S. Patent 2,632,058, March 17, 1953. \bibitem{grossing_zeilinger} G. Gr$\ddot{o}$ssing and A. Zeilinger, \newblock Quantum cellular automata, \newblock \textit{Complex Systems,} 2, pages 197--208, 1988. \bibitem {holland} J. H. Holland, \newblock \textit{Adaptation in Natural and Artificial Systems}, MIT Press, 1975. \bibitem {koza} J. Koza, \newblock \textit{Genetic Programming: On the Programming of Computers by Means of Natural Selection}, MIT Press, 1992. \bibitem {land_belew} M. Land and R. K. Belew, \newblock No perfect two-state cellular automate for density classification exists, \newblock \textit{Physical Review Letters}, 74(25), pages 5148-5150, 1995. \bibitem {meyer} D. Meyer, \newblock From quantum cellular automata to quantum lattice gases. \newblock \textit{Journal of Statistical Physics}, Vol. 85, Issues 5-6, pages 554--574, 1996. \bibitem{mch} M. Mitchell and J. P. Crutchfield and P. T. Hraber, \newblock Evolving cellular automata to perform computations: mechanisms and impediments. \newblock \textit{Physica D}, 75, pages 361--391, 1994. \bibitem{packard} N. H. Packard, \newblock Adaptation toward the edge of chaos, \newblock In \textit{Dynamic Patterns in Complex Systems}, \newblock eds. J. A. S. Kelso, A. J. Mandell, and M. F. Shlesinger, pages 293--301, World Scientific, Singapore, 1988. \bibitem {cheung} C. A. P$\acute{e}$rez-Delgado and D. Cheung, \newblock Local unitary quantum cellular automata. \newblock \textit{Physical Review A (Atomic, Molecular, and Optical Physics)}, 76:032320, 2007. \bibitem {miszczak_etal} B. Schumacher and R. F. Werner, \newblock Reversible Quantum Cellular Automata. \newblock \textit{Quantum Information Processing}, 2011. \bibitem {toffoli_margolus} T. Toffoli and N. H. Margolus \newblock Invertible cellular automata: A review. \newblock \textit{Physica D: Nonlinear Phenomena,}, 45:1-2, pages 229--253, September 1990. \bibitem {watrous} J. Watrous, \newblock On one-dimensional quantum cellular automata, \newblock In \textit{Proceedings of the 36th Annual Symposium on Foundations of Computer Science}, pages 528-537, October 1995. \bibitem {wolfram} S. Wolfram, \newblock Statistical mechanics of cellular automata \newblock \textit{Reviews of Modern Physics}, 55:3, pages 601-644, 1983. \end{thebibliography} \section*{Appendix A} \small \texttt{\#include <stdio.h>\\ \#include <stdlib.h>\\ \#include <math.h>\\ \#include <complex.h>\\ double radian(double degree)\{return degree*M\_PI/180.0;\}\\ complex double e(double phi)\{return (cos(radian(phi))+ I * sin(radian(phi)));\}\\ double norm(complex double x)\{return creal(x)*creal(x)+cimag(x)*cimag(x);\}\\ /* Dimension of problem and boundaries for variables */\\ int D = 5;\\ /* Xl defines lower limit; Xu defines upper limit */\\ double Xl[5] = \{0.0, -180.0, -180.0, -180.0, 0.0\}; \\ double Xu[5] = \{90.0, 180.0, 180.0, 180.0, 4294967296.0\};\\ double SD[5] = \{0.05, 0.1, 0.1, 0.1, 20.0\};\\ /*majority binary code*/\\ const int majority[32]=\{0,0,0,0,0,0,0,1,0,0,0,1,0,1,1,1,0,0,0,1,0,1,1,1,0,1,1,1,1,1,1,1\};\\ /*majority gray code*/\\ //const int majority[32]=\{0,0,0,0,0,1,0,0,0,1,1,1,0,1,0,0,0,1,1,1,1,1,1,1,0,1,1,1,0,1,0,0\};\\ int func(double *rule) // rule=($\theta,\alpha,\beta,\gamma$,Z)\\ \{\\ \indent // Correction of boundary constraint violations \\ \indent for (i=0; i<D; i++)\\ \indent \{\\ \indent\indent if (rule[i] < Xl[i] || rule[i] > Xu[i]) \\ \indent\indent\indent rule[i] = Xl[i] + (Xu[i] - Xl[i])*((double)rand()/((double)RAND\_MAX + 1.0)); \indent \}\\ \\ \indent // Converting Z to Zbinary\\ \indent int z\_binary[32]=\{0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0\};\\ \indent int index;\\ \indent int L=32;\\ \indent unsigned long Z = (unsigned long)rule[D-1];\\ \indent for(index=0; index < L; index++)\\ \indent \indent if(Z < 2)\\ \indent\indent \{\\ \indent\indent\indent z\_binary[index]=(int) Z;\\ \indent\indent\indent index =L;\\ \indent\indent \}\\ \indent \indent else\\ \indent \indent \{\\ \indent \indent \indent z\_binary[index]=(int)(Z\ \indent \indent \indent Z=Z/2;\\ \indent \indent \} \\ \indent // QPA simulation\\ \indent int i, j, step, start, k, n, o; \\ \indent int fitness=0;\\ \indent int M=2048;\\ \indent int SAMPLES = 32;\\ \indent double phiSum = 180.0;\\ \indent complex double U2[2][2];\\ \indent complex double ca[2049][32];\\ \indent U2[0][0]=sin(radian(rule[0]))*e(rule[1]);\\ \indent U2[0][1]=cos(radian(rule[0]))*e(rule[2]);\\ \indent U2[1][0]=cos(radian(rule[0]))*e(rule[3]);\\ \indent U2[1][1]=sin(radian(rule[0]))*e(phiSum-rule[1]+rule[2]+rule[3]);\\ \indent ca[0][0]=1.0;\\ \indent for (j=1; j < L; j++)\\ \indent \{\\ \indent \indent ca[0][j]=0.0;\\ \indent \}\\ \indent for (step=1; step <= M; step++)\\ \indent \{\\ \indent \indent start = step \ \indent \indent if(start == 1)\\ \indent \indent \{ //odd step \\ \indent \indent\indent for(k = 0; k < L; k=k+2)\\ \indent \indent\indent \{\\ \indent \indent\indent\indent ca[step][k] = U2[0][0]*ca[step-1][k]+U2[0][1]*ca[step-1][k+1];\\ \indent \indent\indent\indent ca[step][k+1] = U2[1][0]*ca[step-1][k]+U2[1][1]*ca[step-1][k+1];\\ \indent \indent \indent \}\\ \indent \indent \}\\ \indent \indent else /*shift right*/\\ \indent \indent \{ \\ \indent \indent\indent for(k = 1; k < L; k=k+2)\\ \indent \indent \indent \{\\ \indent \indent \indent \indent ca[step][k] = U2[0][0]*ca[step-1][k]+U2[0][1]*ca[step-1][k+1];\\ \indent \indent \indent \indent ca[step][k+1] = U2[1][0]*ca[step-1][k]+U2[1][1]*ca[step-1][k+1];\\ \indent \indent \indent \}\\ \indent \indent \indent ca[step][L-1] = U2[0][0]*ca[step-1][L-1]+U2[0][1]*ca[step-1][0];\\ \indent \indent \indent ca[step][0] = U2[1][0]*ca[step-1][L-1]+U2[1][1]*ca[step-1][0];\\ \indent \indent \}\\ \\ \indent \indent /*check each of the 32 samples*/\\ \indent \indent for(i = 0 ; i < SAMPLES; i++) \\ \indent \indent \{\\ \indent \indent \indent double zero=0.0, one=0.0;\\ \indent \indent \indent int cell=(L-i)\ \indent \indent \indent for(index=0; index < L; index++)\\ \indent \indent \indent \{\\ \indent \indent \indent \indent if(z\_binary[cell]==0)\\ \indent \indent \indent \indent zero+=norm(ca[step][index]);\\ \indent \indent \indent \indent cell=(cell+1)\ \indent \indent \indent \} \\ \indent \indent \indent one=1.0-zero;\\ \indent \indent \indent if(majority[i] == 0 \&\& zero > 0.5)\\ \indent \indent \indent \{\\ \indent \indent \indent \indent fitness++;\\ \indent \indent \indent \}\\ \indent \indent\indent else if(majority[i] == 1 \&\& one > 0.5)\\ \indent \indent \indent \{\\ \indent \indent \indent \indent fitness++;\\ \indent \indent \indent \}\\ \indent \indent \}//finish evaluating 32 samples\\ \\ \indent \indent if( fitness == 32) \\ \indent \indent \{ \\ \indent \indent\indent for(n=0; n <= step; n++)\\ \indent \indent\indent \indent for(o=0; o < L; o++) \\ \indent \indent\indent \indent \indent printf(``\ \indent \indent\indent step = M;\\ \indent \indent \}\\ \indent \} //close M steps loop\\ \indent return fitness;\\ \} //close function } |
1501.04061 | Title: Superfield Effective Potential for the Supersymmetric Topologically
Massive Gauge theory in Four Dimensions
Abstract: We explicitly calculate the one-loop effective potential for the
supersymmetric topologically massive gauge theory in four dimensions, where the
chiral scalar superfield is directly coupled to the field strength for the
gauge spinor superfield.
Body: \title{Superfield Effective Potential for the Supersymmetric Topologically Massive Gauge theory in Four Dimensions} \author{F. S. Gama} \email{fgama@fisica.ufpb.br} \affiliation{Departamento de F\'{\i}sica, Universidade Federal da Para\'{\i}ba\\ Caixa Postal 5008, 58051-970, Jo\~ao Pessoa, Para\'{\i}ba, Brazil} \author{M. Gomes} \email{mgomes@fma.if.usp.br} \affiliation{Departamento de F\'{\i}sica Matem\'atica, Universidade de S\~ao Paulo,\\ Caixa Postal 66318, 05314-970, S\~ao Paulo, SP, Brazil} \author{J. R. Nascimento} \email{jroberto@fisica.ufpb.br} \affiliation{Departamento de F\'{\i}sica, Universidade Federal da Para\'{\i}ba\\ Caixa Postal 5008, 58051-970, Jo\~ao Pessoa, Para\'{\i}ba, Brazil} \author{A. Yu. Petrov} \email{petrov@fisica.ufpb.br} \affiliation{Departamento de F\'{\i}sica, Universidade Federal da Para\'{\i}ba\\ Caixa Postal 5008, 58051-970, Jo\~ao Pessoa, Para\'{\i}ba, Brazil} \author{A. J. da Silva} \email{ajsilva@fma.if.usp.br} \affiliation{Departamento de F\'{\i}sica Matem\'atica, Universidade de S\~ao Paulo,\\ Caixa Postal 66318, 05314-970, S\~ao Paulo, SP, Brazil} \begin{abstract} We explicitly calculate the one-loop k\"{a}hlerian effective potential for the supersymmetric topologically massive gauge theory in four dimensions which involves two gauge superfields, the usual scalar one and the spinor one originally introduced by Siegel, coupled to a chiral scalar matter. \end{abstract} \maketitle \section{Introduction} It is well known (see f.e. ) that in a four-dimensional space-time, three types of constrained superfields exist, consisting of irreducible representations of the supersymmetry algebra, that is, chiral, antichiral and linear superfields. The scalar chiral and antichiral superfields are well studied being the basic ingredients of the Wess-Zumino model and many other field theories . Another well-studied important example is a real gauge superfield which is a reducible one since it is unconstrained, represents itself as a natural superfield extension of the usual gauge field, being thus a basic ingredient for supergauge theories such as super-QED and super-Yang-Mills theory (for different aspects of supergauge theories, see and many other textbooks). However, these models do not exhaust the set of physically interesting theories. In this paper, we consider a model for a spinor chiral superfield coupled to an usual chiral matter. Originally, the spinor chiral superfield was introduced in where it was shown to correspond to the so-called tensor multiplet and to allow for introducing, first, a new supergauge model, and second, a topological mass term in the case of coupling of the spinor gauge superfield to the usual real gauge scalar superfield. One more interesting feature of this model is that the gauge invariant strength, corresponding to spinor chiral and antichiral superfields, is just a linear superfield, differently from the chiral one, occurring for the real scalar superfield . While in only the free theory has been considered, we study here its coupling to a chiral matter. Classical aspects of this model were discussed in . An alternative coupling for the linear superfield and tensor multiplet has been discussed in , where some of its string-related aspects were considered (for applications of this multiplet see also the references therein). Using the previously developed superfield effective potential methodology , we calculate the one-loop superfield effective potential for this theory. We emphasize that, up to now, there were no examples of quantum calculations involving the chiral spinor superfields. The structure of the paper looks like follows. In the section II, we discuss the classical action of the chiral spinor gauge superfield, coupled to the usual scalar gauge superfield and a chiral matter. In the section III we calculate the one-loop effective potential in this theory, and the section IV contains the summary of our results. \section{Supersymmetric topologically massive gauge theory} Let us start our study with the supersymmetric topologically massive gauge theory which will be used to find the one-loop K\"{a}hlerian effective potential (KEP). In the pure gauge sector, we have \bea S_{G}=\frac{1}{2}\int d^6zW^\alpha W_\alpha-\frac{1}{2}\int d^8zG^2-m\int d^8zVG \ , \eea where $m$ is a constant with mass dimension equal to 1, and \bea W_\alpha=i\bar D^2D_\alpha V \ , \ G=-\frac{1}{2}(D^\alpha\psi_\alpha+\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}}) \ , \eea where $\psi_{\alpha}$, $\bar{\psi}_{\dot{\alpha}}$ are chiral and antichiral spinor superfield corresponding to the tensor multiplet , and $V$ is an usual real gauge superfield. Actually, $G$ is a linear superfield satisfying the relation $D^2G=\bar{D}^2G=0$. The superfield strengths $W_\alpha$, $G$, and the action () are invariant under the Abelian gauge transformations: \bea \delta V=i(\bar\Lambda-\Lambda) \ , \ \delta\psi_\alpha=i\bar D^2D_\alpha L \ , \ \delta\bar\psi_{\dot{\alpha}}=-iD^2\bar D_{\dot{\alpha}}L \ , \eea where $\Lambda$ is a chiral superfield, and $\bar\Lambda$ is an antichiral one, and $L=\bar L$ is a real scalar one . Let us show that the theory () describes a massive gauge theory. For this, let us extract the equations of motion by varying the action () with respect to the superfields $V$ and $\psi_\alpha$. Then, we get \bea \frac{\delta S_{G}}{\delta V}&=&iD_\alpha W^\alpha-mG=0 \ ,\\ \frac{\delta S_{G}}{\delta \psi_\alpha}&=&\bar D^2D^\alpha G-imW^\alpha=0 \ . \eea On the one hand, if we multiply eq. () by $\bar D^2D^\beta$ and use $\bar D^2D^\beta D_\alpha W^\alpha=\Box W^\beta$, we get \bea (\Box-m^2)W^\alpha=0 \ . \eea On the other hand, if we multiply eq. () by $D_\alpha$ and use $D_\alpha\bar D^2 D^\alpha G=\Box G$, we get \bea (\Box-m^2)G=0 \ . \eea Therefore, from eqs. () and (), we can conclude that the superfield strengths $W_\alpha$ and $G$ satisfy massive Klein-Gordon equations. In order to perform quantum calculations, we must add to () a gauge-fixing term. In particular, we will consider the following one : \bea S_{GF}=-\frac{1}{2\alpha}\int d^8zV\{D^2,\bar D^2\}V-\frac{1}{8\beta}\int d^8z(D^\alpha\psi_\alpha-\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})^2 \ , \eea where $\alpha$ and $\beta$ are the gauge-fixing parameters. The ghosts are completely factorized since the theory is Abelian. Now, let us introduce interaction between the (anti-)chiral scalar superfield and the gauge superfields . Under the usual gauge transformation, the chiral and antichiral matter superfields transform as \bea \Phi^{\prime}=e^{2ig\Lambda}\Phi \ , \ \bar\Phi^{\prime}=\bar\Phi e^{-2ig\bar\Lambda} \ . \eea The interaction term that we will consider in this paper, which is invariant under the combined transformations () and (), is given by \bea S_M=\int d^8z\bar\Phi e^{2gV}\Phi e^{4hG} \ . \eea The coupling constants $g$ and $h$ have mass dimensions zero and $-1$, respectively. The model $S_G+S_M$ [see eqs. () and ()] was considered in in the study of the formation of cosmic strings. It follows from this expression that the tree-level KEP is \bea K^{(0)}=\Phi\bar{\Phi}. \eea Finally, the supersymmetric topologically massive gauge theory that we will study in this work follows from (), (), and (): \bea S&=&-\frac{1}{2}\int d^8z V(-D^\alpha \bar D^2D_\alpha+\frac{1}{\alpha}\{D^2,\bar D^2\})V-\frac{1}{8}\int d^8z \Big\{\big(1+\frac{1}{\beta}\big)[\psi_\alpha D^\alpha D^\beta\psi_\beta\nonumber\\ &&+\bar\psi_{\dot{\alpha}}\bar D^{\dot{\alpha}}\bar D^{\dot{\beta}}\bar\psi_{\dot{\beta}}]+2\big(1-\frac{1}{\beta}\big)\psi_\alpha D^\alpha\bar D^{\dot{\beta}}\bar\psi_{\dot{\beta}}\Big\}+\frac{m}{2}\int d^8zV(D^\alpha\psi_\alpha+\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})\nonumber\\ &&+\int d^8z\bar\Phi e^{2gV}\Phi e^{-2h(D^\alpha\psi_\alpha+\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})} \ , \eea where we explicitly wrote the gauge superfields. The standard method of calculating the effective action is based on the methodology of the loop expansion . To do this, we make a shift $\Phi\rightarrow\Phi+\phi$ in the superfield $\Phi$ (together with the analogous shift for $\bar\Phi$), where now $\Phi$ is a background (super)field and $\phi$ is a quantum one. We assume that the gauge superfields $V$, $\psi_\alpha$, and $\bar\psi_{\dot{\alpha}}$ are quantum. In order to calculate the effective action at the one-loop level, we have to keep only the quadratic terms in the quantum superfieds. By using this prescription, we get from () \bea &&S_2[\bar\Phi,\Phi;\bar\phi,\phi,\psi_\alpha,\bar\psi_{\dot{\alpha}},V]=S_q+S_{int} \ ,\\ &&S_q=\frac{1}{2}\int d^8z\big[ -V\Box(\Pi_{1/2}+\frac{1}{\alpha}\Pi_0)V-\frac{1}{4}\big[\big(1+\frac{1}{\beta}\big)(\psi_\alpha D^\alpha D^\beta\psi_\beta+\bar\psi_{\dot{\alpha}}\bar D^{\dot{\alpha}}\bar D^{\dot{\beta}}\bar\psi_{\dot{\beta}})\nonumber\\ && \ \ \ \ +2\big(1-\frac{1}{\beta}\big)\psi_\alpha D^\alpha\bar D^{\dot{\beta}}\bar\psi_{\dot{\beta}}\big]+2\bar\phi\phi\big] \ ,\\ &&S_{int}=\frac{1}{2}\int d^8z\big\{(m-8gh\bar\Phi\Phi)V(D^\alpha\psi_\alpha+\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})+2(2g)\bar\Phi V\phi+2(2g)\Phi\bar\phi V\nonumber\\ && \ \ \ \ +(2g)^2\bar\Phi\Phi V^2-4h\bar\Phi(D^\alpha\psi_\alpha+\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})\phi-4h\Phi\bar\phi(D^\alpha\psi_\alpha+\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})\nonumber\\ && \ \ \ \ +(2h)^2\bar\Phi\Phi[(D^\alpha\psi_\alpha)D^\beta\psi_\beta+(\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}})\bar D^{\dot{\beta}}\bar\psi_{\dot{\beta}}+2(D^\alpha\psi_\alpha)\bar D^{\dot{\alpha}}\bar\psi_{\dot{\alpha}}]\big\} \ , \eea where the irrelevant terms were omitted, including those involving covariant derivatives of the background (anti-)chiral superfields. Moreover, we used the projection operators $\Pi_{1/2}\equiv-\Box^{-1}D^\alpha\bar D^2D_\alpha$ and $\Pi_{0}\equiv\Box^{-1}\{D^2,\bar D^2\}$. The one-loop approximation does not depend on how we break the Lagrangian into free and interacting parts . However, by convenience, we will extract the propagators from the terms that are independent of the background superfields and the vertices from the ones in which the quantum superfields interact with the background ones. In the gauges $\alpha=0$ and $\beta=-1$, we obtain from $S_q$ the propagators \bea \langle V(1)V(2)\rangle=-\frac{1}{p^2}(\Pi_{1/2})_1\delta_{12} \ , \ \langle\psi_\alpha(1)\bar\psi_{\dot{\alpha}}(2)\rangle=\frac{2p_{\alpha\dot{\alpha}}}{p^4}\delta_{12} \ , \ \langle\phi(1)\bar\phi(2)\rangle=\frac{1}{p^2}\delta_{12} \ . \eea Before we start the calculation of the one-loop supergraphs, we first notice from () that there is a factor $D^\alpha\bar D^2$ in a vertex at one end of the propagator $\langle\psi_\alpha(1)\bar\psi_{\dot{\alpha}}(2)\rangle$, and there is a factor $\bar D^{\dot{\alpha}}D^2$ in the other vertex at the other end of the same propagator. Here the factors $\bar{D}^2$ and $D^2$ are present in the vertices due to the chirality (antichirality) of the superfield $\psi_{\alpha}$ ($\bar{\psi}_{\dot{\alpha}}$) just as in the usual Wess-Zumino model, because of the properties of the variational derivatives with respect to the chiral superfields (see ), and the $D^{\alpha}$, $\bar{D}^{\dot{\alpha}}$ arise from the explicit form of the vertices. It is convenient to go from the above used formulation of propagators where the derivatives $D^2$, $\bar{D}^2$ are associated with the vertices, to a formulation where these derivatives are incorporated into the propagators (these two manners to introduce the Feynman supergraphs exist also in the Wess-Zumino model, see f.e. ). In other words, we associate the covariant derivatives with the propagator $\langle\psi_\alpha(1)\bar\psi_{\dot{\alpha}}(2)\rangle$ (instead to the vertices) and defining a new scalar field $\psi=D^{\alpha}\psi_{\alpha}$ with the propagator: \bea \langle\psi(1)\bar\psi(2)\rangle \equiv D^\alpha_1\bar D^2_1\bar D^{\dot{\alpha}}_2 D^2_2\langle\psi_\alpha(1)\bar\psi_{\dot{\alpha}}(2)\rangle =2(\Pi_{1/2})_1\delta_{12} \ , \eea where we used the fact that $\bar D^{\dot{\alpha}}_2 D^2_2\delta_{12}=-D^2_1\bar D^{\dot{\alpha}}_1\delta_{12}$, and the factors $D^2$, $\bar{D}^2$ emerged due to properties of variational derivatives. We can also apply the same reasoning for the propagator $\langle\phi(1)\bar\phi(2)\rangle$ and for the vertices involving the scalar (anti-)chiral superfields. In summary, by transferring all covariant derivatives from the vertices () to the propagators (), we get \bea \langle V(1)V(2)\rangle&=&-\frac{1}{p^2}(\Pi_{1/2})_1\delta_{12} \ , \\ \langle\psi(1)\bar\psi(2)\rangle&=&\langle\bar\psi(1)\psi(2)\rangle=2(\Pi_{1/2})_1\delta_{12} \ , \\ \langle\phi(1)\bar\phi(2)\rangle&=&-(\Pi_-)_1\delta_{12} \ , \ \langle\bar\phi(1)\phi(2)\rangle=-(\Pi_+)_1\delta_{12} \ , \eea where $\Pi_-\equiv\Box^{-1}\bar D^2D^2$ and $\Pi_+\equiv\Box^{-1}D^2\bar D^2$ are projection operators. These propagators will connect the following new vertices: \bea \tilde S_{int}&=&\frac{1}{2}\int d^8z\big\{2MV(\psi+\bar\psi)+2(2g)\bar\Phi V\phi+2(2g)\Phi\bar\phi V+(2g)^2\bar\Phi\Phi V^2\nonumber\\ &&-4h\bar\Phi(\psi+\bar\psi)\phi-4h\Phi\bar\phi(\psi+\bar\psi)+(2h)^2\bar\Phi\Phi[\psi^2+\bar\psi^2+2\psi\bar\psi]\big\} \ , \eea where $M\equiv\frac{1}{2}(m-8gh\bar\Phi\Phi)$. Therefore, now the vertices involve only scalar superfields. In the next section, we will perform the calculations of the one-loop supergraphs using the propagators (-), written in terms of projection operators, and the vertices (), written only in terms of scalar superfields, instead of the original propagators () and the original vertices (). \section{One-loop calculations} Now, let us start the calculations of the one-loop supergraphs contributing to the KEP. Since $\Pi_{1/2}\Pi_-=\Pi_-\Pi_{1/2}=\Pi_{1/2}\Pi_+=\Pi_+\Pi_{1/2}=0$, it follows from (-) that there can be no mixed contributions containing both gauge and matter propagators at one-loop order. Therefore, the basic supergraphs contributing to the effective action in the theory under consideration are of three types: first, those with internal lines composed of propagators $\langle\psi(1)\bar\psi(2)\rangle$ only; second, those composed of propagators $\langle V(1)V(2)\rangle$ only; third, those involving alternating propagators $\langle\psi(1)\bar\psi(2)\rangle$ and $\langle V(1)V(2)\rangle$. In our graphical notation, the dashed line is for $<\psi\bar{\psi}>$ propagator, the wavy line is for $<VV>$ propagator, and the double one is for $\Phi$ or $\bar{\Phi}$ background fields. It is easy to verify that the contribution to the effective action generated by the sum of supergraphs at the Fig. 1, with simple propagators (), and the vertices $2(2h)^2(\Phi\bar{\Phi})\psi\bar{\psi}$ is zero. Indeed, it is equal to \bea \Gamma_0=\sum\limits_{n=1}^{\infty}\frac{1}{2n}[4(2h^2)\Phi\bar{\Phi}<\psi\bar{\psi}>]^n, \eea where the coefficient 4 is caused by two different contractions. Using the explicit form of the propagators (), we get \bea \Gamma_0=\sum\limits_{n=1}^{\infty}\int d^8z_1\frac{1}{2n}[4(2h^2)\Phi\bar{\Phi}\Pi_{1/2}]^n\delta^8(z_1-z_2)|_{z_1=z_2}, \eea Then, we take into account that $(\Pi_{1/2})^n=\Pi_{1/2}$, and $\Pi_{1/2}\delta^8(z_1-z_2)|_{z_1=z_2}=-2\frac{1}{\Box}\delta^4(x_1-x_2)|_{x_1=x_2}$. Carrying out the Fourier transform, we have \bea \Gamma_0=\sum\limits_{n=1}^{\infty}\frac{1}{2n}\int d^8z[4(2h^2)\Phi\bar{\Phi}]^n\int \frac{d^4k}{(2\pi)^4}\frac{1}{k^2}, \eea but within the dimensional regularization framework implemented through the replacement $d^4k\to \mu^{4-2\omega}d^{2\omega}k$, one has $\int\frac{d^{2\omega}k}{(2\pi)^{2\omega}}\frac{1}{k^2}=0$. Hence, this contribution vanishes. Now, let us sum over the vertices $(2h)^2\bar\Phi\Phi\psi^2$ and $(2h)^2\bar\Phi\Phi\bar\psi^2$. The corresponding supergraphs again exhibit structures similar to Fig. 1 with only even number of vertices. However, it is worth to point out that we can insert an arbitrary number of vertices $(2h)^2\bar\Phi\Phi\psi\bar\psi$ into the propagators $\langle\psi(1)\bar\psi(2)\rangle$. Therefore, we should firstly introduce a "dressed" propagator. In this propagator, the summation over all vertices $(2h)^2\bar\Phi\Phi\psi\bar\psi$ is performed (see Fig. 2). As a result, this dressed propagator is equal to \bea \langle \psi(1)\bar\psi(2)\rangle_D&=&\langle \psi(1)\bar\psi(2)\rangle+\int d^4\theta_3\langle \psi(1)\bar\psi(3)\rangle[(2h)^2\bar\Phi\Phi]_3\langle \psi(3)\bar\psi(2)\rangle+\int d^4\theta_3d^4\theta_4\nonumber\\ &\times&\langle \psi(1)\bar\psi(3)\rangle[(2h)^2\bar\Phi\Phi]_3\langle \psi(3)\bar\psi(4)\rangle[(2h)^2\bar\Phi\Phi]_4\langle \psi(4)\bar\psi(2)\rangle+\ldots \ . \eea By using (), integrating by parts, and summing the resultant series, we arrive at \bea \langle \psi(1)\bar\psi(2)\rangle_D=\bigg(\frac{2\Pi_{1/2}}{1-2(2h)^2\bar\Phi\Phi}\bigg)_1\delta_{12} \ . \eea Afterwards, we can compute all the contributions by noting that each one-loop supergraph above is formed by $n$ vertices like those ones given by Fig. 3. Hence, the contribution of this vertex is given by \bea Q_{13}&=&\int d^4\theta_2[(2h)^2\bar\Phi\Phi]_1\Big[\Big(\frac{2\Pi_{1/2}}{1-2(2h)^2\bar\Phi\Phi}\Big)_1\delta_{12}\Big][(2h)^2\bar\Phi\Phi]_2\Big[\Big(\frac{2\Pi_{1/2}}{1-2(2h)^2\bar\Phi\Phi}\Big)_2 \delta_{23}\Big]\nonumber\\ &=&\bigg(\frac{2(2h)^2\bar\Phi\Phi}{1-2(2h)^2\bar\Phi\Phi}\Pi_{1/2}\bigg)^2_1\delta_{13} \ . \eea It follows from the result above that the contribution of a supergraph formed by $n$ vertices is given by \bea I_n&=&\int d^4x\frac{1}{2n}\int d^4\theta_1d^4\theta_3\ldots d^4\theta_{2n-1}\int \frac{d^4p}{(2\pi)^4}Q_{13}Q_{35}\ldots Q_{2n-3,2n-1} Q_{2n-1,1} \nonumber\\ &=&\int d^4x\frac{1}{2n}\int d^4\theta_1d^4\theta_3d^4\theta_5\ldots d^4\theta_{2n-1}\int \frac{d^4p}{(2\pi)^4}\Big[\bigg(\frac{2(2h)^2\bar\Phi\Phi}{1-2(2h)^2\bar\Phi\Phi}\Pi_{1/2}\bigg)^2_1\delta_{13}\Big]\nonumber\\ &\times&\Big[\bigg(\frac{2(2h)^2\bar\Phi\Phi}{1-2(2h)^2\bar\Phi\Phi}\Pi_{1/2}\bigg)^2_3\delta_{35}\Big]\ldots \Big[\bigg(\frac{2(2h)^2\bar\Phi\Phi}{1-2(2h)^2\bar\Phi\Phi}\Pi_{1/2}\bigg)^2_{2n-1}\delta_{2n-1,1}\Big] \nonumber\\ &=&\int d^8z\frac{1}{2n}\int \frac{d^4p}{(2\pi)^4}\bigg(\frac{2(2h)^2\bar\Phi\Phi}{1-2(2h)^2\bar\Phi\Phi}\bigg)^{2n}\Pi_{1/2}\delta_{\theta\theta^{\prime}}|_{\theta=\theta^{\prime}}\ . \eea By using $\Pi_{1/2}\delta_{\theta\theta^{\prime}}|_{\theta=\theta^{\prime}}=2/p^2$, we get the effective action \bea \Gamma^{(1)}_1=\sum_{n=1}^{\infty}I_n=-\int d^8z\int \frac{d^4p}{(2\pi)^4}\frac{1}{p^2}\ln\bigg[1-\bigg(\frac{2(2h)^2\bar\Phi\Phi}{1-2(2h)^2\bar\Phi\Phi}\bigg)^{2}\bigg] \ . \eea The integral over the momenta vanishes within the dimensional regularization scheme. Therefore, \bea \Gamma^{(1)}_1=0 \ . \eea We will not calculate explicitly the one-loop supergraphs involving the gauge superfield propagators $\langle V(1)V(2)\rangle$ connecting the vertices $(2g)^2\bar\Phi\Phi V^2$, because the result is already known and described in . Therefore, it is given by \bea \Gamma^{(1)}_2=-\int d^8z\int \frac{d^4p}{(2\pi)^4}\frac{1}{p^2}\ln\bigg[1+\frac{(2g)^2\bar\Phi\Phi}{p^2}\bigg] \ . \eea Finally, let us move on to the last type of one-loop supergraphs, which involve the propagators $\langle\psi(1)\bar\psi(2)\rangle$ and $\langle V(1)V(2)\rangle$ in the internal lines connecting the vertices $MV\psi$ and $MV\bar\psi$ (see Fig. 4). As before, we can insert an arbitrary number of vertices $(2h)^2\bar\Phi\Phi\psi\bar\psi$ into the propagators $\langle\psi(1)\bar\psi(2)\rangle$. Moreover, we can also insert an arbitrary number of pairs of the vertices $(2h)^2\bar\Phi\Phi\psi^2$ and $(2h)^2\bar\Phi\Phi\bar\psi^2$ into $\langle\psi(1)\bar\psi(2)\rangle$. Since $\langle\psi(1)\bar\psi(2)\rangle$ has already been dressed by $(2h)^2\bar\Phi\Phi\psi\bar\psi$ in (-), it follows that the desired dressed propagator $\langle \psi(1)\bar\psi(2)\rangle_{2D}$ is equal to the summation over all pairs of the vertices $(2h)^2\bar\Phi\Phi\psi^2$ and $(2h)^2\bar\Phi\Phi\bar\psi^2$ into $\langle\psi(1)\bar\psi(2)\rangle_D$ (see Fig. 5). Therefore, we get \bea \langle \psi(1)\bar\psi(2)\rangle_{2D}&=&\langle \psi(1)\bar\psi(2)\rangle_D+\int d^4\theta_3d^4\theta_4\langle \psi(1)\bar\psi(3)\rangle_D[(2h)^2\bar\Phi\Phi]_3\langle \bar\psi(3)\psi(4)\rangle_D\nonumber\\ &\times&[(2h)^2\bar\Phi\Phi]_4\langle \psi(4)\bar\psi(2)\rangle_D+\int d^4\theta_3d^4\theta_4d^4\theta_5d^4\theta_6\langle \psi(1)\bar\psi(3)\rangle_D[(2h)^2\bar\Phi\Phi]_3\nonumber\\ &\times&\langle \bar\psi(3)\psi(4)\rangle_D[(2h)^2\bar\Phi\Phi]_4\langle \psi(4)\bar\psi(5)\rangle_D[(2h)^2\bar\Phi\Phi]_5\langle \bar\psi(5)\psi(6)\rangle_D\nonumber\\ &\times&[(2h)^2\bar\Phi\Phi]_6\langle \psi(6)\bar\psi(2)\rangle_D+\ldots \ . \eea After some algebraic work, we find \bea \langle \psi(1)\bar\psi(2)\rangle_{2D}=(2f(\bar\Phi\Phi)\Pi_{1/2})_1\delta_{12} \ , \ \textrm{where} \ f(\bar\Phi\Phi)\equiv\frac{1}{1-4(2h)^2\bar\Phi\Phi} \ . \eea Additionally, we can also insert an arbitrary number of vertices $(2g)^2\bar\Phi\Phi V^2$ into the propagators $\langle V(1)V(2)\rangle$. In this case, the dressed propagator $\langle V(1)V(2)\rangle_D$ is already known in the literature and it is given by \bea \langle V(1)V(2)\rangle_D=\bigg(\frac{-\Pi_{1/2}}{p^2+(2g)^2\bar\Phi\Phi}\bigg)_1\delta_{12} \ . \eea As before, we can compute all the contributions by noting that each supergraph above (Fig. 4) is formed by $n$ fragments, like those depicted in Fig. 6. This fragment yields the contribution \bea R_{13}&=&\int d^4\theta_2(M)_1\Big[\Big(\frac{-\Pi_{1/2}}{p^2+(2g)^2\bar\Phi\Phi}\Big)_1\delta_{12}\Big](M)_2\Big[(2f\Pi_{1/2})_2\delta_{23}\Big]\nonumber\\ &=&\bigg(\frac{-2fM^2\Pi_{1/2}}{p^2+(2g)^2\bar\Phi\Phi}\bigg)_1\delta_{13} \ . \eea It follows from the result above that the contribution of a supergraph formed by $n$ subgraphs is given by \bea J_n&=&\int d^4x\frac{1}{2n}\int d^4\theta_1d^4\theta_3\ldots d^4\theta_{2n-1}\int \frac{d^4p}{(2\pi)^4}R_{13}R_{35}\ldots R_{2n-3,2n-1} R_{2n-1,1} \nonumber\\ &=&\int d^4x\frac{1}{2n}\int d^4\theta_1d^4\theta_3d^4\theta_5\ldots d^4\theta_{2n-1}\int \frac{d^4p}{(2\pi)^4}\Big[\bigg(\frac{-2fM^2\Pi_{1/2}}{p^2+(2g)^2\bar\Phi\Phi}\bigg)_1\delta_{13}\Big]\nonumber\\ &\times&\Big[\bigg(\frac{-2fM^2\Pi_{1/2}}{p^2+(2g)^2\bar\Phi\Phi}\bigg)_3\delta_{35}\Big]\ldots \Big[\bigg(\frac{-2fM^2\Pi_{1/2}}{p^2+(2g)^2\bar\Phi\Phi}\bigg)_{2n-1}\delta_{2n-1,1}\Big] \nonumber\\ &=&\int d^8z\frac{1}{2n}\int \frac{d^4p}{(2\pi)^4}\bigg(\frac{-2fM^2}{p^2+(2g)^2\bar\Phi\Phi}\bigg)^n\Pi_{1/2}\delta_{\theta\theta^{\prime}}|_{\theta=\theta^{\prime}}\ . \eea Again, by using $\Pi_{1/2}\delta_{\theta\theta^{\prime}}|_{\theta=\theta^{\prime}}=2/p^2$, we get the effective action \bea \Gamma^{(1)}_3=\sum_{n=0}^{\infty}J_n=-\int d^8z\int \frac{d^4p}{(2\pi)^4}\frac{1}{p^2}\ln\bigg[1+\frac{2fM^2}{p^2+(2g)^2\bar\Phi\Phi}\bigg] \ . \eea By summing (), (), and () we obtain the total one-loop effective action \bea \Gamma^{(1)}[\bar\Phi,\Phi]=-\int d^8z\int \frac{d^4p}{(2\pi)^4}\frac{1}{p^2}\ln\big[p^2+(2g)^2\bar\Phi\Phi+2fM^2\big] \ . \eea Substituting the explicit form for $M$ and $f$, we arrive to the following result for the KEP: \bea K^{(1)}(\bar\Phi,\Phi)=-\int \frac{d^4p}{(2\pi)^4}\frac{1}{p^2}\ln\big[p^2+\frac{1}{2-8(2h)^2\bar\Phi\Phi}(m-8gh\bar\Phi\Phi)^2+(2g)^2\bar\Phi\Phi\big] \ . \eea The integral above is well-known and can be computed by using the dimensional regularization. Finally, in the limit $\omega\rightarrow2$ we find \bea K^{(1)}(\bar\Phi,\Phi)=K_{div}^{(1)}(\bar\Phi,\Phi)+K_{fin}^{(1)}(\bar\Phi,\Phi) \ , \eea where \bea K_{div}^{(1)}(\bar\Phi,\Phi)&=&\frac{1}{16\pi^2(2-\omega)}\Big[\frac{1}{2-8(2h)^2\bar\Phi\Phi}(m-8gh\bar\Phi\Phi)^2+(2g)^2\bar\Phi\Phi\Big] \ ,\\ K_{fin}^{(1)}(\bar\Phi,\Phi)&=&-\frac{1}{16\pi^2}\Big[\frac{1}{2-8(2h)^2\bar\Phi\Phi}(m-8gh\bar\Phi\Phi)^2+(2g)^2\bar\Phi\Phi\Big]\nonumber\\ &\times&\ln\frac{1}{\mu^2}\Big[\frac{1}{2-8(2h)^2\bar\Phi\Phi}(m-8gh\bar\Phi\Phi)^2+(2g)^2\bar\Phi\Phi\Big] \ , \eea and $\mu$ is an arbitrary scale required on dimensional grounds. Notice that the one-loop KEP (-) is divergent. Moreover, we notice that the divergent part () is given by an infinite power series in $\bar\Phi\Phi$. Therefore, the theory under consideration is non-renormalizable and it must be interpreted as an effective field theory below some energy scale chosen on the basis of phenomenological considerations . In particular, let us take $h=0$ in (). This choice corresponds to a minimal coupling between the gauge scalar superfield and the matter chiral superfields [see ()]. Therefore, \bea K_{div}^{(1)}(\bar\Phi,\Phi)&=&\frac{(2g)^2\bar\Phi\Phi}{16\pi^2(2-\omega)} \ ,\\ K_{fin}^{(1)}(\bar\Phi,\Phi)&=&-\frac{1}{32\pi^2}\big[m^2+2(2g)^2\bar\Phi\Phi\big]\ln\frac{1}{2\mu^2}\big[m^2+2(2g)^2\bar\Phi\Phi\big] \ . \eea In this case, we notice that the divergent term () is proportional to $\bar\Phi\Phi$. Therefore, in order to remove divergences, we can insert a similar one-loop counterterm as the one used in the SQED. Moreover, if we take the massless case in (-), we recover the one-loop KEP for the usual SQED . \section{Summary} We formulated a new theory involving coupling of three superfields of different natures: a chiral spinor gauge superfield originally introduced in together with the usual real scalar gauge superfield and the chiral scalar matter superfield. For this theory, we developed a superfield procedure for calculating the one-loop effective potential which we successfully found. The procedure does not essentially differ from the usual supergauge theories with the rather similar structure of the one-loop contribution. The fact that the new theory is non-renormalizable is not unexpected since many non-polynomial supersymmetric theories are non-renormalizable . We expect that this theory, besides of the classical studies in the cosmic string context, can be used as an ingredient of possible phenomenologically interesting supersymmetric gauge theories involving several gauge (super)fields with some of them being massive. \vspace*{3mm} {\bf Acknowledgments.} This work was partially supported by Conselho Nacional de Desenvolvimento Cient\'{\i}fico e Tecnol\'{o}gico (CNPq). The work by A. Yu. P. has been partially supported by the CNPq project No. 303438/2012-6. \begin{thebibliography}{100} \bibitem{SGRS} S. J. Gates, M. T. Grisaru, M. Rocek, W. Siegel. Superspace or One Thousand and One Lessons in Supersymmetry. Benjamin/Cummings, (1983), hep-th/0108200. \bibitem{BuKu} I. L. Buchbinder and S. M. Kuzenko, Ideas and Methods of Supersymmetry and Supergravity. IOP Publishing, Bristol and Philadelphia, (1998). \bibitem{ourcourse} A. Yu. Petrov, ``Quantum superfield supersymmetry'', hep-th/0106094. \bibitem{Siegel} W. Siegel, Phys. Lett. B85, 333 (1979). \bibitem{Fe} C. N. Ferreira, J. A. Helay\"{e}l-Neto, M. B. D. S. M. Porto, Nucl. Phys. B620, 181 (2002). \bibitem{JL} J. Louis, J. Swiebodzinski, Eur. Phys. J. C51, 731 (2007), hep-th/0702211. \bibitem{WZ} I. L. Buchbinder, S. M. Kuzenko, J. V. Yarevskaya, Nucl. Phys. B411, 665 (1994). \bibitem{WZ1}I. L. Buchbinder, S. M. Kuzenko, A. Yu. Petrov, Phys. Lett. B321, 372 (1994); Phys. At. Nucl. 59, 148 (1996). \bibitem{GC} I. L. Buchbinder, A. Yu. Petrov, Phys. Lett. B461, 209 (1999), hep-th/9905062; I. L. Buchbinder, M. Cvetic, A. Yu. Petrov, Mod. Phys. Lett. A15, 783 (2000), hep-th/9903243; Nucl. Phys. B571, 358 (2000), hep-th/9906141. \bibitem{SYM} B. de Wit, M. Grisaru, M. Rocek, Phys. Lett. B374, 297 (1996), hep-th/9601115; M. Grisaru, M. Rocek, R. von Unge, Phys. Lett. B383, 415 (1996), hep-th/9605149; A. De Giovanni, M. Grisaru, M. Rocek, R. von Unge, D. Zanon, Phys. Lett. B409, 251 (1997), hep-th/9706013. \bibitem{Chris} H. R. Christiansen, M. S. Cunha, J. A. Helay\"{e}l-Neto, L. R. U. Manssur, A. L. M. A. Nogueira, Int. J. Mod. Phys. A14, 147 (1999). \bibitem{BO} R. Jackiw. D9, 1686 (1974); I. L. Buchbinder, S. D. Odintsov, I. L. Shapiro. Effective action in quantum gravity. IOP Publishing, Bristol and Philadelphia, (1992). \bibitem{Coleman} S. Coleman, Aspects of Symmetry, Cambridge Univ. Press, (1985). \bibitem{Our} F. S. Gama, M. Gomes, J. R. Nascimento, A. Yu. Petrov, A. J. da Silva, Phys. Rev. D84, 045001 (2011), arXiv: 1101.0724; F. S. Gama, M. Gomes, J. R. Nascimento, A. Yu. Petrov, A. J. da Silva, Phys.Rev. D89 (2014) 085018, arXiv: 1401.6839. \bibitem{Bur} C. P. Burgess, Ann. Rev. Nucl. Part. Sci. 57, 329 (2007), hep-th/0701053. \bibitem{Brignole} A. Brignole, Nucl. Phys. B579, 101 (2000), hep-th/0001121. \end{thebibliography} |
1501.04062 | Title: Infinite symmetric group and bordisms of pseudomanifolds
Abstract: We consider a category whose morphisms are bordisms of $n$-dimensional
pseudomanifolds equipped with a certain additional structure (coloring). On the
other hand, we consider the product $G$ of $(n+1)$ copies of infinite symmetric
group. We show that unitary representations of $G$ produce functors from the
category of $(n-1)$-dimensional bordisms to the category of Hilbert spaces and
bounded linear operators.
Body: \begin{center} {\bf\Large Infinite symmetric group and bordisms of pseudomanifolds} \bigskip \sc \large Alexander A. Gaifullin \footnote{Supported by RFBR (project 14-01-92612), by grants of the President of the Russian Federation (NSh-4833.2014.1 and MD-2969.2014.1), and by Dmitri Zimin's ``Dynasty'' foundation.}, Yury A. Neretin \end{center} {\small We consider a category whose morphisms are bordisms of $n$-dimensional pseudomanifolds equipped with a certain additional structure (coloring). On the other hand, we consider the product $G$ of $(n+1)$ copies of infinite symmetric group. We show that unitary representations of $G$ produce functors from the category of $(n-1)$-dimensional bordisms to the category of Hilbert spaces and bounded linear operators.} \section{Pseudomanifolds and pseudobordisms} \COUNTERS First, we fix several definitions. \sm {\bf\punct Simplcial cell complexes.} Consider a disjoint union $\coprod \Xi_j$ of a finite collection of simplices $\Xi_j$. We consider a topological quotient space $\Sigma$ of $\coprod \Xi_j$ with respect to certain equivalence relation. The quotient must satisfy the following properties \sm a) For any simplex $\Xi_i$, the tautological map $\xi_i:\Xi_i\to \Sigma$ is an embedding. Therefore we can think of $\Xi_i$ as of a subset of $\Sigma$. \sm b) For any pair of simplices $\Xi_i$, $\Xi_j$, the intersection $\xi_i^{-1}\bigl(\xi_i(\Xi_i)\cap \xi_j(\Xi_j)\bigr)\subset \Xi_i$ is a union of faces of $\Xi_i$ and the partially defined map $$ \Xi_i\stackrel{\xi_i}{\longrightarrow} \Sigma\stackrel{\xi_j^{-1}}{\longrightarrow} \Xi_j $$ is affine on each face. \sm We shall call such quotients {\it simplicial cell complexes}. \sm {\sc Remark on terminology.} There are two similar (and more common) definitions of spaces composed from simplices (see, e.g., ). The first one is a more restrictive definition of ``a simplicial complex''. In this case, a non-empty intersection of two faces is a (unique) face. See examples of simplicial cell complexes, which are not simplicial complexes in Fig.2 and Fig.3.b. A more wide class of simplicial spaces are $\Delta$-complexes, in this case glueing of a simplex with itself along faces is allowed (as for standard 1-vertex triangulations of two-dimensional surfaces), see Fig 1. \hfill $\lozenge$ \sm {\bf\punct Pseudomanifolds.} A {\it pseudomanifold} of dimension $n$ is a simplicial cell complex such that \sm a) Each face is contained in an $n$-dimensional face. We call $n$-dimensional faces {\it chambers.} \sm b) Each $(n-1)$-dimensional face is contained in precisely two chambers. \sm See, e.g., , . \sm {\sc Remark.} Any cycle of singular $\Z$-homologies in a topological space can be realized as an image of a pseudo-manifold (this is more-or-less obvious). Recall that there are cycles in manifolds, which cannot be realized as images of manifolds. \hfill $\lozenge$ \sm {\sc Remark on terminology.} In literature, there exists another variant of a definition of a pseudomanifold. Seifert, Threlfall, impose two additional requiments: a pseudomanifold must be a simplicial complex and must be 'strongly connected'. The latter conditions means that the complement of the union of faces of codimension 2 must be connected. \hfill $\lozenge$ \sm \sm {\bf \punct Normal pseudomanifolds and normalization. } {\it Links.} Let $\Sigma$ be a pseudomanifold, let $\Gamma$ be its $k$-dimensional face. Consider all $(k+1)$-dimensional faces $\Phi_j$ of $\Sigma$ containing $\Gamma$ and choose a point $\phi_j$ in the relative interior of each face $\Phi_j$. For each face $\Psi_m\supset \Gamma$ we consider the convex hull of all points $\phi_j$ that are contained in $\Psi_m$. The link of $\Gamma$ is the simplicial cell complex whose faces are such convex hulls. \sm {\it Normal pseudomanifolds.} A pseudomanifold is {\it normal} if the link of any face of codimension $\ge 2$ is connected. \sm {\sc Example.} Consider a triangulated compact two-dimensional surface $\Sigma$. Let $a$, $b$ be two vertices that are not connected by an edge. Glueing together $a$ and $b$ we get a pseudomanifold which is not normal, see Fig.2. \hfill $\lozenge$ \sm {\it Normalization.} For any pseudomanifold $\Sigma$ there is a unique {\it normalization} (), i.e. a normal pseudomanifold $\wt\Sigma$ and a map $\pi:\wt\Sigma \to \Sigma$ such that \sm --- restriction of $\pi$ to any face of $\wt\Sigma$ is an affine bijective map of faces. \sm --- the map $\pi$ send different $n$-dimensional and $(n-1)$-dimensional faces to different faces. \sm {\it A construction of the normalization.} To obtain a normalization of $\Sigma$ we cut a pseudomanifold $\Sigma$ into a disjoint collection of chambers $\Xi_i$. As above, denote by $\xi_i:\Xi_i\to \Sigma_j$ the embedding of $\Xi_i$ to $\Sigma$. Let $x\in\Xi_i$, $y\in \Xi_j$. We say that $x\sim y$ if $\xi_i(x)=\xi_j(y)$ and this point is contained in a common $(n-1)$-dimensional face of the chambers $\xi_i(\Xi_i)$ and $\xi_j(\Xi_j)$. We extend $\sim$ to an equivalence relation by the transitivity. The quotient of $\coprod \Xi_i$ is the normalization of $\Sigma$. The following way of normalization is more visual. Let $\Sigma$ be non-normal. Let $\Xi$ be a face of codimension $2$ with link consisting of $m$ connected components. Consider a small closed neighborhood $\cO$ of $\Xi$ in $\Sigma$. Then $\cO\setminus \Xi$ is disconnected and consists of $m$ components, say $\cO_1$,\dots, $\cO_m$. Let $\ov\cO_j$ be the closure of $\cO_j$ in $\Sigma$, $\ov\cO_j=\cO_j\cup \Xi_j$. We replace $\cO$ by the disjoint union of $\ov\cO_j$ and get a new pseudomanifold $\Sigma'$ (in Fig.2, we duplicate the upper vertex). Then we repeat the same operation to another stratum with disconnected link. These operation enlarges number of strata of codimension $\ge 2$, the strata of dimension $n$ and $(n-1)$ remain the same (and the incidence of these strata is preserved). Therefore the process is finite and we get a normal pseudomanifold. \hfill $\lozenge$ \smallskip {\bf\punct Colored pseudomanifolds.} Consider an $n$-dimensional {\it normal} pseudomanifold $\Sigma$. A coloring of $\Sigma$ is the following structure \sm a) To any chamber we assign a sign $(+)$ or $(-)$. Chambers adjacent to plus-chambers are minus-chamber and vise versa. \sm b) Choose $n+1$ colors (say, red, blue, green, orange, etc.). Each vertex of the complex is colored in such a way that the colors of vertices of each chamber are pairwise different. \sm c) All $(n-1)$-dimensional faces are colored, in such a way that colors of faces of a chamber are pairwise different, and a color of a face coincides with a color of the opposite vertex of any chamber containing this face. \sm We say that a {\it double-chamber} is a colored $n$-dimensional pseudomanifold obtained from two identical copies $\Delta_1$, $\Delta_2$ of an $n$-dimensional simplex by identification of the corresponding $x\in \Delta_1$, $x\in\Delta_2$ of the boundaries of $\Delta_1$, $\Delta_2$. \sm {\sc Remark.} Colored pseudomanifolds were introduced by Pezzana and Ferri in 1975-1976, see , , , . \hfill $\lozenge$ \sm {\bf\punct Colored pseudobordisms. } Fix $n\ge 1$. We define a category $\PBor$ of pseudobordisms. Its objects are nonnegative integers. A morphism $\beta\to\alpha$ is the following collection of data \sm 1) A colored $n$-dimensional pseudomanifold (generally, disconnected). \sm 2) An injective map of the set $\{1,2,\dots,\alpha\}$ to the set of plus-chambers and an injective map of the set $\{1,2,\dots,\beta\}$ to the set of minus-chambers In other words, we assign labels $1$, \dots, $\alpha$ to some plus-chambers. and labels $1$, \dots, $\beta$ to some minus-chambers. \sm We require that each double-chamber has at least one label. \smallskip {\it Composition.} Let $\Sigma\in \Mor(\beta,\alpha)$, $\Lambda\in\Mor(\gamma,\beta)$. We define their composition $\Sigma\diamond \Lambda$ as follows. Remove interiors of labeled minus-chambers of $\Sigma$ and interiors of labeled plus-chambers of $\Lambda$. Next, for each $s\le \beta$, we glue boundaries of the minus-chamber of $\Sigma$ with label $s$ with the boundary of the plus-chamber of $\Lambda$ with label $s$ according the simplicial structure of boundaries and colorings of $(n-1)$-simplices. Next, we normalize the resulting pseudomanifold. Finally we remove label-less double chambers (such components can arise as a result of gluing of two label-keeping double chambers). \sm {\it Involution.} For a morphism $\Sigma\in \Mor(\beta,\alpha)$ we define the morphism $\Sigma^*\in \Mor(\alpha,\beta)$ by changing of signs on chambers. Thus we get an {\it involution} in the category $\PBor$. For any $T\in \Mor(\beta,\alpha)$, $S\in\Mor(\gamma,\beta)$ we have $$ (S\diamond T)^*=T^*\diamond S^* $$ \sm {\bf \punct Further structure of the paper.} Below we construct a family of functors from the category of pseudobordisms to the category of Hilbert spaces and bounded operators. This means that for each $\alpha$ we construct a Hilbert space $H(\alpha)$ and for each morphism $\Sigma\in\Mor(\beta,\alpha)$ we construct an operator $\rho(\Sigma):\,H(\beta)\to H(\alpha)$ such that for any $\alpha$, $\beta$, $\gamma$ and any $\Sigma\in \Mor(\beta,\alpha)$, $\Xi\in\Mor(\gamma,\beta)$, $$ \rho(\Sigma\diamond \Xi)=\rho(\Sigma)\, \rho(\Xi) .$$ Also, representations obtained below satisfy properties $$ \rho(\Sigma)^*=\rho(\Sigma^*),\qquad \|\rho(\Sigma)\|\le 1 . $$ In fact, such functors arise in a natural way from the representation theory of infinite symmetric groups. In Section 2, we introduce a category of double cosets on the product of $(n+1)$ copies of an infinite symmetric group. In Section 3, we show that the category of double cosets is equivalent to the category of pseudo-bordisms (this is the main statement of this note). In Section 4, we construct a family of representations of this category (statements of this section are more-or-less automatic). For $n=1$ the construction reduces to Olshanski , for $n=2$ it coincides with . \section{Multiplication of double cosets} \COUNTERS {\bf\punct Symmetric groups. Notation. } Let $S(L)$ be the group of permutations of a set with $L$ elements. Denote by $K=S(\infty)$ the group of finitely supported permutations of $\N$. By $\ov K=\ov S(\infty)$ we denote the group of all permutations of $\N$. Denote by $K(\alpha)\subset K$, $\ov K(\alpha)\subset \ov K$ the stabilizers of points $1$, \dots, $\alpha$. We equip $\ov S(\infty)$ with a natural topology assuming that the subgroups $K(\alpha)$ are open. Sometimes we will represent elements of symmetric groups as $0-1$-matrices. \smallskip {\bf\punct Multiplication of double cosets.} Denote the product of $(n+1)$ copies of $S(\infty)$ by $G$. Denote by $K\simeq S(\infty)$ the diagonal subgroup in $G$, its elements have the form $(g,g,\dots,g)$. Consider double cosets $K(\alpha)\setminus G/K(\beta)$, i.e., elements of $G$ defined up to the equivalence $$ g\sim k_1 g k_2,\qquad k_1\in K(\alpha),\, k_2\in K(\beta) $$ We wish to define product of double cosets $$ K(\alpha)\setminus G/K(\beta)\,\,\times\,\, K(\beta)\setminus G/K(\gamma)\,\,\to\,\, K(\alpha)\setminus G/K(\gamma). $$ For this purpose, define elements $\theta_\sigma[j]\in K(\sigma)$ by $$ \theta_\sigma[j]:= \begin{pmatrix} \1_\sigma&0&0&0 \\ 0&0&\1_j&0 \\ 0&\1_j&0&0 \\ 0&0&0&\1_\infty \end{pmatrix} ,$$ where $\1_j$ denotes the unit matrix of order $j$. \begin{proposition} Let $$\frg\in K(\alpha)\setminus G/ K(\beta), \quad\frh\in K(\beta)\setminus G/K(\gamma)$$ be double cosets. Let $g$, $h\in G$ be their representatives. Then the sequence \begin{equation} \frr_j:= K(\alpha)\cdot g \theta_\beta[j] h\cdot K(\gamma)\,\,\,\in K(\alpha)\setminus G/K(\gamma) \end{equation} is eventually constant. The limit value of $\frr_j$ does not depend on a choice of representatives $g\in\frg$ and $h\in\frh$. Moreover, if $g$, $h\in S(L)^{n+1}\subset S(\infty)^{n+1}$, then it is sufficient to consider $j=L-\beta$. \end{proposition} We define the product $$\frg \circ \frh\in K(\alpha)\setminus G/K(\gamma)$$ of double cosets as the limit value of the sequence (). \smallskip {\sc Formula for product.} Represent $g$ as a collection of block matrices $\bigl(g^{(1)},\dots, g^{(n+1)}\bigr)$ of size $$\bigl(\alpha+(L-\alpha)+(L-\beta)+\infty\bigr)\times \bigl(\beta+(L-\beta)+(L-\beta)+\infty\bigr),$$ represent $h$ as a collection of block matrices $\bigl(h^{(1)},\dots, h^{(n+1)}\bigr)$ of size $$\bigl(\beta+(L-\beta)+(L-\beta)+\infty\bigr)\times\bigl(\gamma+(L-\gamma)+(L-\beta)+\infty\bigr)$$ \begin{equation} g^{(k)}= \begin{pmatrix} a^{(k)}&b^{(k)}&0&0\\c^{(k)}&d^{(k)}&0&0\\ 0&0&\1_{L-\beta}&0 \\0&0&0&\1_\infty \end{pmatrix} ,\qquad h^{(k)}= \begin{pmatrix} p^{(k)}&q^{(k)}&0&0\\r^{(k)}&t^{(k)}&0&0 \\0&0&\1_{L-\beta}&0\\0&0&0&\1_\infty \end{pmatrix} . \end{equation} Then we write a representative of the double coset $\frg\circ\frh$ as $$ (g\circ h)^{(k)}:= g\cdot \theta_{\beta}[L-\beta]\cdot h = \begin{pmatrix} a^{(k)}p^{(k)}&a^{(k)}q^{(k)}&b^{(k)}&0\\ c^{(k)}p^{(k)}&c^{(k)}q^{(k)}&d^{(k)}&0\\r^{(k)}&t^{(k)}&0&0\\ 0&0&0&\1_\infty \end{pmatrix} . $$ {\sc Proof.} First, we show that the result does not depend on a choice of $j$ Denote $$\mu=L-\beta, \quad \nu=L-\alpha,\quad \kappa=L-\gamma .$$ Preserving the previous notation for $g^{(k)}$, $h^{(k)}$, we write $$ (g\cdot \theta_\beta[\mu+j]\cdot h)^{(k)}= \begin{pmatrix} a^{(k)}p^{(k)}&a^{(k)}q^{(k)}&0&b^{(k)}&0&0\\ c^{(k)}p^{(k)}&c^{(k)}q^{(k)}&0&d^{(k)}&0&0\\ 0&0&0&0&\1_j&0\\ r^{(k)}&t^{(k)}&0&0&0&0\\ 0&0&\1_j&0&0&0\\ 0&0&0&0&0&\1_\infty \end{pmatrix} .$$ This coincides with \begin{multline*} \begin{pmatrix} \1_\alpha&0&0&0&0&0\\ 0&\1_\nu&0&0&0&0\\ 0&0&0&\1_j& 0&0\\ 0& 0&\1_\mu &0&0&0\\ 0&0&0&0&\1_j&0\\ 0&0&0&0&0&\1_\infty \end{pmatrix} \begin{pmatrix} a^{(k)}p^{(k)}&a^{(k)}q^{(k)}&b^{(k)}&0&0&0\\ c^{(k)}p^{(k)}&c^{(k)}q^{(k)}&d^{(k)}&0&0&0 \\r^{(k)}&t^{(k)}&0&0&0&0\\ 0&0&0&\1_j&0&0\\ 0&0&0&0&\1_j&0\\ 0&0&0&0&0&\1_\infty \end{pmatrix} \times \\ \times \begin{pmatrix} \1_\gamma&0&0&0&0&0\\ 0&\1_\kappa&0&0&0&0\\ 0&0&0&\1_\mu& 0&0\\ 0&0&0&0&\1_j&0\\ 0& 0&\1_j &0&0&0\\ 0&0&0&0&0&\1_\infty \end{pmatrix} . \end{multline*} Next, we show that () does not depend on the choice of representatives of double cosets. To be definite, replace a collection $\{g^{(k)}\}$ in () by $$ \begin{pmatrix} \1_\alpha &0&0\\0&u&0\\0&0&\1_\infty \end{pmatrix} \begin{pmatrix} a^{(k)}&b^{(k)}&0\\c^{(k)}&d^{(k)}&0\\0&0&\1_\infty \end{pmatrix} \begin{pmatrix} \1_\alpha &0&0\\0&v&0\\0&0&\1_\infty \end{pmatrix} = \begin{pmatrix} a^{(k)}&b^{(k)}v&0\\uc^{(k)}&ud^{(k)}v&0\\0&0&\1_\infty \end{pmatrix} . $$ Then $(g\circ h)^{(k)}$ is \begin{multline*} \begin{pmatrix} a^{(k)}p^{(k)}&a^{(k)}q^{(k)}&b^{(k)}v&0\\ uc^{(k)}p^{(k)}&uc^{(k)}q^{(k)}&ud^{(k)}v&0\\r^{(k)}&t^{(k)}&0&0\\ 0&0&0&\1_\infty \end{pmatrix} =\\= \begin{pmatrix} \1_\alpha&0&0&0\\ 0&u&0&0\\0&0&\1_\mu &0\\ 0&0&0&\1_\infty \end{pmatrix} \begin{pmatrix} a^{(k)}p^{(k)}&a^{(k)}q^{(k)}&b^{(k)}&0\\ c^{(k)}p^{(k)}&c^{(k)}q^{(k)}&d^{(k)}&0\\r^{(k)}&t^{(k)}&0&0\\ 0&0&0&\1_\infty \end{pmatrix} \begin{pmatrix} \1_\gamma&0&0&0\\ 0&\1_\kappa &0&0\\0&0&v &0\\ 0&0&0&\1_\infty \end{pmatrix} \end{multline*} This completes the proof. \hfill $\square$ \smallskip \begin{proposition} The $\circ$-product is associative. \end{proposition} {\sc Proof.} Let $g$, $h\in G$ be as above, and let $w=(w^{(1)},\dots, w^{(n+1)})\in G$ be given by $$ w^{(k)}= \begin{pmatrix} x^{(k)}&z^{(k)}&0\\y^{(k)}&u^{(k)}&0\\0&0&\1_\infty \end{pmatrix} . $$ Evaluating $(g\circ h)\circ w$ and $g\circ (h\circ w)$ we get the matrices \begin{equation} \begin{pmatrix} a^{(k)}p^{(k)}x^{(k)}& a^{(k)}p^{(k)}y^{(k)} & a^{(k)}q^{(k)} & b^{(k)}& 0& 0\\ c^{(k)}p^{(k)}x^{(k)}& c^{(k)}p^{(k)}y^{(k)} & c^{(k)}q^{(k)} & d^{(k)}& 0& 0\\ 0&0&0&0&\1&0\\ r^{(k)}x^{(k)} & r^{(k)}y^{(k)} & t^{(k)} &0&0& 0\\ z^{(k)}&u^{(k)}&0&0&0& 0\\ 0&0&0&0& 0&\1_\infty \end{pmatrix} , \end{equation} \begin{equation} \begin{pmatrix} a^{(k)}p^{(k)}x^{(k)}& a^{(k)}p^{(k)}y^{(k)} & a^{(k)}q^{(k)}& 0 & b^{(k)}& 0\\ c^{(k)}p^{(k)}x^{(k)}& c^{(k)}p^{(k)}y^{(k)} & c^{(k)}q^{(k)}& 0 & d^{(k)}& 0\\ r^{(k)}x^{(k)} & r^{(k)}y^{(k)} & t^{(k)}& 0& 0 &0\\ z^{(k)}&u^{(k)}&0&0&0& 0\\ 0&0&0& \1&0& 0\\ 0&0&0&0& 0&\1_\infty \end{pmatrix} . \end{equation} Both matrices are elements of the double coset containing \begin{equation} \begin{pmatrix} a^{(k)}p^{(k)}x^{(k)}& a^{(k)}p^{(k)}y^{(k)} & a^{(k)}q^{(k)} & b^{(k)}& 0& 0\\ c^{(k)}p^{(k)}x^{(k)}& c^{(k)}p^{(k)}y^{(k)} & c^{(k)}q^{(k)} & d^{(k)}& 0& 0\\ r^{(k)}x^{(k)} & r^{(k)}y^{(k)} & t^{(k)} &0&0& 0\\ z^{(k)}&u^{(k)}&0&0&0& 0\\ 0&0&0&0& \1&0\\ 0&0&0&0& 0&\1_\infty \end{pmatrix} ,\end{equation} matrix () is obtained from () by a permutation of rows, and matrix () is obtained from () by a permutation of columns. \hfill $\square$ \smallskip Thus we get a category $\cK$, whose objects are nonnegative integers, and $$ \Mor_\cK(\beta,\alpha):=K(\alpha)\setminus G/K(\beta) . $$ A product is the product of double cosets. \smallskip {\bf\punct Involution.} The map $g\mapsto g^{-1}$ induces the map $\frg\mapsto \frg^*$ of double cosets $$ K(\alpha)\setminus G/K(\beta) \to K(\beta)\setminus G/K(\alpha) .$$ Evidently, $(\frg\circ \frh)^*=\frh^*\circ \frg^*$. \section{Correspondence} \COUNTERS {\bf\punct A correspondence between pseudomanifolds and symmetric groups. } Denote by $S(L)$ the symmetric group of order $L$. Denote by $$ S(L)^{n+1}:=S(L)\times\dots\times S(L) $$ the direct product of $n+1$ copies of $S(L)$, we assign $n+1$ colors, say, red, blue, orange, etc., to copies of $S(L)$. Consider a colored pseudomanifold $\Sigma$ with $2L$ chambers. We say that a labeling of $\Sigma$ is a bijection of the set $\{1,2,\dots,L\}$ with the set of plus-chambers of $\Sigma$ and a bijection of $\{1,2,\dots,L\}$ with the set of minus-chambers of $\Sigma$. \begin{theorem} There is a canonical one-to-one correspondence between the group $S(L)^{n+1}$ and the set of all labeled colored normal $n$-dimensional pseudo-manifolds with $2L$ chambers. \end{theorem} {\sc Remark.} This correspondence for $n=2$ was proposed in . Earlier there was a construction of Pezzana--Ferri (1975-1976), , , . They considered bipartie $(n+1)$-valent graphs whose edges are colored in $(n+1)$ colors, edges adjacent to a given vertex have pairwise different colors. Such graphs correspond to colored pseudomanifolds. In -- there was considered an action of free product $\Z_2*\dots*\Z_2$ of $n$ copies of $\Z_2$ on the set of chambers of a colored pseudomanifold. A construction relative to the present construction was considered in . \hfill $\lozenge$ \sm {\sc Construction of the correspondence.} Indeed, consider a labeled colored normal pseudomanifold $\Sigma$ with $2L$ chambers. Fix a color (say, blue). Consider all blue $(n-1)$-dimensional faces $A_1$, $A_2$, \dots. Each blue face $A_j$ is contained in the plus-chamber with some label $p(j)$ and in the minus-chamber with some label $q(j)$. We take an element of the symmetric group $S(L)$ setting $p(j)\mapsto q(j)$ for all blue faces $A_j$. We repeat the same construction for all colors and obtain a tuple $(g^{(1)},\dots,g^{(n+1)})\in S(L)^{n+1}$. Conversely, consider an element of the group $S(L)^{n+1}$. Consider $L$ labeled copies of a colored chamber (plus-chambers) and another collection of $L$ labeled copies of the same chamber with another orientation (minus-chambers). Let the blue permutation send $\alpha\mapsto\beta$. Then we glue the the plus-chamber with label $\alpha$ with the minus-chamber with label $\beta$ along the blue face (preserving colorings of vertices). The same is done for all colors. The obtained pseudomanifold $\Sigma$ is normal because the normalization procedure from Subsection applied to $\Sigma$ produces $\Sigma$ itself. \hfill $\square$ \smallskip {\bf\punct The multiplication in symmetric group and pseudomanifolds.} Describe the multiplication in $S(L)^{n+1}$ in a geometric language. Consider two labeled colored pseudomanifolds $\Sigma$, $\Xi$. Remove interiors of minus-chambers of $\Sigma$ remembering a minus-label on each face of a removed chamber, denote the topological space obtained in this way by $\Sigma_-$. All $(n-1)$-faces of $\Sigma_-$ are colored and labeled. In the same way, we remove plus-chambers from $\Xi$ and get a complex $\Xi_+$. Next, we glue the corresponding faces of $\Sigma_-$ and $\Xi_+$ (with coinciding colors and labels according coloring of vertices). In this way, we get a pseudomanifold and consider its normalization. \smallskip {\bf\punct Infinite case.} We say that an {\it infinite pseudo-manifold} is a disjoint union of a countable collection of compact pseudomanifolds such that all but a finite number of its components are double-chambers. We define a colored infinite pseudo-manifold as above. A labeled pseudomanifold is a colored pseudomanifold with a numbering of plus-chambers by natural numbers and a numbering of minus-chambers by natural numbers such that all but a finite number of double-chambers have the same labels on both chambers. \begin{theorem} There is a canonical one-to-one correspondence between the group $S(\infty)^{n+1}$ and the set of all labeled colored normal infinite pseudomanifolds. \end{theorem} The correspondence is given by the same construction obtained as above. \sm {\bf \punct Equivalence of categories.} \begin{theorem} The category $\cK$ of double cosets and the category $\PBor$ of pseudo-bordisms are equivalent. The equivalence is given by the following construction. \end{theorem} {\sc Correspondence $\Mor_\cK(\beta,\alpha)\longleftrightarrow \Mor_\PBor(\beta,\alpha)$.} Let $\frg\in K(\alpha)\setminus G/ K(\beta)$ be a double coset. Let $g\in \frg$ be its representative. Consider the corresponding labeled colored pseudomanifold. A left multiplication $g\mapsto ug$ by an element $u\in K(\alpha)$ is equivalent to a permutation $u$ of labels $\alpha+1$, $\alpha+2$, \dots on plus-chambers. A right multiplication $g\mapsto gv$ by an element $v\in K(\beta)$ is equivalent to a permutation of labels $\beta+1$, $\beta+2$, \dots on minus-chambers. Thus passing to double cosets is equivalent to forgetting labels $>\alpha$ on plus-chambers and labels $>\beta$ on minus-chambers. Notice that all but a finite number of double-chambers are label-less. Such label-less double chambers can be forgotten. Thus we get a pseudobordism. \smallskip {\sc Correspondence of products.} Let $g$, $h$ be representatives of double cosets. Let $\Sigma$, $\Xi$ be the corresponding infinite labeled colored pseudomanifolds. Let $\Sigma'$ correspond to $g \theta_\beta[j]$ , where $j$ is large. We multiply $g \theta_\beta[j]$ by $h$ according to the rule in Subsection . Notice that minus-chambers of $\Sigma'$ with labels $>\beta$ are glued with double-chambers. Plus-chambers of $\Xi$ with labels $>\beta$ are also glued with double chambers. Both operations yield a changing of labels on chambers. This means that in fact we glue together only chambers with labels $\le\beta$, in remaining cases we change labels on chambers only. Afterwards we forget all labels which are grater than $\beta$ and get the operation described in Subsection . \hfill $\square$ \section{Representations} \COUNTERS Here we construct a family of representations of the group $G$. This produces representations of the category of double cosets and therefore representations of the category of pseudobordisms. The construction is a special case of (where the case $n=2$ was considered), more ways of constructions of representations of the group $G$, see in , . \sm {\bf\punct The group $\G$.} We define an 'intermediate' group $\G$, $$ S(\infty)^{n+1}\subset \G\subset \ov S(\infty)^{n+1} ,$$ consisting of tuples $(g_1,\dots,g_{n+1})\in \ov S(\infty)^{n+1}$ such that $g_i g_j^{-1}\in S(\infty)$ for all $i$, $j$. Denote by $\K\simeq \ov S(\infty)$ the diagonal subgroup consisting of tuples $(g,\dots,g)$. Define the subgroup $\K(\alpha)$ to be the group of all $(h,\dots,h)$, where $h$ fixes $1$, \dots, $\alpha$. Define the topology on $\G$ assuming that subgroups $\K(\alpha)$ are open. Obviously, there is the identification of double cosets $$ K(\alpha)\setminus G/K(\beta)\simeq \K(\alpha)\setminus \G/\K(\beta) . $$ \smallskip {\bf\punct A family of representation of $\G$.} Consider $(n+1)$ Hilbert spaces \footnote{We admit arbitrary, finite-dimensional or infinite-dimensional, separable Hilbert spaces.} $V_{red}$, $V_{orange}$, $V_{blue}$, \dots. Consider their tensor product $$ \cV=V_{red}\otimes V_{blue} \otimes V_{green}\otimes \dots $$ Fix a unit vector $\xi\in \cV$. Consider a countable tensor product of Hilbert spaces \begin{multline} \frV= (\cV,\xi)\otimes (\cV,\xi)\otimes (\cV,\xi)\otimes \dots =\\= (V_{red}\otimes V_{blue} \otimes \dots\,\,,\xi)\otimes (V_{red}\otimes V_{blue} \otimes\dots\,\,,\xi)\otimes \dots \end{multline} (for a definition of tensor products, see ). Denote $$ \frv=\xi\otimes \xi\otimes\dots \in \frV . $$ We define a representation $\nu$ of $\G$ in $\frV$ in the following way. The 'red' copy of $S(\infty)$ acts by permutations of factors $V_{red}$. The 'blue' copy $S_\infty$ acts by permutation of factors $V_{blue}$, etc. Thus we get an action of the group $S(\infty)^{n+1}$. The diagonal $\K=\ov S(\infty)$ acts by permutations of factors $\cV$. \sm {\sc Remark.} For type I groups $H_1$, $H_2$ irreducible unitary representations of $H_1\times H_2$ are tensor products of representations of $H_1$ and $H_2$ (see, e.g., 13.1.8). However, $S(\infty)$ is not a type I group. {\it Representations of $S(\infty)^{n+1}$ constructed above are not tensor products of representations of $S(\infty)$.} \hfill $\lozenge$ \smallskip {\bf \punct Representations of the category $\cK$.} Consider a unitary representation $\rho$ of the group $\G$ in a Hilbert space $H$. For $\alpha=0$, $1$, $2$, \dots consider the subspace $H_\alpha$ of $\K(\alpha)$-fixed vectors in $H$. Denote by $P_\alpha$ the operator of orthogonal projection to $H_\alpha$. Let $\frg\in \K(\alpha)\setminus \G/\K(\beta)$ be a double coset, and let $g\in \G$ be its representative. We define an operator $$ \ov\rho(\frg): H_\beta\to H_\alpha $$ by $$ \ov\rho(\frg)=P_\alpha \rho(g)\Bigr|_{H_\beta} $$ \begin{theorem} The operator $\ov\rho(\frg)$ does not depend on the choice of a representative $g\in \frg$. For any $\alpha$, $\beta$, $\gamma$, $$ \frg\in \K(\alpha)\setminus \G/\K(\beta),\quad \frh\in \K(\beta)\setminus \G/\K(\gamma) $$ we have $$ \ov \rho(\frg)\ov\rho(\frh)=\ov\rho(\frg\circ \frh) $$ \end{theorem} See a proof for $n=2$ in , the general case is completely similar (also this is a special case of , Theorem VIII.5.1) \begin{theorem} Let $\pi$ be a representation of the category $\cK$ in Hilbert spaces compatible with the involution and satisfying $\|\pi(\frg)\|\le 1$ for all $\frg$. Then $\pi$ is equivalent to some representation $\ov\rho$, where $\rho$ is a unitary representation of $\G$. \end{theorem} This is a special case of , Theorem VIII.1.10. \smallskip {\bf\punct Spherical functions.} In the above example we have $$ \frV_\alpha=\underbrace{(\cV,\xi)\otimes \dots (\cV,\xi)}_{ \text{$\alpha$ times}}\otimes \xi\otimes \xi\dots\simeq \cV^{\otimes \alpha} ,$$ in particular $$ \frV_0=\frv. $$ We wish to write an explicit formula for the spherical function $$ \Phi(g)=\la \nu(g) \frv,\frv\ra. $$ Choose an orthonormal basis in each space $V_{red}$, $V_{blue}$, $V_{green}$, etc. $$ e_{red}^{i}\in V_{red}, \quad e_{blue}^j\in V_{blue},\quad e^k_{green}\in V_{green},\dots $$ This determines the basis \begin{equation*} e_{red}^{i}\otimes e_{blue}^j\otimes e^k_{green}\otimes \dots \end{equation*} in $\cV$. Expand $\xi$ in this basis, \begin{equation} \xi=\sum x_{ijk\dots} e_{red}^{i}\otimes e_{blue}^j\otimes e^k_{green}\otimes \dots \end{equation} Consider the double coset $\frg$ containing $g$ and the corresponding colored pseudomanifold $\Sigma$. Assign to each $(n-1)$-face an element of the basis of the corresponding color (in arbitrary way). Fix such arrangement. Consider a chamber $\Delta$, on its faces we have certain basis vectors $e_{red}^{i}$, $e_{blue}^j$, $e^k_{green}$, ... . Then we assign the number $x(\Delta):=x_{ijk\dots}$ (see the last formula) to $\Delta$. \begin{proposition} $$ \Phi(g)= \sum\limits_{\begin{matrix}\text{\scriptsize arangements}\\ \text{\scriptsize of basis elements} \end{matrix}} \prod\limits_{\text{plus-chambers $\Delta$}} x(\Delta) \cdot \prod\limits_{\text{minus-chambers $\Gamma$}} \ov{x(\Gamma)} $$ \end{proposition} Proof coincides with proof of Proposition 4.2 in . \begin{thebibliography}{cc} \bibitem{BG} Bukhshtaber, V. M.; Gaifullin, A. A. {\it Representations of m-valued groups on triangulations of manifolds.} Russian Math. Surveys 61 (2006), no. 3, 560-562 \bibitem{Dix} Dixmier, J. {\it Les $C^*$-algebr\'es et leurs representations.} Gauthier Villars, 1964. \bibitem{Fer} Ferri, M. {\it Una rappresentazione delle n-varieta topologiche triangolabili mediante grafi {\rm (n+1)}-colorati}. Boll. Un. Mat. Ital. B (5) 13 (1976), no. 1, 250-260. \bibitem{FGG} Ferri, M.; Gagliardi, C.; Grasselli, L. {\it A graph-theoretical representation of PL-manifolds: a survey on crystallizations.} Aequationes Math. 31 (1986), no. 2-3, 121-141. \bibitem{Gai} Gaifullin, A. A. {\it The Manifold of Isospectral Symmetric Tridiagonal Matrices and Realization of Cycles by Aspherical Manifolds}, Proc. Steklov Inst. Math., 263 (2008), 38-56 \bibitem{Gai1} Gaifullin, A. {\it Universal realisators for homology classes.} Geom. Topol. 17 (2013), no. 3, 1745-1772. \bibitem{Gai3} Gaifullin, A. {\it Combinatorial realisation of cycles and small covers.} European Congress of Mathematics (Krak\'ow, 2--7 July, 2012), eds. R. Latala et al., European Mathematical Society, 2013, 315--330, \bibitem{GM} Goresky, M.; MacPherson, R., {\it Intersection homology theory}, Topology 19 (1980), no. 2, 135-162. \bibitem{Hat} Hatcher, A. {\it Algebraic topology.} Cambridge University Press, Cambridge, 2002. \bibitem{Ner-cat} Neretin, Yu. A. {\it Categories of symmetries and infinite-dimensional groups.} Oxford University Press, New York, 1996. \bibitem{Ner-comb} Neretin, Yu. A. {\it Infinite symmetric group and combinatorial descriptions of semigroups of double cosets}. arXiv:1106.1161 \bibitem{Ner-tri} Neretin, Yu. {\it Infinite tri-symmetric group, multiplication of double cosets, and checker topological field theories.} Int. Math. Res. Not. IMRN 2012 (2012), no. 3, 501-523. \bibitem{Ner-rem} Neretin, Yu.A. {\it A remark on representations of infinite symmetric groups.} J. Math. Sci., New York 190, No. 3, 464-467 (2013); translation from Zap. Nauchn. Semin. POMI 403, 103-109 (2012). \bibitem{Olsh1} Olshanski, G. I. {\it Unitary representations of $(G,K)$-pairs that are connected with the infinite symmetric group $S(\infty)$.} Leningrad Math. J. 1 (1990), no. 4, 983-1014. \bibitem{Olsh2} Olshanski, G. I. {\it On semigroups related to infinite-dimensional groups.} Topics in representation theory, 67-101, Adv. Soviet Math., 2, Amer. Math. Soc., Providence, RI, 1991. \bibitem{Pez} Pezzana, M. {\it Diagrammi di Heegaard e triangolazione contratta.} (Italian) Collection in memory of Enrico Bompiani. Boll. Un. Mat. Ital. (4) 12 (1975), no. 3, suppl., 98-105. \bibitem{ST} Seifert, H.; Threlfall, W. {\it Lehrbuch der Topologie.} Leipzig und Berlin: B. G. Teubner. VII, 353., 1934; English transl. Academic Press, 1980. \bibitem{vN} von Neumann, J. {\it On infinite direct products.} Compos. Math. 6, 1-77 (1938). Reprinted in von Neumann {\it Collected works, V.3}, Pergamon Press, 1963, 323-399. \end{thebibliography} \tt \noindent A.Gaifullin: Steklov Mathematical Institute, Moscow, Russia;\\ MechMath Dept., Moscow State University\\ Kharkevich Institute for Information Transmission Problems, Moscow, Russia\\ agaif@mi.ras.ru \noindent Yu.Neretin: Math. Dept., University of Vienna; \\ Institute for Theoretical and Experimental Physics; \\ MechMath Dept., Moscow State University \\ neretin(at)mccme.ru *yjdsq fh,fn 13? 319 <sandra.schoenach@wien.gv.at> 20-ref@ma35.wien.gv.at |
1501.04063 | Title: Do transitive preferences always result in indifferent divisions?
Abstract: The transitivity of preferences is one of the basic assumptions used in the
theory of games and decisions. It is often equated with rationality of choice
and is considered useful in building rankings. Intransitive preferences are
considered paradoxical and undesirable. This problem is discussed by many
social and natural sciences. The paper discusses a simple model of sequential
game in which two players in each iteration of the game choose one of the two
elements. They make their decisions in different contexts defined by the rules
of the game. It appears that the optimal strategy of one of the players can
only be intransitive! (the so-called \textsl{relevant intransitive
strategies}.) On the other hand, the optimal strategy for the second player can
be either transitive or intransitive. A quantum model of the game using pure
one-qubit strategies is considered. In this model, an increase in importance of
intransitive strategies is observed -- there is a certain course of the game
where intransitive strategies are the only optimal strategies for both players.
The study of decision-making models using quantum information theory tools may
shed some new light on the understanding of mechanisms that drive the formation
of types of preferences.
Body: \title{DO CLASSICAL (OR QUANTUM) TRANSITIVE PREFERENCES ALWAYS RESULT IN INDIFFERENT DIVISIONS?} \author{Marcin Makowski $^{1,}$*, Edward W. Piotrowski $^{1}$** and Jan S\l{}adkowski $^{2}$\\ \small 1. Institute of Mathematics, University of Bia\l ystok,\\\small Akademicka 2, PL-15424, Bia{\l}ystok, Poland \\\small $*$ makowski.m@gmail.com, $**$ qmgames@gmail.com \\ \small 2. Institute of Physics, University of Silesia,\\\small Uniwersytecka 4, 40-007 Katowice, Poland\\ \small jan.sladkowski@us.edu.pl} \maketitle \begin{abstract} The transitivity of preferences is one of the basic assumptions used in the theory of games and decisions. It is often equated with rationality of choice and is considered useful in building rankings. Intransitive preferences are considered paradoxical and undesirable. This problem is discussed by many social and natural sciences. The paper discusses a simple model of sequential game in which two players in each iteration of the game choose one of the two elements. They make their decisions in different contexts defined by the rules of the game. It appears that the optimal strategy of one of the players can only be intransitive! (the so-called \textsl{relevant intransitive strategies}.) On the other hand, the optimal strategy for the second player can be either transitive or intransitive. A quantum model of the game using pure one-qubit strategies is considered. In this model, an increase in importance of intransitive strategies is observed -- there is a certain course of the game where intransitive strategies are the only optimal strategies for both players. The study of decision-making models using quantum information theory tools may shed some new light on the understanding of mechanisms that drive the formation of types of preferences.\end{abstract} \section{Introduction} Games have always fascinated scholars, often contributing to the development of new theories . In fact, the attempts to construct a systematic theory of rational behaviour are focused on games as simple examples of human rationality. The attractiveness of such approach to the analysis the interactions between rational players (a problem reflected in many fields of science) reveals itself in its various applications . Game theory methods have been used in areas such as military science, biology, economics and other social sciences. Since the very beginning, the game theory has been closely connected with the information theory . Therefore, during the development of the theory of quantum information , a quantum game theory has emerged in a natural way. In its general form, pure strategies are identified with Hilbert space vectors (pure states) --- qubits or qubit systems. Mixed strategies are represented by convex combinations of pure states. Replacing classical probabilities used in the game theory with quantum probability amplitudes provides many interesting opportunities arising from superposition and entanglement. The idea of constructing quantum models of games is intensely developing nowdays. Numerous examples show the rapid rate with which the situation of players changes after obtaining access to quantum technology and how the fundamental limitations of classical models can be overcomed thanks to the properties of quantum processes . This paper considers a simple model of a repeated game in which, at each stage, both players divide between themselves a set consisting of three goods. With reference to the earlier results influenced by the remarks of Hugo Steinhaus , the players are referred to as 'cats' and goods are referred to as 'foods'. The main goal of our analysis is to examine how rules of the game affect the type of optimal strategies for the players (whether they are transitive or intransitive). The differences arising from the adoption of a different way of sourcing strategies are also examined. One of them is a mixed strategy based on one bit (classical variant) and the other is one-qubit pure state (quantum variant). Before we proceed to the formal description of the game, let us highlight the main ideas of transitivity --- intransitivity. This is done in the next subsection. The second subsection of this introduction is devoted to review of recent results that have led us to formulate the main results. \subsection{Intransitivity} Any relation $\succ$ between the elements of a certain set is called transitive if \begin{equation} A\succ B \wedge B\succ C \Rightarrow A\succ C. \end{equation} \noindent is fulfilled for any three elements $A$, $B$, $C$. If this condition is not fulfilled then the relation will be called intransitive (not transitive). The problem of transitivity (intransitivity) stems from various fields of research. There is an opinion that people who make decisions relying on rational reasoning, should make decisions in determined and linear order . Transitivity of preferences indicates the way of choosing according to the ``logical order''. There is also a hypothesis that many animals (also people) follow transitive inference rules (choosing \textit{A} over \textit{C} on the basis of knowing that \textit{A} is better than \textit{B} and \textit{B} is better than \textit{C}). This type of reasoning has been confirmed in several animal species . One of the main arguments against intransitiveness is the so-called ``money pump'' . On the other hand, some modification of utility theory and decision theory which dispense of the transitivity assumption have been considered . The beginnings of research on intransitive orders probably dates back to 1785 when Jean Condorcet published his work: \textit{Essay on the Application of Analysis to the Probability of Majority Decisions} (1785), in which he analyzed the paradox of voting. He concluded that collective preferences can be intransitive even if the preferences of individual voters are not. Analysis of this paradox led Kenneth Arrow (Nobel Prize winner in economy) to prove that an election procedure which would perfectly agree with basic postulates of democracy does not exist . In psychology, which also attempts to explain, inter alia, the decision making process, the special interest is focused on a broadly understood relation of superiority or domination (preference) . Does the fact that A dominates over B and B dominates over C, imply that A dominates over C? It turns out that the answer is not obvious and depends on a particular situation. The discussion on the concept of dominance has also continued using the tools provided by exact sciences with probabilistic models as good examples, e.g.~Efron's Dice . The game proposed by Walter Penney is another example of intransitivity in probabilistic models. Intransitivity may explain the processes occurring in Nature. Rivalry between species may be intransitive. For example, in the case of fungi, Phallus impudicus replaced Megacollybia platyphylla, M.~platyphylla replaced Psathyrella hydrophilum, but P.~hydrophilum replaced P.~impudicus . Similarly, we can explain the stability of the population of lizards or experiments with bees which make intransitive choices between flowers . Intransitivity models appear also in many seemingly distant sciences including philosophy , operations research , thermodynamics , quantum theory , logic . Research on the rational decision-making process is complex and consists of many various concepts. There is a large literature with the discussions on rational choice . This is a difficult problem because it turns out that the assumptions made by theoretician (such as completeness, transitivity, independence of irrelevant alternatives or others) are often broken in the course of experiments with humans. Thus, people have constructed theories of choice without transitivity or other assumptions. In this paper we focus on one of the most important issues in the theory of decision --- the transitivity/intransitivity of the choice. Reflections on this occurs in the context of studies of various types of preferences, state-dependent preferences , context-dependent preferences . Intransitive issue is still not well understood. It is worth to study this using tools of quantum game theory. This new look can lead to many interesting conclusions that can be used in research on decision-making process simulation and other biological mechanisms. \subsection{Earlier results. ``I cut, you choose'' game.} In his diary, Hugo Steinhaus mentions Pitts experiments with cats. It turns out that that the cat, facing the choice between fish, meat and milk prefers fish to meat, meat to milk, and milk to fish! Steinhaus thought that the cat provided itself with a balanced diet, thanks to the above-mentioned food preferences. This is one of the key factors needed to maintain good health. Hence, in our model of the game, the players are cats choosing among foods. Obviously, this is only an illustration of the problem which can be interpreted in different ways (e.g. in relation to other goods, or as an electoral issue). In the paper was considered a classic model of a game in which the player (cat) are offered three types of foods, every time in pairs of two types. Optimal strategy was defined as one that leads to a balanced diet (equal distribution of the frequency of the occurrence of a particular food in cat's diet). The offering player (Nature) was not interested in the result of the game. The quantitative analysis of various types (intransitiv or transitive) optimal strategies indicated the advantage of the transitive strategies. Intransitive strategies represent a significant part of all optimal strategies, but in the situation that favors optimal strategies we can always find the strategy that determines the transitive order (with identical result, under the same conditions). This situation changed in the quantum variant of the game , which reveals the existence of the so-called \textit{relevant intransitive strategies}, defined as follows : \begin{df} The intransitive strategy will be called the relevant strategy, if there is no transitive strategy of the same consequences with the same assumption. \end{df} In the paper was considered electoral interpretation of the game (see also ). The decrease of importance of intransitive orders which accompanies the growth of support for one of the candidates turned out to be an interesting property of the quantum game model. The use of an entangled state to the model construction causes an increase in the importance of transitive strategies. It turns out that intransitive relevant strategies are not only characteristic of quantum models. In the paper offering player (Nature) has been replaced by a rational player. In this game both players (Cat 1 and Cat 2) divide the set of three foods according to the ''I cut, you choose'' method (see Fig.). The first player, chooses and rejects one of the foods. Then the second player, selects and consumes one of the remaining two foods. The first player eats the food that is left. The optimum consists (as in previous models) in not distinguishing any of the three foods. Each of the foods is equally important to each of the players. In this variant of the game, only the second player (Cat 2) makes choices between the pairs of foods offered to him. Therefore, transitive and intransitive strategies can be analysed only in case of this player. It appears that his optimal strategies are only intransitive strategies (\textsl{relevant intransitive strategies}.). In this article, a modification of the above game is analysed. It consists in letting the player choose the food selected at the first stage of the game, as well as the food rejected by the other player. This modification allows both players to determine their preferences with regard to the pairs of foods. It will enable exploring the types of strategies and their availability for individual players. \section{Classical model} The two players(\textit{Cat 1} i \textit{Cat 2}) are offered three foods: food number 0, food number 1, food number 2, always in pairs of two, in accordance with the following procedure. The first player (\textit{Cat 1}) selects and keeps one of the three foods. The second player (\textit{Cat 2}) chooses between the two remaining foods. The food kept in the first move and the food rejected by Cat 2 now form a pair from which a selection is made by \textit{Cat 1} ($2^{nd}$ move) . Diagram of the game is shown in Figure . The optimal behaviour of the players is not to favour any of the three foods (each being equally attractive and important). Both participants try to achieve a balanced diet. This assumption allows to extract meaning of intransitive orders, because all the elements that we sort of are equally important. \subsection{Mathematical description} The first move of \textsl{Cat 1} is limited to select one of the three food. This movement can be described by the point $(P_0,P_1,P_2)$ of a three dimensional simplex, where $P_j$ denotes the frequency of the choice of food numbered by \textit{j}. Let us denote by $P(C_{k} | B_{j})$ the probability of choosing (by the Cat 2) food number \textit{k} when the offered pair of dishes does not contain food number \textit{j}. Analogously, number $Q(C_{k} | B_{j})$ denotes the probability of choosing by the \textit{Cat 1} food number \textit{k} when the offered pair of dishes does not contain food number \textit{j}. Six conditional probabilities $P(C_{k} | B_{j})$ (or $Q(C_{k} | B_{j})$) determine the behaviour of the players with respect to the proposed them in pairs food portions. Let $\lambda_k$ and $\omega_k$ denot the frequencies of appearance of the particular foods in the \textit{Cat 1} and the \textsl{Cat 2} diet, respectively. For \textit{Cat 2} we obtain: \begin{eqnarray} \omega_0&=&P(C_{0} | B_{1})P_1+P(C_{0} | B_{2})P_2,\nonumber\\ \omega_1&=&P(C_{1} | B_{0})P_0+P(C_{1} | B_{2})P_2,\\ \omega_2&=&P(C_{2} | B_{0})P_0+P(C_{2} | B_{1})P_1.\nonumber \end{eqnarray} From the perspective of the \textit{Cat 1}: \begin{eqnarray} \lambda_0&=&Q(C_{0} | B_{1})[P(C_{1} | B_{0})P_0+P(C_{1} | B_{2})P_2]+Q(C_{0} | B_{2})[P(C_{2} | B_{0})P_0+P(C_{2} | B_{1})P_1],\nonumber\\ \lambda_1&=&Q(C_{1} | B_{2})[P(C_{2} | B_{0})P_0+P(C_{2} | B_{1})P_1]+Q(C_{1} | B_{0})[P(C_{0} | B_{1})P_1+P(C_{0} | B_{2})P_2],\\ \lambda_2&=&Q(C_{2} | B_{1})[P(C_{1} | B_{0})P_0+P(C_{1} | B_{2})P_2]+Q(C_{2} | B_{0})[P(C_{0} | B_{1})P_1+P(C_{0} | B_{2})P_2].\nonumber \end{eqnarray} Let us introduce the parameterization of the conditional probabilities. Let us assume: \begin{eqnarray} P(C_2|B_0)=\frac{1+l_0}{2},&& P(C_1|B_0)=\frac{1-l_0}{2},\nonumber \\ P(C_0|B_1)=\frac{1+l_1}{2},&& P(C_2|B_1)=\frac{1-l_1}{2}, \\ P(C_1|B_2)=\frac{1+l_2}{2},&& P(C_0|B_2)=\frac{1-l_2}{2}\,.\nonumber \end{eqnarray} and \begin{eqnarray} Q(C_2|B_0)=\frac{1-L_0}{2},&& Q(C_1|B_0)=\frac{1+L_0}{2},\nonumber \\ Q(C_0|B_1)=\frac{1-L_1}{2},&& Q(C_2|B_1)=\frac{1+L_1}{2}, \\ Q(C_1|B_2)=\frac{1-L_2}{2},&& Q(C_0|B_2)=\frac{1+L_2}{2}\,.\nonumber \end{eqnarray} \subsection{Optimal strategies} The optimal strategy for the two players is to get the equal distribution of frequency of the occurrence of a particular food in diet. It is required that the following conditions are held: \begin{eqnarray} \lambda_0=\lambda_1=\lambda_2&=&\frac{1}{3}\,,\\ \omega_0=\omega_1=\omega_2&=&\frac{1}{3}\,. \end{eqnarray} The () condition takes the form: \begin{eqnarray} l_1 P_1-l_2 P_2=\frac{2}{3}-(P_1+P_2),\nonumber \\ l_2 P_2-l_0 P_0=\frac{2}{3}-(P_0+P_2), \\ l_1 P_1-l_3 P_3=\frac{2}{3}-(P_0+P_3)\, ,\nonumber \end{eqnarray} and its solution: \begin{eqnarray} P_0=\frac{-(-1 - l_1 + l_2 - 3\,l_1 l_2)}{3 (1 + l_0 l_1 + l_0 l_2 + l_1 l_2)},\nonumber \\ P_1=\frac{-(-1 + l_0 - l_2 - 3\,l_0 l_2)}{3 (1 + l_0 l_1 + l_0 l_2 + l_1 l_2)}, \\ P_2= \frac{-(-1 - l_0 + l_1 - 3\,l_0 l_1)}{3 (1 + l_0 l_1 + l_0 l_2 + l_1 l_2)}\, ,\nonumber \end{eqnarray} defines a mapping of the three-dimensional cube $[-1,1]^3$ in the space of parameters $l_j$ (that define conditional probabilities $P(C_{k} | B_{j})$) into a triangle--two-dimensional simplex. The barycentric coordinates of a point of this triangle are interpreted as the probabilities $(P_0,P_1,P_2)$. Note that \textit{Cat 2} is able to achieve the optimum effect () if for every $j$ ($j=0,1,2$) the condition $P_j\leq\tfrac{2}{3}$ is satisfied. Indeed, if $P_0>\tfrac{2}{3}$ then $P_1+P_2<\tfrac{1}{3}$ and $\omega_0 <\tfrac{1}{3}$ (similarly in other cases). Let us now consider the situation of (\textit{Cat 1}). This player moves first in each iteration of the game. The () condition can be written as follows: \begin{eqnarray} (1 - L_1) (1 - l_0) P_0 + (1 + L_2) (1 + l_0) P_0 + (1 + L_2) (1 - l_1) P_1 + (1 - L_1) (1 + l_2) P_2 = 4/3,\nonumber \\ (1 - L_2) (1 + l_0) P_0 + (1 + L_0) (1 + l_1) P_1 + (1 - L_2) (1 - l_1) P_1 + (1 + L_0) (1 - l_2) P_2 = 4/3, \\ (1 + L_1) (1 - l_0) P_0 + (1 - L_0) (1 + l_1) P_1 + (1 - L_0) (1 - l_2) P_2 + (1 + L_1) (1 + l_2) P_2 = 4/3\, ,\nonumber \end{eqnarray} Let us suppose there are exist {3}$ for any $j$.} parameters $l_i$ such that condition () is satisfied -- Cat 2 reaches the diet completeness. Substituting formula () on the probability $P_j$ into formula () immediately gives \begin{eqnarray} L_1 = L_2,\nonumber\\ L_0 = L_2, \\\nonumber L_0 = L_1. \end{eqnarray} These conditions describe a set of \textit{Cat's 1} optimal strategy. This player must use strategy characterized by three (independent) equal conditional probabilities. In the next section we analyze the optimal strategies for both players in terms of their division into two types - intransitive and transitive. \section{Optimal intransitive and transitive strategies} We say that a player prefers food no. 1 to food no. 0 ($1\succ 0$) when he/she is willing to choose food no. 1 more often than food no. 0 from the offered pair $(0,1)$ ($P(C_1|B_2)>P(C_0|B_2)$). The situation corresponds to an intransitive choice if one of the following two conditions is satisfied: \vspace{0.1cm} \begin{enumerate} \item $P(C_0|B_2)<P(C_1|B_2)$, $P(C_1|B_0)<P(C_2|B_0)$, $P(C_2|B_1)<P(C_0|B_1)$\,, \vspace{0.3cm} \item $P(C_0|B_2)>P(C_1|B_2)$, $P(C_1|B_0)>P(C_2|B_0)$, $P(C_2|B_1)>P(C_0|B_1)$\,. \end{enumerate} These two conditions can be written (by using normalization probability measure to 1) in the following form: \begin{eqnarray} P(C_0|B_2)<\frac{1}{2}, P(C_1|B_0)<\frac{1}{2}, P(C_2|B_1)<\frac{1}{2}\,,\\ P(C_0|B_2)>\frac{1}{2}, P(C_1|B_0)>\frac{1}{2}, P(C_2|B_1)>\frac{1}{2}\,. \end{eqnarray} Let's see how different types of strategies of individual players are achievable. The conditions under which \textit{Cat 2} makes his decision are similar to those considered in the work (then it was a game with Nature). This player can select its optimal strategy if only $P_j\leq \tfrac{2}{3}$, for $j=0,1,2$. It turns out that under this assumption, this strategy can always be transitive, but not always intransitive. This can be illustrated graphically. Figure presents the areas of frequency $(P_0,P_1,P_2)$ for which \textit{Cat 2} optimal strategies exist. It is the range (for 10,000 randomly selected points) of mapping (defined by equations () ) of the three-dimensional cube of parameters $t_j$ into a triangle $(P_0,P_1,P_2)$. Probabilities $ (P_0, P_1, P_2) $, for which \textit{Cat 2} can identify the optimal intransitive strategy, form a six-armed star composed of two triangles (any of them corresponding to one of two possible intransitive orders). Both the optimal strategies of any type and optimal transitive strategies correspond to regular hexagon specified by conditions $P_j\leq \tfrac{2}{3}$, $j=0,1,2$. The other player (\textit{Cat 1}) situation is completely different. \textit{Cat 1} selects at the final stage of each iteration of the game. This movement reveals his preference for one food over another from pair. As pointed out above, \textit{Cat 1} optimal strategies (assuming that the second player reaches the optimal strategy) satisfy the condition (), therefore: \begin{eqnarray} Q(C_0|B_2)=Q(C_1|B_0)=Q(C_2|B_1). \end{eqnarray} This means that either () or () is satisfied. It means that \textit{Cat 1} has to make intransitive choices in order to achieve the optimal result! \section{Quantum cats} \subsection{One-qubit pure strategies} In this section we use different method of obtaining conditional probabilities (provided by the quantum games theory) which describe the preferences of decision makers,\textit{Cat 1} and \textit{Cat 2}, over pairs of food. This method was first used in the work and is based on the concept of the so-called mutually unbiased: \begin{df}Two orthonormal bases $\mathcal{A}\equiv \{ \,|\psi_0 \rangle\negthinspace\,\ldots\,|\psi_{N-1} \rangle\negthinspace\,\}$ and~$\mathcal{B}\equiv \{ \,|\varphi_0 \rangle\negthinspace\,\ldots\,|\varphi_{N-1} \rangle\negthinspace\,\}$ in Hilbert space $ \mathbb{C}^N$ are mutually unbiased if \begin{displaymath} |\langle \psi_i | \varphi_j \rangle|^2 = \frac{1}{N}, \end{displaymath} for any $0\le i,j \le N-1$. \end{df} For two-dimensional Hilbert space, three mutually unbiased bases is given as follows \begin{align} &\left\{ | 0 \rangle,| 1 \rangle \right\},&\nonumber \\ &\left\{ \frac{| 0 \rangle+| 1 \rangle}{\sqrt{2}},\frac{| 0 \rangle-| 1 \rangle}{\sqrt{2}} \right\},& \\ &\left\{ \frac{| 0 \rangle+i | 1 \rangle}{\sqrt{2}},\frac{| 0 \rangle-i| 1 \rangle}{\sqrt{2}} \right\}.& \nonumber \end{align} This set is the most important for our considerations. It is worth to mention here that mutually unbiased bases led Wiesner to begin research into quantum cryptography, before asymmetric key cryptography was invented! These bases play also an important role in universality of quantum market games . Let us turn to the construction of conditional probabilities that define the strategy of the players. Let us denote three different mutually unbiased bases of two-dimensional Hilbert space as \begin{displaymath} \{| 1 \rangle\negthinspace_{\,0}, | 2 \rangle\negthinspace_{\,0}\}, \{| 0 \rangle\negthinspace_{\,1}, | 2 \rangle\negthinspace_{\,1}\}, \{| 0 \rangle\negthinspace_{\,2}, | 1 \rangle\negthinspace_{\,2}\}=\{ (1,0)^{T},(0,1)^{T}\} \end{displaymath} Strategy of choosing the food number \textit{k}, when the offered food pair not contain the food of number \textsl{l} is denoted by $| k \rangle\negthinspace_{\,l}$ ($k$,\,$l=0,1,2$ i~$k\ne l$). A family $\{|z\rangle\negthinspace\}$ ($z \in \overline{\mathbb{C}}$) of convex vectors: \begin{displaymath} | z \rangle\negthinspace:=| 0 \rangle\negthinspace_{\,2}+z |1\rangle\negthinspace_{\,2}=| 0 \rangle\negthinspace_{\,1}+\frac{1-z}{1+z} |2\rangle\negthinspace_{\,1}=| 1 \rangle\negthinspace_{\,0}+\frac{1+iz}{1-iz} |2\rangle\negthinspace_{\,0}, \end{displaymath} defined by the parameters of the heterogeneous coordinates of the projective space $\mathbb{C}P^{1}$, represents all strategies spanned by the base vectors. The coordinates of the same strategy $|z\rangle\negthinspace$ read (measured) in three different bases () define quantum cats preferences toward a food pair represented by the base vectors. Squares of their moduli, after normalization, measure the conditional probability of quantum cat�s making decision in choosing a particular product, when the choice is related to the suggested food pair. Diagram of the quantum variant of game is shown in Figure . It is similar to the previous one , since the difference lies only in the process of implementing the game. The player (\textit{Cat 1} or \textit{Cat 2} in our game) makes a decision to choose the right food from pair with the following probabilities: \begin{align} P(C_0|B_2) & =\frac{1}{1+|z|^{2}},& P(C_1|B_2) &=\frac{|z|^{2}}{1+|z|^{2}}, \nonumber \\ P(C_0|B_1)&=\frac{1}{1+|\frac{1-z}{1+z}|^2}, & P(C_2|B_1) &= \frac{|\frac{1-z}{1+z}|^2}{1+|\frac{1-z}{1+z}|^2}, \\ P(C_1|B_0)&=\frac{1}{1+|\frac{1+iz}{1-iz}|^2},& P(C_2|B_0)&= \frac{|\frac{1+iz}{1-iz}|^2}{1+|\frac{1+iz}{1-iz}|^2}\,.\nonumber \end{align} It is convenient to parameterized $|z\rangle\negthinspace$ by the sphere $S_2\backsimeq \overline{\mathbb{C}}$ by using stereographic projection which establishes bijection between elements of $\overline{\mathbb{C}}$ and the points of $S_2$. The conditional probabilities can now be written in the following form: \begin{align} P(C_0|B_2)&=\frac{1-x_3}{2},& P(C_1|B_2)&=\frac{1+x_3}{2},&\nonumber \\ P(C_0|B_1)&=\frac{1+x_1}{2},& P(C_2|B_1)&=\frac{1-x_1}{2},& \\ P(C_1|B_0)&=\frac{1+x_2}{2},& P(C_2|B_0)&=\frac{1-x_2}{2}.&\nonumber \end{align} Note that the probability () are parameterized similarly to the classic model. But in this case we have a sphere points, so the condition $x_1 ^ 2 + x_2^ 2 + x_3^ 2 = 1 $ must be fulfilled. A careful reader certainly noticed that we abstained from introducing to much of the game-theoretical terminology. The problem can be easily reformulated in the language of quantum game theory . Another interesting approach would be to adopt the framework developed by A.~Lambert-Mogiliansky, I.~Martinez-Martinez , c.f.~also . This framework allows one to incorporate type indeterminacy of agents . In our model the agents are of definite type. \subsection{Availability of different types of one-qubit strategies} Availability of various types of \textit{Cat 2}'s one-qubit optimal strategies is illustrated in Figure (see ). Here we see a fundamental contrast between quantum and classical model. First of all, the area corresponding to the optimal transitive strategies does not cover the entire area of the optimal strategy of any type (transitiv or intransitiv). Transitive optimal strategies do not appear within the boundaries of hexagon-like figure in the central part of the triangle. This means that for some frequencies $(P_0,P_1,P_2)$ \textit{Cat 2} is able to achieve optimal results only thanks to the intransitive strategy! and ) is similar to that in work . Readers interested in this, please refer to this paper.} Let's analyze the \textit{Cat 1} situation in the case of one-qubit pure strategy, at the final stage of each iteration of the game. The probabilities $Q (C_k | B_j) $ are in the form (). To distinguish it from the strategy of the other player instead of $x_i$, we will use $ X_i $. Assuming the \textit{Cat 2} achieves the optimal strategy, the optimality conditions () for \textit{Cat 1} will simplify to: \begin{eqnarray} X_1-X_3=0,\nonumber\\ X_2 +X_3=0, \\\nonumber X_1+X_2=0. \end{eqnarray} Since $ X_1^ 2 + X_2^2 + X_3^2 = 1 $ we obtain two optimal strategies that are characterized by two points $(X_1, X_2, X_3)$ of two-dimensional sphere: \begin{eqnarray} X_1= -\frac{1}{\sqrt{3}}, X_2=\frac{1}{\sqrt{3}}, X_3=-\frac{1}{\sqrt{3}},\nonumber\\\nonumber X_1= \frac{1}{\sqrt{3}}, X_2=-\frac{1}{\sqrt{3}}, X_3=\frac{1}{\sqrt{3}}. \end{eqnarray} Both of these strategies are intransitive. It is worth noting that in the language of quantum game theory one-qubit strategies are pure strategies, have an equal informative values (the zero entropy). Hence, treating them in an equal way is natural and does not raise any controversy similar of those of the constant measure (the Laplace's principle of insufficient reason) in the classical model. In this case, we used a mixed strategies which may provide different pieces of information. Pure strategies that we can identify with eight choice functions: \begin{displaymath} f_k:\{(0,1), (0,2),(1,2)\}\to \{0,1,2\},\qquad k=0,\ldots,7, \end{displaymath} do not lead us to identify the optimal strategy. \section{Discussion and conclusions} The variant of the game considered in the paper allows to explore the availability of different types of optimal strategies of both players (as opposed to earlier models which related to the situation of one player only). They make their decisions in a variety of contexts (depending on the decision made by the other player), which affects their possibility to use optimal intransitive and transitive strategies. The most interesting variant of the game has been analysed, i.e. such variant where the circumstances provide both players with the opportunity to achieve their optimal strategies. It appears that the only optimal strategies of \textit{Cat 1} are intransitive strategies. This is indeed an interesting result. This player performs two moves in the game. In the first one, he chooses (with frequency $P_j$) one of the foods. The food is included in the pair from which he makes the final selection. Therefore, he has a partial influence on the composition of this pair does not have this possibility .} It might seem that such a possibility should help him achieve an optimal strategy in the sense that he has more freedom in choosing the type of strategy. However, it is quite the opposite; he can only use an intransitive strategy. The other player (\textit{Cat 2}) in the classical variant of the game can always use a transitive strategy (in certain specific conditions also intransitive strategy, see Fig. ). Replacing classical strategies of the players with one-qubit quantum strategies will increase the importance of intransitive strategies. At certain $(P_0,P_1,P_2)$ frequencies, in order to achieve an optimal effect, both players must apply intransitive strategies! Moreover, the one-qubit strategies are pure strategies -- have an equal mathematical and information status (in contrast to the mixed strategies in classical model). The model discussed in the paper, despite its mathematical simplicity, clearly illustrates how the type of preferences depends on the context of the decision to be made and on the behaviour of other players. The rejection of intransitive strategies as undesirable strategies is an unnecessary limitation. A distinction between intransitive and transitive strategies does not necessarily mean a distinction between irrational and rational strategies. However, it may well describe the decision-making process from an entirely different point of view -- the context of choice (e.g. whether and how the freedom of choice with regard to the decisions we take is affected by the adopted rules or the behaviour of other players). It is worth mentioning that at present we can observe rapidly growing literature where ideas of Quantum Mechanics are proposed to explain a problems of behavioral and social science . The formalism of quantum mechanics can be useful in providing explanation to violations of transitivity in decision-making process (see ). We are still not fully aware of the implications of quantum theory in computer science models, machine learning and especially in decision-making. The work that has been done in this field indicates significant progress in the development this part of our knowledge. \section*{Acknowledgments} This work was supported by the Polish National Science Centre under the project number \textbf{DEC-2011/01/B/ST6/07197}. \begin{thebibliography}{999} \bibitem{ecbi} von Neumann, J.; Morgenstern, O. {\em Theory of Games and Economic Behavior}. Princeton University Press, Princeton, US, 1944. \bibitem{stt} Straffin, P. D. {\em Game Theory and Strategy}. The Mathematical Association of America, Washington DC, US, 1992. \bibitem{inf} Rasmusen, E. {\em Games and Information: An Introduction to Game Theory}. Wiley-Blackwell, Malden, US, 2007. \bibitem{rq} Nielsen, M. A,; Chuang, I. L. {\em Quantum Computation and Quantum information}. Cambridge University Press, New York, US, 2000. \bibitem{re4} Eisert, J.; Wilkens, M.; Lewenstein, M. Quantum games and quantum strategies. {\em Phys. Rev. Lett.} {\bf 1999}, {\em 83}, 3077-3080. \bibitem{rd5} Meyer, D. Quantum strategies. {\em Phys. Rev. Lett.} {\bf 1999}, {\em 8}, 1052-1055. \bibitem{rA6} Flitney, A. P.; Abbott, D. An introduction to quantum game theory. {\em Fluct. and Noise Lett.} {\bf 2002}, {\em 2} R175-R187. \bibitem{rP7} Piotrowski, E. W.; S\l{}adkowski, J. An invitation to quantum game theory. {\em Inter. J. of Theor. Phys.} {\bf 2003}, {\em 42}, 1089-1099. \bibitem{r8} Piotrowski, E. W.; S\l{}adkowski, J. The next stage: quantum game theory, In {\em Mathematical Physics Research at the Cutting Edge}, C. V. Benton, C. V. Eds.; Nova Science Publishers: New York, US, 2004; pp. 247-268. \bibitem{r1} Piotrowski, E. W.; Makowski, M. Cat's dilemma - transitivity vs. intransitivity. {\em Fluct. and Noise Lett.} {\bf 2005}, {\em 5}, L85-L96. \bibitem{r2} Makowski, M.; Piotrowski, E. W. Quantum cat's dilemma: an example of intransitivity in a quantum game. {\em Phys. Lett. A} {\bf 2006}, {\em 355}, 250-254. \bibitem{r2n} Makowski, M. Transitivity vs. intransitivity in decision making process - an example in quantum game theory {\em Phys. Lett. A} {\bf 2009}, {\em 373}, 2125-2130. \bibitem{En} Makowski, M.; Piotrowski, E. W. Transitivity of an entangled choice. {\em J. Phys. A Math. Theor.} {\bf 2011}, {\em 44}, {\em 075301}, 1-12. \bibitem{El} Makowski, M.; Piotrowski, E. W. Decisions in elections - transitive or intransitive quantum preferences. {\em J. Phys. A Math. Theor.} {\bf 2011}, {\em 44, 215303}, 1-10. \bibitem{div} Makowski, M.; Piotrowski, E. W. When I cut, you choose method implies intransitivity. {\em Physica A} {\bf 2014}, {\em 415}, 189-193. \bibitem{ST} Steinhaus, H. {\em Memoires and Notes} (in Polish), Aneks, London, UK, 1992. \bibitem{trra} Anand, P.; Pattanaik, P. K.; Puppe, C. Introduction, in: {\em Handbook of Rational and Social Choice}; Anand, P.; Pattanaik, P. K. and Puppe, C. Eds.; Oxford University Press, US, 2009; pp.1--20. \bibitem{wr} Von Wright, G. H. The Logic of Preference: an Essay, Edinburgh University Press, Edinburgh, UK, 1963. \bibitem{r33} Grosenick, L.; Clement, T. S.; Fernald, R. D. Fish can infer social rank by observation alone. {\em Nature}, {\bf 2007} {\em 445}, 429-432. \bibitem{tull} Tullock, G. The irrationality of intransitivity. {\em Oxford Economic Papers} {\bf 1964}, {\em 16}, 401--406. \bibitem{r23} Fishburn, P. C. Nontransitive Preferences in Decision Theory. {\em J. of Risk and Uncertainty} {\bf 1991} {\em 4(2)}, 113-134. \bibitem{tw} Fishburn, P. C. Nontransitive measurable utility. {\em J. of Math. Psychology} {\bf 1982}, {\em 26}, 31-67. \bibitem{r24} Arrow, K. J. {\em Social Choice and Individual Values}, Yale Univ. Press, New York, US, 1951. \bibitem{lek} Poddiakov, A. N. Intransitive Character of Superiority Relations and Decision-making. {\em Psychology. The J. of the Higher School of Econ.} {\bf 2006}, {\em Vol. 3 N 3}, 88-111. \bibitem{T} Tversky, A. Intransitivity of preferences. {\em Psychological Rev.} {\bf 1969}, {\em 76(1)}, 31-48. \bibitem {Gar} Gardner, M. Mathematical Games: The Paradox of the Nontransitive Dice and the Elusive Principle of Indifference. {\em Sci. Amer.} {\bf Dec. 1970}, {\em 223}, 110-114. \bibitem{r35} Penney, W. Problem 95: penney-ante. {\em J. of Rec. Math.} {\bf Oct. 1969}, 241-258. \bibitem{r15} Boddy, L. Interspecific combative interactions between wood-decaying basidiomycetes. {\em FEMS Microbiol Ecol.} {\bf 2000}, {\em 31}, 185-194. \bibitem{jasz} Sinervo, B.; Lively, C. M. The rock-paper-scissors game and the evolution of alternative male strategies. {\em Nature} {\bf 1996}, {\em 380}, 240-243. \bibitem{r16} Shafir, S. Intransitivity of preferences in honey bees: support for'comparative'evaluation of foraging options. {\em Animal Behaviour} {\bf 1994}, {\em 48}, 55-67. \bibitem{ope} Pahikkala, T.; Waegeman, W.; Tsivtsivadze, E.; Salakoski, T.; De Baets, B. Learning intransitive reciprocal relations with kernel methods. {\em European J. Oper. Res.} {\bf 2010}, {\em 206(3)}, 676-685. \bibitem{Ki} Klimenko, A. Y. Abstract competition and competitive thermodynamics. {\em Phil. Trans. R. Soc. A} {\bf 2013}, {\em 371(1982)}, 1-16. \bibitem{H} Halpern, J. Y. Intransitivity and vagueness. {\em Rev. Symbol. Log.} {\bf 2008}, {\em 1(04)}, 530-547. \bibitem{Trv} Tversky, A.; Kahneman, D. {\em Choices, Values, and Frames}. Cambridge Univ. Press, New York, US, 2000. \bibitem{St} Karni, E. {\em Decision Making Under Uncertainty: The Case of State-Dependent Preference}. Harvard Univ. Press, Cambridge, UK, 1985. \bibitem{trS} Tversky, A.; Simonson, I. Context-dependent preferences. {\em Management Science} {\bf 1993}, {\em 39}, 1179-1189. \bibitem{De} Aerts, D.; Sozzo, S.; Veloz, T. Quantum Structure in Cognition and the Foundations of Human Reasoning. {\bf 2014}, arXiv:1412.8704v1 [cs.AI], http://arxiv.org/pdf/1412.8704.pdf. \bibitem{kier} Khrennikova, P,; Haven, E.; Khrennikov, A. An application of the theory of open quantum systems to model the dynamics of party governance in the US Political System. {\em Int. J. of Theor. Phys.} {\bf 2013}, {\em 53 (4)}, 1346-1360. \bibitem{Brams} Brams, S. J.; Taylor, A. D. {\em Fair Division - From Cake-cutting to Dispute Resolution}. Cambridge Univ. Press, New York, US, 1996. \bibitem{ST1} Steinhaus, H. {\em Mathematical snapshots}. Oxford Univ. Press, New York, US, 1996. \bibitem{r27} Wiesner, S. Conjugate coding. {\em Sigact News} {\bf 1983}, {\em 15}, 78-88. \bibitem{r11} Piotrowski, E. W.; S\l{}adkowski, J. Quantum Market Games. {\em Physica A} {\bf 2002}, {\em 312}, 208-216. \bibitem{r12} Piotrowski, E. W.; S\l{}adkowski, J. Quantum-like Approach to Financial Risk: Quantum Anthropic Principle, {\em Acta Physica Polonica B} {\bf 2001}, {\em 32}, 3873-3879. \bibitem {28} Paku\l{}a, I., Piotrowski, E. W.; S\l{}adkowski, J. Universality of measurements on quantum markets. {\em Physica A} {\bf 2007}, {\em 385}, 397-405. \bibitem{29} Lambert-Mogiliansky, A.; Martinez-Martinez, I. Basic Framework for Games with Quantum-like Players. {\em PSE Working Papers} {\bf 2014}, {\em 2014-42}, 1-21. \bibitem{30} Lambert-Mogiliansky, A.; Zamir, S.; Zwirn, H. Type indeterminacy: A model of the KT(Kahneman-Tversky)-man. {\em J. of Math. Psych.} {\bf 2009}, {\em 53 (5)}, 349-361. \bibitem{31} Kvam, P. D.; Busemeyer, J. R.; Lambert-Mogiliansky, A. An Empirical Test of Type-Indeterminacy in the Prisoner's Dilemma. In {\em Quantum Interaction}; Atmanspacher, H., Haven, E., Kitto, K., Raine, D., Eds; Springer Berlin Heidelberg; 2014; pp 213-224. \bibitem{32} Lambert-Mogiliansky, A.; Busemeyer, J. R. Quantum type indeterminacy in dynamic decision-making: Self-control through identity management. {\em Games} {\bf 2012}, {\em 3(2)}, 97-118. \bibitem{33} Brandenburger, A., La Mura, P. Team Decision Problems with Classical and Quantum Signals. arXiv:1107.0237v3 [quant-ph] {\bf 2015}, 1-18. \bibitem{34} Busemeyer, J. R.; Wang, Z.; Townsend, J. T. Quantum dynamics of human decision making. {\em J. of Math. Psych.} {\bf 2006}, {\em 50}, 220-241. \bibitem{35} Brandenburger, A. The relationship between quantum and classical correlation in games. {\em Games and Economic Behavior} {\bf 2010}, {\em 69}, 175-183. \bibitem{36} La Mura, P. A Double-Slit Experiment for Non-Classical Interference Effects in Decision Making. {\em Topics in Cognitive Science}, {\bf 2014}, {\em 6}, 58-62. \bibitem{37} Yukalov, V.; Sornette, D. Conditions for Quantum Interference in Cognitive Sciences. {\em Topics in Cognitive Science}, {\bf 2014}, {\em 6}, 79-90. \bibitem{38} Busemeyer, J. R.; Bruza, P. D. {\em Quantum Models of Cognition and Decision}. Cambridge Univ. Press, New York, US, 2012. \end{thebibliography} |
1501.04070 | Title: An Information-Theoretic Alternative to the Cronbach's Alpha Coefficient
of Item Reliability
Abstract: We propose an information-theoretic alternative to the popular Cronbach alpha
coefficient of reliability. Particularly suitable for contexts in which
instruments are scored on a strictly nonnumeric scale, our proposed index is
based on functions of the entropy of the distributions of defined on the sample
space of responses. Our reliability index tracks the Cronbach alpha coefficient
uniformly while offering several other advantages discussed in great details in
this paper.
Body: \begin{frontmatter} \title{An Information-Theoretic Alternative to the Cronbach's Alpha Coefficient of Item Reliability} \runtitle{Information-Theoretic Measure of Reliability} \begin{aug} \author{\fnms{Ernest} \snm{Fokou\'e}\thanksref{t1,m1}\ead[label=e1]{epfeqa@rit.edu}}, \author{\fnms{Necla} \snm{G\"und\"uz}\thanksref{m2}\ead[label=e2]{ngunduz@gazi.edu.tr}} \thankstext{t1}{Corresponding author} \runauthor{Fokou\'e and G\"und\"uz} \affiliation{Rochester Institute of Technology\thanksmark{m1} and Gazi University\thanksmark{m2}} \address{\thanksmark{m1}School of Mathematical Sciences\\ Rochester Institute of Technology, Rochester, New York, USA\\ \printead{e1}} \address{\thanksmark{m2}Department of Statistics, Faculty of Science\\ Gazi University, Ankara, Turkey\\ \printead{e2}} \end{aug} \begin{abstract} We propose an information-theoretic alternative to the popular Cronbach alpha coefficient of reliability. Particularly suitable for contexts in which instruments are scored on a strictly nonnumeric scale, our proposed index is based on functions of the entropy of the distributions of defined on the sample space of responses. Our reliability index tracks the Cronbach alpha coefficient uniformly while offering several other advantages discussed in great details in this paper. \end{abstract} \begin{keyword}[class=AMS] \kwd[Primary ]{62H30} \kwd[; secondary ]{62H25} \end{keyword} \begin{keyword} \kwd{Cronbach} \kwd{Entropy} \kwd{Uncertainty} \kwd{Reliability} \kwd{Variation of Information} \end{keyword} \end{frontmatter} \section{Introduction} Suppose that we are given a dataset represented by an $n \times p$ matrix $\bfX$ whose $i$th row $\vx_i^\top \equiv (\rx_{i1}, \rx_{i2},\cdots,\rx_{ip})$ denotes the $p$-tuple of characteristics, with each $\rx_{ij} \in \{1,2,3,4,5\}$ representing the Likert-type level (order) of preference of respondent $i$ on item $j$. This Likert-type score is obtained by translating/mapping the response levels $\{{\tt Strong \, Disagree}, {\tt Disagree}, {\tt Neutral}, {\tt Agree}, {\tt Strongly \, Agree}\}$ into pseudo-numbers $\{1,2,3,4,5\}$. \begin{center} \begin{tabular}{|c|c|c|c|c|} \toprule {\tt Strong Disagree} & {\tt Disagree} & {\tt Neutral} & {\tt Agree} & {\tt Strongly Agree} \\ \hline $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ \\ \hline $1$ & $2$ & $3$ & $4$ & $5$ \\ \bottomrule \end{tabular} \end{center} \noindent A usually crucial part in the analysis of questionnaire data is the calculation of Cronbach's alpha coefficient which measures the internal consistency or reliability/quality of the data. Let $X= (X_1,X_2,\cdots,X_p)^\top$ be a $p$-tuple representing the $p$ items of a questionnaire. Initially proposed by and later used and re-explained extensively by thousands of researchers and practitioners like Cronbach's alpha coefficient is a function of the ratio of the sum of the idiosyncratic item variances over the variance of the sum of the items, and is given by $$ \alpha = \left(\frac{p}{p-1}\right)\left[1-\frac{\sum_{j=1}^p{\mathbb{V}(X_j)}}{\mathbb{V}\left(\sum_{\ell=1}^p{X_\ell}\right)}\right]. $$ \noindent The coefficient of Cronbach $\alpha$ will be $1$ if the items are all the same and $0$ if none is related to another. Because it is depend on the variance of the sum of a group of independent variables and the sum of their variances. If the variables are positively correlated, the variance of the sum will be increased. If the items making up the score are all identical and so perfectly correlated, all the ${\mathbb{V}(X_j)}$ will be equal and ${\mathbb{V}\left(\sum_{\ell=1}^p{X_\ell}\right)}=p^2 {\mathbb{V}(X_j)}$, so that $\frac{\sum_{j=1}^p{\mathbb{V}(X_j)}}{\mathbb{V}\left(\sum_{\ell=1}^p{X_\ell}\right)} = \frac{1}{p}$ and $\alpha = 1$.\\ The empirical version of Cronbach's alpha coefficient of internal consistency is given by $$ \widehat{{\alpha}} = \left(\frac{p}{p-1}\right)\left[1-\frac{\displaystyle \sum_{j=1}^p{\sum_{i=1}^n{\left(\rx_{ij}-\frac{1}{n}\sum_{i=1}^n{\rx_{ij}}\right)^2}}} {\displaystyle \sum_{i=1}^n{\left(\sum_{j=1}^p{\rx_{ij}}-\frac{1}{n}\sum_{i=1}^n{\sum_{j=1}^p{\rx_{ij}}}\right)^2}}\right]. $$ \begin{definition} Let $\mathcal{D}=\{\vx_1,\vx_2,\cdots,\vx_n\}$ be a dataset with $\vx_i^\top = (\rx_{i1}, \rx_{i2},\cdots,\rx_{ip})$. An observation vector $\vx_i$ will be called a {\it zero variation} vector if $\rx_{ij}= {\tt constant},\,\, j=1,\cdots,p$. Respondents with {\it zero variation} response vectors will be referred to as {\tt single minded} respondents/evaluators. \end{definition} In fact, {\it zero variation} responses essentially reduce a $p$ items survey to a single item survey. \begin{theorem} Let $X= (X_1,X_2,\cdots,X_p)^\top$ be a $p$-tuple representing the $p$ items of a questionnaire. If $X$ is {\it zero variation}, then the Cronbach's alpha coefficient will be equal to $1$. \end{theorem} \begin{proof} If $X= (X_1,X_2,\cdots,X_p)^\top$ is {\it zero variation}, then $X_j=W$ for $j=1,\cdots,p$, and $\sum_{j=1}^p{X_j}=pW$. As a result, $\sum_{j=1}^p{\mathbb{V}(X_j)}=p\mathbb{V}(W)$ and $\mathbb{V}\left(\sum_{j=1}^p{X_j}\right)=\mathbb{V}(pW)=p^2\mathbb{V}(W)$. Therefore, $$\alpha = \left(\frac{p}{p-1}\right)\left[1-\frac{p\mathbb{V}(W)}{p^2\mathbb{V}(W)}\right] = \left(\frac{p}{p-1}\right)\left[1-\frac{1}{p}\right] = 1 $$ \end{proof} We use a straightforward adaptation of the Cronbach's alpha coefficient to measure {\it respondent reliability}. \begin{definition} Let $\mathcal{D}=\{\vx_1,\vx_2,\cdots,\vx_n\}$ be a dataset with $\vx_i^\top = (\rx_{i1}, \rx_{i2},\cdots,\rx_{ip})$. Let the estimated variance of the $i$th respondent be ${\tilde{S}_i^2} = \sum_{j=1}^p{(\rx_{ij}-\bar{\rx}_i)^2/(p-1)}$. Let $Z_j = \sum_{i=1}^n{\rx_{ij}}$ represent the sum of the scores given by all the $n$ respondents to item $j$. Our respondent reliability is estimated by $$ \widehat{\tilde{\alpha}} = \left(\frac{n}{n-1}\right)\left[1-\frac{\displaystyle \sum_{i=1}^n{\sum_{j=1}^p{\left(\rx_{ij}-\frac{1}{p}\sum_{j=1}^p{\rx_{ij}}\right)^2}}} {\displaystyle \sum_{j=1}^p{\left(\sum_{i=1}^n{\rx_{ij}}-\frac{1}{p}\sum_{j=1}^p{\sum_{i=1}^n{\rx_{ij}}}\right)^2}}\right] $$ \end{definition} Given a data matrix $\bfX$, respondent reliability can be computed in practice by simply taking the Cronbach's alpha coefficient of $\bfX^\top$, the transpose of the data matrix $\bfX$. Let $m$ be the number of {\it nonzero variation}. If $m \ll p$ and $m/n$ is very small, then respondent reliability will be very poor. Despite its widespread use of Likert-type data since it creation, Cronbach's alpha coefficient is rigorously speaking not suitable for categorical data for the simple reason that averages on ordinal measurements are often difficult to interpret at best and misleading at worst. For many years researchers working on the clustering of Likert-type inappropriately resorted to average-driven methods like kMeans clustering. Fortunately, there has been a surge of contributions to the clustering of categorical data whereby appropriate methods have been used. At the heart of the clustering of categorical data is the need to define appropriate measure of similarity. Recognizing the possibility to preprocess Likert-type questionnaire data into a collection of estimate probability distributions over the sample spaces of responses, many authors have developed powerful, scalable and highly techniques for clustering categorical data, most of them based on information-theoretic concepts like entropy , , , , , , , , , mutual information, variation of information , along with many other distances and measures on probability distributions like the Bhattacharya distance , , , , , , the Kullback-Leibler divergence and the Hellinger distance just to name a few. In this paper, we use information-theoretic tools and concepts to create several measures of internal consistency of questionnaire data. \section{Information-Theoretic Measures of Internal Data Consistency} Let $X_j$ represent one of the questions on the questionnaire, and consider the $n$ responses, $\{\rx_{1j}, \cdots, \rx_{ij}, \cdots,\rx_{nj}\}$ provided by the $n$ evaluators. Let $\vv_j = (\rv_{j1}, \cdots,\rv_{jk},\cdots,\rv_{jK})^\top$ denote the vector containing the relative frequencies of each Likert level for question $j$. With a total of $n$ questionnaires collected, we have \begin{eqnarray} \widehat{\rv}_{jk} = \frac{1}{n}\sum_{i=1}^n{I({\rm x}_{ij}=k)}, \qquad k=1,2,\cdots,K \quad \text{and} \quad j=1,2,\cdots,p. \end{eqnarray} Using \eqref{eq:entr:item:0}, one can then form probabilistic vectors $\widehat{\vv}_j^\top = (\widehat{\rv}_{j1}, \cdots, \widehat{\rv}_{jk}, \cdots,\widehat{\rv}_{jK})$, for $j=1,2,\cdots,p$. Each vector $\widehat{\vv}_j$ essentially represents an approximate probability distribution on the sample space made up of the $K$ response levels. Using this probabilistic representation of each question $j$, we can compare the variability of each item of the questionnaire using the entropy, specifically \begin{eqnarray} {H}(\widehat{\vv}_j) = -\sum_{k=1}^K{\widehat{\rv}_{jk}\log_2(\widehat{\rv}_{jk})} \end{eqnarray} We can imagine a transformation of the $n \times p$ data matrix $\mX$ into a probabilistic $p \times K$ counterpart $\mV$ where each row represent the approximate probability distribution of the corresponding question (item). The entropy of each question indicates the variability of the answers given by students on that question. For a given course and a given instructor, a small value of this entropy would indicate a greater degree of agreement of his/her student on that item, and therefore suggest a more careful examination of the scores on that item. As far as the relationship between items is concerned, information theory also provides a wealth of measures. The symmetrized Kullback-Leibler divergence given by $$ {\tt KL}_2(\vv_i, \vv_j) = \frac{1}{{2}}\Big\{{\tt KL}(\vv_i, \vv_j) +{\tt KL}(\vv_j, \vv_i)\Big\} = \frac{1}{{2}}{\sum_{k=1}^K\left\{\rv_{ik}\log\left(\frac{\rv_{ik}}{\rv_{jk}}\right)+ \rv_{jk}\log\left(\frac{\rv_{jk}}{\rv_{ik}}\right)\right\}}, $$ where $$ {\tt KL}(\vv_i, \vv_j) = \sum_{k=1}^K{\rv_{ik}\log\left(\frac{\rv_{ik}}{\rv_{jk}}\right)} \quad \textrm{and} \quad {\tt KL}(\vv_j, \vv_i) = \sum_{k=1}^K{\rv_{jk}\log\left(\frac{\rv_{jk}}{\rv_{ik}}\right)}, $$ is usually the default measure used by most authors. The {\it Kullback-Leibler} divergence is closely related the {\it mutual information} $$ I(\vv_i, \vv_j) = \sum_{k=1}^K\left\{\sum_{l=1}^K\left\{\rv_{ik,jl}\log_2\left(\frac{\rv_{ik,jl}}{\rv_{ik}\rv_{jl}}\right)\right\}\right\}, $$ which has been used extensively in machine learning to define a distance known as the {\it Variation of Information}, and defined by $$ {\tt VI}(\vv_i, \vv_j) = H(\vv_i) + H(\vv_j)- 2 I(\vv_i, \vv_j). $$ Many other non-information-theoretic similarity and variation measures operating on probabilistic vectors can be used to further investigate several aspects of the categorical data at hand. One that have been extensively used in the machine learning and data mining community is the Bhattacharya distance is given by $$ {\tt BC}(\vv_i, \vv_j) = -\log {\tt F}(\vv_i, \vv_j), $$ where $$ {\tt F}(\vv_i, \vv_j) = \sum_{k=1}^K{\sqrt{\rv_{ik}\rv_{jk}}}, $$ is known as the Bhattacharya coefficient or Fidelity coefficient. The Bhattacharya distance ${\tt BC}(\vv_i, \vv_j)$ measures the overlap between $\vv_i$ and $\vv_j$. The Bhattacharya distance has been immensely used in various data mining and machine learning applications , , , . It is interesting to note that the Bhattachrya distance is related to {\it total variation} measure defined by $$ \Delta(\vv_i, \vv_j) = \frac{1}{{2}}\sum_{k=1}^K{|\rv_{ik}-\rv_{jk}|} = \frac{1}{{2}}\|\vv_i-\vv_j\|_1 $$ where $\|\cdot\|_1$ is the $\ell_1$ norm. Another very commonly used distance is the Hellinger distance between $\vv_i$ and $\vv_j$ is given by $$ {\tt Hellinger}(\vv_i, \vv_j) = \frac{1}{\sqrt{2}}\sqrt{\sum_{k=1}^K{(\sqrt{\rv_{ik}}-\sqrt{\rv_{jk}})^2}} =\frac{1}{\sqrt{2}}\|\sqrt{\vv_i}-\sqrt{\vv_j}\|_2, $$ where $\|\cdot\|_2$ is the Euclidean norm or $\ell_2$ norm, $\sqrt{\vv_i} = (\sqrt{\rv_{i1}},\cdots,\sqrt{\rv_{iK}})$ and $\sqrt{\vv_j} = (\sqrt{\rv_{j1}},\cdots,\sqrt{\rv_{jK}})$. \begin{definition} Let $Q$ denote an instrument (questionnaire) for which the realized matrix of obtained responses is given by $\bfX$ with entries $\rx_{ij} \in \{1,2, \cdots, K\}$. We propose an information-theoretic measure of the reliability of $Q$, referred to as the information consistency ratio of $Q$ and given by \begin{eqnarray} \varphi = 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}} {\underset{\vz}{\mathtt{max}}\Big\{H\left({\vz}\right)\Big\}} = 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}}{H\left(\frac{1}{K},\cdots,\frac{1}{K}\right)}= 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}}{\log_2(K)}, \end{eqnarray} where each $\widehat{\vz}_i = \{\widehat{\rz}_{ik}, \, k=1,2,\cdots,K\}$ defines an approximate probability distribution on the sample space of possible responses, and $H(\cdot)$ is the entropy function, with \begin{eqnarray} \widehat{\rz}_{ik} = \frac{1}{p}\sum_{j=1}^p{I({\rm x}_{ij}=k)} \quad \text{and} \quad {H}(\widehat{\vz}_i) = -\sum_{k=1}^K{\widehat{\rz}_{ik}\log_2(\widehat{\rz}_{ik})}. \end{eqnarray} \end{definition} \begin{lemma} Let $\vz$ denote any probability measure defined on some $K$-dimensional sample space, with each $\rz_k = \Pr\{E_k\}, \, k=1.2,\cdots,K$. Let $H(\cdot)$ denote the entropy function, such that for every $\vz$, we have $H(\vz) = -\sum_{k=1}^K{{\rz}_{k}\log_2({\rz}_{k})}$. Then $$ \underset{\vz}{\mathtt{max}}\Big\{H\left({\vz}\right)\Big\} = \log_2(K). $$ \end{lemma} \begin{proof} Since entropy essentially measures uncertainty (disturbance), the probability measure for which the uncertainty is the largest is the probability measure $\vz^*$ in which all the events are equally likely, i.e., $\rz_k^* = \Pr\{E_k\}=\frac{1}{K}, \, k=1.2,\cdots,K$. \begin{eqnarray*} \underset{\vz}{\mathtt{max}}\Big\{H\left({\vz}\right)\Big\} = H(\vz^*) = H\left(\frac{1}{K},\cdots,\frac{1}{K}\right) = -\sum_{k=1}^K{\frac{1}{K}\log_2\left(\frac{1}{K}\right)}=\log_2(K). \end{eqnarray*} \end{proof} \begin{proposition} Let $Q_0$ denote a special questionnaire whose items are all mutually independent (unrelated). Then the corresponding information consistency ratio $\varphi_0$ of $Q_0$, is such that $$ \underset{p \rightarrow \infty}{\lim}{\varphi_0} = 0. $$ \end{proposition} \begin{proof} With $Q_0$ denoting a questionnaire whose items that are all mutually independent (unrelated), the matrix of realized responses has entries $\rx_{ij}$ that a realization of the discrete uniform distribution on $\{1,2,\cdots,K\}$, or specifically, $\rx_{ij} \sim {\tt uniform}(1,2,\cdots,K)$. It follows that for each $i=1,2,\cdots,n$, we must have $$ \underset{p \rightarrow \infty}{\lim}{\widehat{\rz}_{ik}} = \underset{p \rightarrow \infty}{\lim}\left\{\frac{1}{p}\sum_{j=1}^p{I({\rm x}_{ij}=k)}\right\} = \frac{1}{K}, \quad k=1,2,\cdots,K. $$ In other words, given enough questions (items), the empirical proportion of answers will converge to its theoretical counterpart by the law of large number. We therefore have the uniform generation of answers, the limiting distribution $$ \underset{p \rightarrow \infty}{\lim}{\widehat{\vz}_{i}} = \vz^* = \left(\frac{1}{K},\cdots,\frac{1}{K}\right). $$ Finally, since all the response distributions will tend to converge to the same maximal measure $\vz^*$, i.e. $\widehat{\vz}_{i} \overset{\mathscr{D}}{\rightarrow} \vz^*$, for $i=1,2,\cdots,n$, we must have $$ \underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\} \overset{P}{\rightarrow} H(\vz^*) = \underset{\vz}{\mathtt{max}}\Big\{H\left({\vz}\right)\Big\}, $$ and therefore $$ \underset{p \rightarrow \infty}{\lim}{\varphi_0} = 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}} {\underset{\vz}{\mathtt{max}}\Big\{H\left({\vz}\right)\Big\}} = 1 - \frac{H(\vz^*)}{H(\vz^*)} = 1-1 = 0. $$ \end{proof} \begin{proposition} Let $Q_+$ denote a special questionnaire whose items are all identical. Then the corresponding information consistency ratio $\varphi_+$ of $Q_+$, is such that ${\lim}{\varphi_+} = 1$. \end{proposition} \begin{proof} With $Q_+$ denoting a questionnaire whose items that are all identical, the matrix of realized responses has entries $\rx_{ij} = c$, for some constant $c \in \{1,2,\cdots,K\}$. Then for each $i=1,2,\cdots,n$, there exists $k_+ \in \{1,2,\cdots,K\}$ such that $$ \widehat{\rz}_{ik} = \left\{ \begin{array}{ll} 1 & \quad k = k_+ \\ 0 & \quad k \neq k_+ \end{array}\right. $$ In other words, with $Q_+$, the approximate distributions $\widehat{\vz}_i$ of the answers of each respondent are of the form $(1,0,\cdots,0)$, or $(0,1,\cdots,0)$ or $(0,0,\cdots,1)$. Therefore, for $Q_+$, we must have $H(\widehat{\vz}_i) = 0, \quad i=1,\cdots,n$, with the result being $\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}=0$, and therefore $$ \varphi_+ = 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}} {\underset{\vz}{\mathtt{max}}\Big\{H\left({\vz}\right)\Big\}} = 1 - \frac{0}{H(\vz^*)} = 1-0 = 1. $$ \end{proof} \begin{definition} Let $Y_i$ represent the most frequently occurring answer in respondent $i$'s vector of $p$ answers. It is easy to see that $Y_i$ has the same sample space as each question/item, namely the same Likert scale in our case. Using the random variables $Y_i$, we can then define $\widehat{\vw} = (\widehat{\rw}_{1}, \cdots,\widehat{\rw}_{k},\cdots,\widehat{\rw}_{K})^\top$ in the same manner that we define $\vv_j$ earlier. More specifically, we have \begin{eqnarray} Y_i = \underset{k=1,\cdots,K}{\mathtt{argmax}}\left\{\frac{1}{p}\sum_{j=1}^p{I(X_{ij}=k)}\right\} \quad \text{and} \quad \widehat{\rw}_{k} = \frac{1}{n}\sum_{i=1}^n{I(Y_{i}=k)}. \end{eqnarray} The entropy of $\widehat{\vw}$ is given by \begin{eqnarray} H(\widehat{\vw}) = -\sum_{k=1}^K{\widehat{\rw}_{k}\log_2(\widehat{\rw}_{k})}. \end{eqnarray} {\it The random variable $Y_i$ is maximal in a set-theoretic sense, and and can be thought of as the categorical analogue of the sum of numeric $X_j$'s.} Using $\widehat{\vw}$, an alternative definition of the {\it information consistency ratio} $\varphi$ is \begin{eqnarray} \varphi= 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{min}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}} {H\left({\widehat{\vw}}\right)}. \end{eqnarray} An even more stringent measure of the information consistency ratio is given by \begin{eqnarray} \varphi= 1 - \frac{\underset{i=1,\cdots,n}{\mathtt{max}}\Big\{H\left(\widehat{\vz}_i\right)\Big\}} {H\left({\widehat{\vw}}\right)}. \end{eqnarray} \end{definition} \section{Demonstration of Properties of $\varphi$} We use a simple simulation setup to empirically compare the different measures presented in this paper. We set $p=50$ and $n=1000$ and we vary the ratio of perfectly reliable components from $10\ $100\ $\{1,2,\cdots,K\}$, that is, $$ {\tt Draw} \quad \rx_{ij} \sim {\tt uniform}(1,2,\cdots,K). $$ Randomly replace $100{\tt c}\ ${\tt c} \in \{0.1,0.2,\cdots,0.9,1\}$. Table \eqref{tab:empirical:entr:cronb:1} shows the simulated values of the information consistency ratio and Cronbach's alpha coefficient for different fractions of of reliable components in the instrument. Figure \eqref{fig:entropy:vs:cronbach:1} is a direct pictorial representation of the numbers from Table \eqref{tab:empirical:entr:cronb:1}, and we can see that the Cronbach alpha coefficient is less strick than the information consistency ratio. \begin{table}[!htbp] {\begin{tabular}{@{\extracolsep{5pt}} lrrrrrrrrrr} \\[-1.8ex]\toprule \\[-1.8ex] Fraction & $10$ & $20$ & $30$ & $40$ & $50$ & $60$ & $70$ & $80$ & $90$ & $100$ \\ \hline $\varphi_1$ & $0.230$ & $0.270$ & $0.330$ & $0.440$ & $0.520$ & $0.630$ & $0.740$ & $0.900$ & $1.000$ & $1.000$ \\ $\varphi_2$ & $0.000$ & $0.000$ & $0.020$ & $0.080$ & $0.140$ & $0.240$ & $0.360$ & $0.520$ & $0.720$ & $1.000$ \\ $\varphi_3$ & $0.230$ & $0.270$ & $0.330$ & $0.440$ & $0.520$ & $0.630$ & $0.740$ & $0.900$ & $1.000$ & $1.000$ \\ $\varphi_4$ & $0.000$ & $0.000$ & $0.020$ & $0.080$ & $0.140$ & $0.240$ & $0.360$ & $0.520$ & $0.720$ & $1.000$ \\ ${\tt Cronbach}$ & $0.380$ & $0.700$ & $0.820$ & $0.910$ & $0.940$ & $0.960$ & $0.980$ & $0.990$ & $1.000$ & $1.000$ \\ \bottomrule \\[-1.8ex] \end{tabular}} \caption{Simulated values of the information consistency ratio and Cronbach's alpha coefficient for different fractions of reliable components in the instrument.} \end{table} \section{Conclusion and Discussion} We have proposed and developed an information-theoretic measure of internal data consistency et demonstrated via straightforward simulation that it does indeed capture the amount of information potentially contained in the data for,the purposes of performing all kinds of pattern for the data. We have also provided several many other measures of similarity over probabilistic vectors that we intend to use for further refined our proposed information consistency ratio $\varphi$. We intend to conduct a larger simulation study to establish our proposed measure on a stronger footing. We also plan to compare the predictive power of ICR to Cronbach's alpha coefficient on various real and simulated data. \section*{Acknowledgements} Ernest Fokou\'e wishes to express his heartfelt gratitude and infinite thanks to Our Lady of Perpetual Help for Her ever-present support and guidance, especially for the uninterrupted flow of inspiration received through Her most powerful intercession. \bibliographystyle{chicago} \bibliography{fg-entropy-ref} \def\reduce#1{{\small #1}} \def\c#1{\mathcal #1} \def\foot#1{} \newcommand{\nmathbf}{\bm} \def\b#1{{\bf #1}} \def\bfA{\nmathbf A} \def\bfB{\nmathbf B} \def\bfC{\nmathbf C} \def\bfD{\nmathbf D} \def\bfE{\nmathbf E} \def\bfF{\nmathbf F} \def\bfG{\nmathbf G} \def\bfH{\nmathbf H} \def\bfI{\nmathbf I} \def\bfJ{\nmathbf J} \def\bfK{\nmathbf K} \def\bfL{\nmathbf L} \def\bfM{\nmathbf M} \def\bfN{\nmathbf N} \def\bfO{\nmathbf O} \def\bfP{\nmathbf P} \def\bfQ{\nmathbf Q} \def\bfR{\nmathbf R} \def\bfS{\nmathbf S} \def\bfT{\nmathbf T} \def\bfU{\nmathbf U} \def\bfV{\nmathbf V} \def\bfW{\nmathbf W} \def\bfX{\nmathbf X} \def\bfY{\nmathbf Y} \def\bfZ{\nmathbf Z} \def\bfa{\nmathbf a} \def\bfb{\nmathbf b} \def\bfc{\nmathbf c} \def\bfd{\nmathbf d} \def\bfe{\nmathbf e} \def\bff{\nmathbf f} \def\bfg{\nmathbf g} \def\bfh{\nmathbf h} \def\bfi{\nmathbf i} \def\bfj{\nmathbf j} \def\bfk{\nmathbf k} \def\bfl{\nmathbf l} \def\bfm{\nmathbf m} \def\bfn{\nmathbf n} \def\bfo{\nmathbf o} \def\bfp{\nmathbf p} \def\bfq{\nmathbf q} \def\bfr{\nmathbf r} \def\bfs{\nmathbf s} \def\bft{\nmathbf t} \def\bfu{\nmathbf u} \def\bfv{\nmathbf v} \def\bfw{\nmathbf w} \def\bfx{\nmathbf x} \def\bfy{\nmathbf y} \def\bfz{\nmathbf z} \def\bfalpha {\nmathbf \alpha} \def\bfbeta {\nmathbf \beta} \def\bfgamma {\nmathbf \gamma} \def\bfdelta {\nmathbf \delta} \def\bfepsilon{\nmathbf \epsilon} \def\bfzeta {\nmathbf \zeta} \def\bfeta {\nmathbf \eta} \def\bftheta {\nmathbf \theta} \def\bfiota {\nmathbf \iota} \def\bfkappa {\nmathbf \kappa} \def\bflambda {\nmathbf \lambda} \def\bfmu {\nmathbf \mu} \def\bfnu {\nmathbf \nu} \def\bfxi {\nmathbf \xi} \def\bfomicron{\nmathbf \omicron} \def\bfpi {\nmathbf \pi} \def\bfrho {\nmathbf \rho} \def\bfsigma {\nmathbf \sigma} \def\bftau {\nmathbf \tau} \def\bfupsilon{\nmathbf \upsilon} \def\bfphi {\nmathbf \phi} \def\bfpsi {\nmathbf \psi} \def\bfchi {\nmathbf \chi} \def\bfomega {\nmathbf \omega} \def\bfAlpha {\nmathbf \Alpha} \def\bfBeta {\nmathbf \Beta} \def\bfGamma {\nmathbf \Gamma} \def\bfDelta {\nmathbf \Delta} \def\bfEpsilon{\nmathbf \Epsilon} \def\bfZeta {\nmathbf \Zeta} \def\bfEta {\nmathbf \Eta} \def\bfTheta {\nmathbf \Theta} \def\bfIota {\nmathbf \Iota} \def\bfKappa {\nmathbf \Kappa} \def\bfLambda {\nmathbf \Lambda} \def\bfMu {\nmathbf \Mu} \def\bfNu {\nmathbf \Nu} \def\bfXi {\nmathbf \Xi} \def\bfOmicron{\nmathbf \Omicron} \def\bfPi {\nmathbf \Pi} \def\bfRho {\nmathbf \Rho} \def\bfPsi {\nmathbf \Psi} \def\bfSigma {\nmathbf \Sigma} \def\bfTau {\nmathbf \Tau} \def\bfUpsilon{\nmathbf \Upsilon} \def\bfPhi {\nmathbf \Phi} \def\bfChi {\nmathbf \Chi} \def\bfOmega {\nmathbf \Omega} \newcommand{\ttheta}{\tilde{\theta}} \newcommand{\bfzero}{{\nmathbf 0}} \newcommand{\bfone}{{\nmathbf 1}} \newcommand{\vareps}{\varepsilon} \def\bfvareps{\nmathbf \varepsilon} \newcommand{\tgamma}{\tilde\gamma} \newcommand{\cfA}{\c{A}} \newcommand{\cfB}{\c{B}} \newcommand{\cfC}{\c{C}} \newcommand{\cfD}{\c{D}} \newcommand{\cfE}{\c{E}} \newcommand{\cfF}{\c{F}} \newcommand{\cfG}{\c{G}} \newcommand{\cfH}{\c{H}} \newcommand{\cfI}{\c{I}} \newcommand{\cfJ}{\c{J}} \newcommand{\cfK}{\c{K}} \newcommand{\cfL}{\c{L}} \newcommand{\cfM}{\c{M}} \newcommand{\cfN}{\c{N}} \newcommand{\cfO}{\c{O}} \newcommand{\cfP}{\c{P}} \newcommand{\cfQ}{\c{Q}} \newcommand{\cfR}{\c{R}} \newcommand{\cfS}{\c{S}} \newcommand{\cfT}{\c{T}} \newcommand{\cfU}{\c{U}} \newcommand{\cfV}{\c{V}} \newcommand{\cfX}{\c{X}} \newcommand{\cfY}{\c{Y}} \newcommand{\cfZ}{\c{Z}} \def\boldfacefake#1{\kern-4pt \hbox{ \mathsurround=0pt \hbox to 0.4pt{$#1$\hss}\hbox to 0.4pt{$#1$\hss}\hbox {$#1$}}} \newcommand{\bcfA}{\boldsymbol{\mathcal{A}}} \newcommand{\bcfB}{\boldsymbol{\mathcal{B}}} \newcommand{\bcfC}{\boldsymbol{\mathcal{C}}} \newcommand{\bcfD}{\boldsymbol{\mathcal{D}}} \newcommand{\bcfE}{\boldsymbol{\mathcal{E}}} \newcommand{\bcfF}{\boldsymbol{\mathcal{F}}} \newcommand{\bcfG}{\boldsymbol{\mathcal{G}}} \newcommand{\bcfH}{\boldsymbol{\mathcal{H}}} \newcommand{\bcfI}{\boldsymbol{\mathcal{I}}} \newcommand{\bcfJ}{\boldsymbol{\mathcal{J}}} \newcommand{\bcfK}{\boldsymbol{\mathcal{K}}} \newcommand{\bcfL}{\boldsymbol{\mathcal{L}}} \newcommand{\bcfM}{\boldsymbol{\mathcal{M}}} \newcommand{\bcfN}{\boldsymbol{\mathcal{N}}} \newcommand{\bcfO}{\boldsymbol{\mathcal{O}}} \newcommand{\bcfP}{\boldsymbol{\mathcal{P}}} \newcommand{\bcfQ}{\boldsymbol{\mathcal{Q}}} \newcommand{\bcfR}{\boldsymbol{\mathcal{R}}} \newcommand{\bcfS}{\boldsymbol{\mathcal{S}}} \newcommand{\bcfT}{\boldsymbol{\mathcal{T}}} \newcommand{\bcfU}{\boldsymbol{\mathcal{U}}} \newcommand{\bcfV}{\boldsymbol{\mathcal{V}}} \newcommand{\bcfW}{\boldsymbol{\mathcal{W}}} \newcommand{\bcfX}{\boldsymbol{\mathcal{X}}} \newcommand{\bcfY}{\boldsymbol{\mathcal{Y}}} \newcommand{\bcfZ}{\boldsymbol{\mathcal{Z}}} \newcommand{\g}{\,\vert\,} \newcommand{\bg}{\;\bigg\vert\;} \newcommand{\p}{\mbox{P}} \newcommand{\D}{\mbox{D}} \newcommand{\E}{\mbox{E}} \newcommand{\Mo}{\mbox{Mo}} \newcommand{\Me}{\mbox{Me}} \newcommand{\Cov}{\mbox{Cov}} \newcommand{\Var}{\mbox{Var}} \newcommand{\Corr}{\mbox{Corr}} \newcommand{\Q}{\mbox{Q}} \newcommand{\Bb}{\mbox{Bb}} \newcommand{\Be}{\mbox{Be}} \newcommand{\Bi}{\mbox{Bi}} \newcommand{\Br}{\mbox{Br}} \newcommand{\Ca}{\mbox{Ca}} \newcommand{\Di}{\mbox{Di}} \newcommand{\Ex}{\mbox{Ex}} \newcommand{\Fs}{\mbox{Fs}} \newcommand{\Ge}{\mbox{Ge}} \newcommand{\Gg}{\mbox{Gg}} \newcommand{\Hy}{\mbox{Hy}} \newcommand{\Ig}{\mbox{Ig}} \newcommand{\Ip}{\mbox{Ip}} \newcommand{\Lo}{\mbox{Lo}} \newcommand{\Mu}{\mbox{Mu}} \newcommand{\N}{\mbox{N}} \newcommand{\Nb}{\mbox{Nb}} \newcommand{\Ng}{\mbox{Ng}} \newcommand{\Nw}{\mbox{Nw}} \newcommand{\Pa}{\mbox{Pa}} \newcommand{\Po}{\mbox{Po}} \newcommand{\Pg}{\mbox{Pg}} \newcommand{\Pn}{\mbox{Pn}} \newcommand{\Ra}{\mbox{Ra}} \newcommand{\St}{\mbox{St}} \newcommand{\Un}{\mbox{Un}} \newcommand{\Wi}{\mbox{Wi}} \newcommand{\dd}[1]{\,d#1} \newcommand{\barx}{\mbox{$\overline x$}} \newcommand{\comb}[2]{{#1\choose#2}} \newcommand{\ontop}[2]{{#1\atop#2}} \newcommand{\h}{\hbox{{\small$1\over2$}}} \newcommand{\ok}{\hfill\fbox{}} \newcommand{\brow}[2]{\mbox{$\{{#1}_1,\ldots,{#1}_{#2}\}$}} \newcommand{\prow}[2]{\mbox{$({#1}_1,\ldots,{#1}_{#2})$}} \newcommand{\row}[2]{\mbox{${#1}_1,\ldots,{#1}_{#2}$}} \newcommand{\data}{\row{x}{n}} \newcommand{\bdata}{\prow{x}{n}} \newcommand{\ie}{\emph{i.e.},\ } \newcommand{\co}{\emph{cf.}\ } \newcommand{\eg}{\emph{e.g.}, } \newcommand{\etalc}{\emph{et al.},\ } \newcommand{\etal}{\emph{et al.}\ } \newcommand{\be}{\begin{eqnarray}} \newcommand{\ee}{\end{eqnarray}} \newcommand{\ba}{\begin{eqnarray*}} \newcommand{\ea}{\end{eqnarray*}} \newcommand{\go}{\rightarrow} \newcommand{\gor}{\rightarrow} \newcommand{\goi}{\rightarrow \infty} \newcommand{\ol}{\overline} \newcommand{\fr}{\frac} \newcommand{\pn}{\par\noindent} \newcommand{\nc}{\nonumber\\} \newcommand{\ssum}{\mbox{$\sum$}} \newcommand{\hhline}{\hline\hline} \newcommand{\reals}{\mbox{\rm I\kern-.20em R}} \newcommand{\sreals}{\mbox{\small \rm I\kern-.20em R}} \newcommand{\mylinel}{\renewcommand{\baselinestretch}{1.8}\tiny\small} \newcommand{\goto}{\rightarrow} \newcommand{\expect}{\E} \newcommand{\pp}[1]{\,\vskip2mm\noindent{\it #1.}~\ignorespaces} \newcommand{\frot}{\frac{1}{2}} \newcommand{\PIN}{\mbox{PIN}} \newcommand{\mSSE}{\mbox{SSE}} |
1501.04073 | Title: Antineutrino induced Lambda(1405) production off the proton
Abstract: We have studied the strangeness changing antineutrino induced reactions
$\bar{\nu}_{l} p \rightarrow l^+ \phi B $, with $\phi B = K^-p$, $\bar{K}^0n$,
$\pi^0\Lambda$, $\pi^0\Sigma^0$, $\eta\Lambda$, $\eta\Sigma^0$,
$\pi^+\Sigma^-$, $\pi^-\Sigma^+$, $K^+\Xi^-$ and $K^0\Xi^0$, using a chiral
unitary approach. These ten coupled channels are allowed to interact strongly,
using a kernel derived from the chiral Lagrangians. This interaction generates
two $\Lambda(1405)$ poles, leading to a clear single peak in the $\pi \Sigma$
invariant mass distributions. At backward scattering angles in the center of
mass frame, $\bar{\nu}_{\mu} p \rightarrow \mu^+ \pi^0 \Sigma^0$ is dominated
by the $\Lambda(1405)$ state at around 1420~MeV while the lighter state becomes
relevant as the angle decreases, leading to an asymmetric line shape. In
addition, there are substantial differences in the shape of $\pi \Sigma$
invariant mass distributions for the three charge channels. If observed, these
differences would provide valuable information on a claimed isospin I=1,
strangeness S=-1 baryonic state around 1400 MeV. Integrated cross sections have
been obtained for the $\pi \Sigma$ and $\bar K N$ channels, investigating the
impact of unitarization in the results. The number of events with
$\Lambda(1405)$ excitation in $\bar\nu_\mu p$ collisions in the recent
antineutrino run at the MINERvA experiment has also been obtained. We find that
this reaction channel is relevant enough to be investigated experimentally and
to be taken into account in the simulation models of future experiments with
antineutrino beams.
Body: \title{Antineutrino induced $\mathbf{\Lambda(1405)}$ production off the proton} \author{Xiu-Lei Ren} \email[E-mail: ]{xiulei.ren@buaa.edu.cn} \affiliation{School of Physics and Nuclear Energy Engineering \& International Research Center for Nuclei and Particles in the Cosmos, Beihang University, Beijing 100191, China} \affiliation{Departamento de F\'{\i}sica Te\'{o}rica and IFIC, Centro Mixto Universidad de Valencia-CSIC, Institutos de Investigaci\'{o}n de Paterna, Apartado 22085, 46071 Valencia, Spain} \affiliation{Institut de Physique Nucl\'{e}aire, IN2P3-CNRS and Universit\'{e} Paris-Sud, F-91406 Orsay Cedex, France} \author{E. Oset} \email[E-mail: ]{oset@ific.uv.es} \affiliation{Departamento de F\'{\i}sica Te\'{o}rica and IFIC, Centro Mixto Universidad de Valencia-CSIC, Institutos de Investigaci\'{o}n de Paterna, Apartado 22085, 46071 Valencia, Spain} \author{L. Alvarez-Ruso} \email[E-mail: ]{alvarez@ific.uv.es} \affiliation{Instituto de F\'isica Corpuscular (IFIC), Centro Mixto Universidad de Valencia-CSIC, Institutos de Investigaci\'{o}n de Paterna, Apartado 22085, 46071 Valencia, Spain} \author{M. J. Vicente Vacas} \email[E-mail: ]{vicente@ific.uv.es} \affiliation{Departamento de F\'{\i}sica Te\'{o}rica and IFIC, Centro Mixto Universidad de Valencia-CSIC, Institutos de Investigaci\'{o}n de Paterna, Apartado 22085, 46071 Valencia, Spain} \begin{abstract} We have studied the strangeness changing antineutrino induced reactions $\bar{\nu}_{l} p \rightarrow l^+ \phi B $, with $\phi B = K^-p$, $\bar{K}^0n$, $\pi^0\Lambda$, $\pi^0\Sigma^0$, $\eta\Lambda$, $\eta\Sigma^0$, $\pi^+\Sigma^-$, $\pi^-\Sigma^+$, $K^+\Xi^-$ and $K^0\Xi^0$, using a chiral unitary approach. These ten coupled channels are allowed to interact strongly, using a kernel derived from the chiral Lagrangians. This interaction generates two $\Lambda(1405)$ poles, leading to a clear single peak in the $\pi \Sigma$ invariant mass distributions. At backward scattering angles in the center of mass frame, $\bar{\nu}_{\mu} p \rightarrow \mu^+ \pi^0 \Sigma^0$ is dominated by the $\Lambda(1405)$ state at around 1420~MeV while the lighter state becomes relevant as the angle decreases, leading to an asymmetric line shape. In addition, there are substantial differences in the shape of $\pi \Sigma$ invariant mass distributions for the three charge channels. If observed, these differences would provide valuable information on a claimed isospin $I=1$, strangeness $S=-1$ baryonic state around $1400$~MeV. Integrated cross sections have been obtained for the $\pi \Sigma$ and $\bar K N$ channels, investigating the impact of unitarization in the results. The number of events with $\Lambda(1405)$ excitation in $\bar\nu_\mu p$ collisions in the recent antineutrino run at the MINER$\nu$A experiment has also been obtained. We find that this reaction channel is relevant enough to be investigated experimentally and to be taken into account in the simulation models of future experiments with antineutrino beams. \end{abstract} \pacs{25.30.pt, 12.15.-y, 12.39.Fe, 13.15.+g, 14.20.Gk} \keywords{Neutrino interactions with Hadrons, Baryon resonances, Chiral Lagrangians} \maketitle \section{Introduction} The $\Lambda(1405)$ resonance is a cornerstone in hadron physics, challenging the standard view of baryons made of three quarks. Long ago it was already suggested that the $\Lambda(1405)$ could be a kind of molecular state arising from the interaction of the $\pi\Sigma$ and $\bar{K}N$ channels~. This view has been recurrent~, but only after the advent of unitary chiral perturbation theory (UChPT) has it taken a more assertive tone~. In this framework, a kernel (potential) derived from the chiral Lagrangians is the input into the Bethe-Salpeter equation in coupled channels. Sometimes the interaction is strong enough to generate poles, denominated as dynamically generated states, which can be interpreted as hadronic molecules with components on the different channels (see Ref.~ for a review). It came as a surprise that UChPT predicts two $\Lambda(1405)$ states~, studied in detail in Ref.~. Two poles appear, one around 1420~MeV with a width of about 40~MeV and another one around 1385~MeV with a larger width of about 150~MeV. These findings have been reconfirmed in more recent studies with potentials that include higher order terms of the chiral Lagrangians~. From the experimental perspective, the old experiments~ produced $\pi\Sigma$ invariant mass distributions where a single $\Lambda(1405)$ peak is seen around $1405$~MeV. According to Ref.~, this single peak results from the overlap of the two pole contributions. It has also been suggested that reactions induced by $K^- p$ pairs show a peak around 1420~MeV because the pole at 1420~MeV couples mostly to $\bar{K}N$, while the one at $1385$ MeV does it more strongly to $\pi\Sigma$. This would be the case of $K^-p\rightarrow \gamma \pi \Sigma$~ and $K^-p\rightarrow \pi^0\pi^0\Sigma^0$. The latter one, measured at Crystal Ball~ and analyzed in Ref.~, confirmed the existence of the state at $1420$~MeV. Another reaction that has proved its existence is $K^- d \rightarrow n \pi \Sigma$~, which was studied in Ref.~. The issues raised in Ref.~ were addressed in detail in Ref.~ reconfirming the findings of Ref.~. It is somewhat surprising that the two poles emerge in the theory even when only data on $K^-p$ scattering and $K^-p$ atoms~, which are above the $\Lambda(1405)$ pole masses, are fitted. Nevertheless, it is clear that the best information on the $\Lambda(1405)$ properties should come from processes where the $\Lambda(1405)$ is produced close to its pole masses. In this sense, the abundant $\Lambda(1405)$ photoproduction data obtained by CLAS with the $\gamma p\rightarrow K^+\pi^+\Sigma^-$, $K^+\pi^0\Sigma^0$, $K^+\pi^-\Sigma^+$ reactions~ add much information to the earlier data of Ref.~, bringing new light into the subject. A fit to these data imposing unitarity in the $\pi\Sigma, \bar{K}N$ channels and allowing only small variations in the kernel of the chiral Lagrangians~ has reconfirmed the existence of the two poles, in agreement with the UChPT predictions. The wide range of energies investigated and the simultaneous measurement of the three $\pi\Sigma$ charged channels were the key to the solutions found in Refs.~ and, more recently, in Ref.~. Studies of $p \, p \rightarrow p \, K^+ \, \Lambda(1405)$ performed at ANKE show again a superposition of the contributions from the two poles~, and can be explained with the theoretical framework of UChPT~. More recent measurements~ show the $\Lambda(1405)$ peak at a lower energy than in the ANKE experiment~. Some reasons for this behavior have been suggested in Ref.~. If more data for this reaction on different conditions became available, a global analysis like the one of Ref.~ for photoproduction would be advisable. In between, $\Lambda(1405)$ electroproduction~ data [$e \, p \rightarrow e' K^+ \, \Lambda(1405)$] have unexpectedly revealed a two-peak structure, albeit with large uncertainties. Previous measurements with different reactions have only observed a single peak coming from the superposition of the two poles, with different shapes depending on the weight of either pole, as determined by the dynamics of each process. Lattice QCD simulations have also brought new light into the $\Lambda(1405)$ properties. Using three-quark interpolators, a state associated with the $\Lambda(1405)$ is produced~. The vanishing strange quark contribution to the $\Lambda(1405)$ magnetic moment for light quark masses close to the physical ones has been interpreted~ as an evidence of a large $\bar K N$ component in the wave function of the $\Lambda(1405)$. Further work along these lines was reported in Ref.~ using {\it synthetic} lattice results from $\bar K N$ and $\pi \Sigma$ interpolators. These lead to the right description of the meson-baryon amplitudes in the continuum and contain the two poles in the complex plane. Until now, the weak excitation of $\Lambda(1405)$ has never been investigated. It is remarkable that while its production in strong and electromagnetic processes has to involve an extra strange particle (usually a $K^-$ in the initial state or a $K^+$ in the final one), the direct excitation of $\Lambda(1405)$ induced by antineutrinos $\bar\nu_l p \rightarrow l^+ \Lambda(1405)$ is allowed although Cabibbo suppressed. Notice that in $\Lambda(1405)$ photo and electroproduction there are line shape distortions due to final state interactions between the $K^+$ and the $\Lambda(1405)$ decay products, which are absent in the weak reaction. Stimulated by the precision needs of neutrino oscillation experiments, there is a significant ongoing effort aimed at a better understanding of neutrino cross sections with nucleons and nuclei. The goal is to develop better interaction models to reduce systematic errors in the detection process, constrain irreducible backgrounds and achieve a better neutrino energy determination. In the recent past, several experiments have produced valuable cross section measurements (see Ref.~ for a comprehensive review of the available data). The MINER$\nu$A experiment~ at FNAL, fully dedicated to the study of neutrino interactions with different target materials has recently completed data taking and started to produce interesting results~. In the few-GeV energy region, where several of the current and future experiments operate, quasielastic scattering and single pion production have the largest cross sections but strange particle production is also relevant. The charged-current $\Delta S =-1$ quasielastic hyperon ($Y =\Lambda,\Sigma$) production by antineutrinos has been investigated~ and found to be a non-negligible source of pions through the $Y \rightarrow N \, \pi$ decay~. Among the inelastic processes, associated ($\Delta S =0$) production of $\bar K$ and $\Sigma$ or $\Lambda$ baryons is the dominant one but has a high threshold. Below it, single $K$ ($\Delta S =1$) and single $\bar K$ ($\Delta S =-1$) can be produced in charged current interactions induced by $\nu$ and $\bar \nu$ respectively. These processes have been recently studied using SU(3) chiral Lagrangians at leading order~. The weak hadronic currents and the corresponding cross sections at threshold are constrained by chiral symmetry with couplings extracted from pion and hyperon semileptonic decays. As stressed in Ref.~, while the derived $K$ production cross section is a robust prediction at threshold, the situation could be different for $\bar K$ production due to the presence of the $\Lambda(1405)$ resonance just below the $\bar K N$ threshold. Another, so far unexplored, $\Delta S = -1$ reaction that can occur below the associated production threshold, $\bar\nu_l \, p \rightarrow l^+ \, \Sigma \, \pi$, is bound to get an important contribution from $\Lambda(1405)$ excitation. Here we report the first study of the antineutrino induced reactions $\bar{\nu}_l p \rightarrow l^+ \phi B $ with $\phi B =$ $K^-p$, $\bar{K}^0n$, $\pi^0\Lambda$, $\pi^0\Sigma^0$, $\eta\Lambda$, $\eta\Sigma^0$, $\pi^+\Sigma^-$, $\pi^-\Sigma^+$, $K^+\Xi^-$, $K^0\Xi^0$ in coupled channels, paying special attention to the role of the $\Lambda(1405)$. In Sect.~ we describe the theoretical framework. The results are presented in Sect.~ followed by our conclusions. \section{Theoretical Framework} \subsection{Effective Lagrangians} At tree level, the process $\bar{\nu}_l p \rightarrow l^+ \phi B$, with $\phi$ and $B$ being the meson and baryon in the final state, proceeds as depicted in the diagrams of Fig.~. There are also baryon-pole terms (see Fig.~1 of Ref.~) which contribute predominantly to the $p$-wave state of the $\phi B$ system. Since our aim is to generate the $\Lambda(1405)$, which appears in $\phi B$ s-wave, we neglect these terms. All mechanisms in Fig.~ consist of a leptonic and a hadronic currents that interact via the exchange of a $W$ boson. The leptonic part is provided by the Standard Model Lagrangian \begin{equation} \mathcal{L}=-\frac{g}{2\sqrt{2}} \left[ \bar{\psi}_\nu \gamma_\mu (1-\gamma_5) \psi_l W^\mu + \bar{\psi}_l \gamma_\mu (1-\gamma_5) \psi_\nu {W^{\dag}}^\mu\right] \,, \end{equation} where $\psi_\nu$, $\psi_l$ and $W$ denote the neutrino, charged lepton and gauge boson $W$ fields, respectively; $g$ is the gauge coupling, related to the Fermi constant by $G_F=\sqrt{2} g^2/(8M_W^2) = 1.16639(1)\times 10^{-5}$ GeV$^{-2}$. The hadronic current is derived from chiral Lagrangians~ at leading order. As mentioned above, in this work we are only concerned about the $s$-wave contribution. In the meson sector, required for CT and MF diagrams, the lowest order SU(3) Lagrangian is given by \begin{equation} \mathcal{L}_\phi^{(2)} = \frac{F_0^2}{4}\langle D_\mu U (D^\mu U)^\dag\rangle + \frac{F_0^2}{4}\langle \chi U^\dag + U\chi^\dag\rangle, \end{equation} where $\langle \ldots \rangle$ stands for the trace in flavor space; $F_0$ is the pseudoscalar meson decay constant in the chiral limit. The quantity $\chi=2B_0\mathcal{M}$, with the quark-mass matrix $\mathcal{M} = \mathrm{diag}(m_u,m_d,m_s)$, represents the explicit breaking of chiral symmetry. The function $U = \mathrm{exp}\left(i \phi/F_0\right)$ is the SU(3) representation of the meson fields \begin{equation} \phi = \left(\begin{array}{ccc} \pi^0 + \frac{1}{\sqrt{3}} \eta & \sqrt{2} \pi^+ & \sqrt{2} K^+ \\ \sqrt{2} \pi^- & -\pi^0 + \frac{1}{\sqrt{3}} \eta & \sqrt{2} K^0 \\ \sqrt{2} K^- & \sqrt{2} \bar{K}^0 & -\frac{2}{\sqrt{3}} \eta \end{array}\right) \,, \end{equation} and its covariant derivative $D_\mu U$ can be written as \begin{equation} D_\mu U = \partial_\mu U - i r_\mu U + i U l_\mu, \end{equation} where $l_\mu$ and $r_\mu$ correspond to left- and right-handed currents. For the charged current weak interaction \begin{equation} r_\mu = 0,\quad l_\mu = \frac{g}{\sqrt{2}} ( W_\mu^\dagger T_+ + W_{\mu} T_-), \end{equation} with \begin{equation} T_+ = \left(\begin{array}{ccc} 0 & V_{ud} & V_{us} \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right), \quad T_- = \left(\begin{array}{ccc} 0 & 0 & 0 \\ V_{ud} & 0 & 0 \\ V_{us} & 0 & 0 \end{array}\right). \end{equation} Here, $V_{ij}$ are the relevant elements of the Cabibbo-Kobayashi-Maskawa matrix. Their magnitudes are $|V_{ud}|=\cos \theta_c =0.97425\pm0.00022$ and $|V_{us}|= \sin \theta_c =0.2252\pm0.0009$~, with $\theta_c$ the Cabibbo angle. The lowest order chiral effective Lagrangian describing the interaction between the octet of pseudoscalar mesons and the octet of baryons can be written as \begin{equation}~ \mathcal{L}_{\phi B}^{(1)} = \langle\bar{B}(i\slashed D-M_B)B\rangle+\frac{D}{2} \langle\bar{B}\gamma^{\mu}\gamma_5\{u_{\mu},B\}\rangle +\frac{F}{2} \langle\bar{B}\gamma^{\mu}\gamma_5[u_{\mu},B]\rangle \,, \end{equation} with the baryon fields arranged in the matrix \begin{equation} B = \left(\begin{array}{ccc} \frac{1}{\sqrt{2}} \Sigma^0 + \frac{1}{\sqrt{6}} \Lambda & \Sigma^+ & p \\ \Sigma^- & -\frac{1}{\sqrt{2}} \Sigma^0 + \frac{1}{\sqrt{6}} \Lambda & n \\ \Xi^- & \Xi^0 & -\frac{2}{\sqrt{6}} \Lambda \end{array}\right) \,; \end{equation} $M_B$ denotes the baryon octet mass in the chiral limit; $D=0.804$ and $F=0.463$ are the axial-vector coupling constants, which are determined from the baryon semi-leptonic decays~. The covariant derivative of the baryon field is defined as \begin{equation} D_{\mu}B=\partial_{\mu}B+[\Gamma_{\mu} , B], \end{equation} \begin{equation} \Gamma_{\mu}=\frac{1}{2}\left\{u^{\dag}(\partial_{\mu}-i r_{\mu})u + u(\partial_{\mu}-i l_{\mu})u^{\dag}\right\}, \end{equation} and $u_{\mu}$ is given by \begin{equation} u_{\mu}=i\left\{u^{\dag}(\partial_{\mu}-i r_{\mu})u-u(\partial_{\mu}-i l_{\mu})u^{\dag}\right\} , \end{equation} where $u=\sqrt{U}$. \subsection{Chiral Unitary Theory} As discussed in the introduction, the $\Lambda(1405)$ is dynamically generated by the interaction of $S=-1$ $s$-wave meson-baryon pairs in coupled channels. This can be achieved by solving the Bethe-Salpeter equation with the interaction potential provided by the chiral Lagrangian of Eq.~(). In the diagrams of Fig.~, the outgoing meson and baryon can interact producing the resonance. Therefore, one must consider the diagrams depicted in Fig.~. The solid square in the figures represents the different $T_{ij \rightarrow \phi B}$ amplitudes, where the pair of indices $i j = K^-p, \, \bar{K}^0n, \, \pi^0\Lambda,\, \pi^0\Sigma^0,\, \eta\Lambda,\, \eta\Sigma^0,\, \pi^+\Sigma^-,\, \pi^-\Sigma^+,\, K^+\Xi^-,\, K^0\Xi^0$ denote any of the ten allowed channels. Following the approach of Ref.~ for the strong interaction in the $S=-1$ sector, \begin{equation} T = V + VG T = [1-VG]^{-1}V \,, \end{equation} where the lowest-order interaction amplitude $V$, extracted from the lowest order chiral Lagrangian $\mathcal{L}^{(1)}_{\phi B}$, is given by \begin{equation} V_{ij} = -C_{ij}\frac{1}{4F_{\phi}^2}(k^0 + {k'}^0) \end{equation} after a nonrelativistic reduction. Here, $k^0$ and ${k'}^{0}$ are the energies of the incoming and outgoing mesons in the $\phi B$ center of mass (CM) frame; $F_0$ has been replaced by the average value of the physical decay constants $F_{\phi}=1.15 f_{\pi}$ with $f_\pi=93$ MeV as in Ref.~. The $10 \times 10$ matrix of coefficients $C_{ij}$ can be found in Table~1 of Ref.~. The meson-baryon loop function $G_{ij}$ is given by \begin{eqnarray} G_{ij} &=& i\int \frac{d^4 q}{(2\pi)^4} \frac{M_j}{E_j(\vec{q\ })} \frac{1}{k^0+p^0-q^0-E_j(\vec{q\ }) + i\epsilon} \frac{1}{q^2 - m_i^2 + i\epsilon}\,,\nonumber\\ &=& \int \frac{d^3 q}{(2\pi)^3} \frac{1}{2\omega_{j}(\vec{q\ })}\frac{M_j}{E_j(\vec{q\ })}\frac{1}{p^0+k^0 -\omega_i(\vec{q\ })-E_j(\vec{q\ })+i\epsilon}\,, \end{eqnarray} where $m_i$, $M_j$ are the physical meson and baryon masses of the $ij$ state while $\omega_i = (m_i^2+\vec{q}^{\,2})^{1/2}$, $E_j=(M_j^2+\vec{q}^{\,2})^{1/2}$ are the corresponding energies. It is a function of the CM energy $M_\mathrm{inv} = p^0+k^0$. In Ref.~, the loop function is regularized with a cutoff $q_\mathrm{max}=630$~MeV. \subsection{Cross section} The reaction under consideration is \begin{equation} \bar{\nu}_l(k_{\bar{\nu}}) + p(p) \rightarrow l^{+}(k_l) + \phi(k^{\prime}) + B(p^{\prime}), \end{equation} where $k_{\bar{\nu}}=(k^0_{\bar\nu},\vec{k}_{\bar\nu})$ [$k_l=(k^0_l,\vec{k}_l)$] is the 4-momentum of the incoming neutrino [outgoing charged lepton] while $p=(E_p,\vec{p})$, $p'=(E_B,\vec{p}\,')$ and $k'=(\omega_\phi,\vec{k}')$ denote the momenta of the initial proton, final baryon and final meson, in this order. Its cross section is given by \begin{equation} \sigma = \frac{2 M_p m_{\bar\nu}}{\lambda^{1/2}(s,m^2_{\bar\nu},M_p^2)} \int\frac{d^3k_l}{(2\pi)^3} \frac{m_l}{k^0_l}\int\frac{d^3k'}{(2\pi)^3}\frac{1}{2\omega_\phi} \int\frac{d^3p'}{(2\pi)^3}\frac{M_B}{E_B} (2\pi)^4\delta^4(p+k_{\bar{\nu}}-k_l-k'-p')\overline{\sum} |t|^2 \,, \end{equation} where $\lambda(x,y,z) = x^2 + y^2 + z^2 - 2 x y - 2 x z - 2 y z$ and $s=(p+k_{\bar\nu})^2$; $\overline{\sum}$ denotes the sum over final state polarizations and average over the initial ones. It is convenient to perform the integrals over $\vec{p}\,'$ and $\vec{k}'$ in the $\phi B$ CM frame, taking advantage of the fact that the amplitude is projected onto the $s$-wave state of the $\phi B$ pair. The last integration over $\vec{k}_l$ is carried out in the global ($\bar\nu p$) CM frame. We obtain \begin{equation} \sigma = \frac{2}{(2\pi)^3}\frac{m_{\bar{\nu}} m_l M_p M_B}{\sqrt{s}(s-M_p^2)} \int_{m_\phi + M_B}^{\sqrt{s}-m_l} dM_\mathrm{inv}\int_{-1}^{+1} d\cos\theta |\vec{k}_l|_{\bar{\nu} p} |\vec{k}'|_{\phi B}\overline{\sum} |t|^2 \,, \end{equation} where $\theta$ is the angle between $\vec{k}_l$ and $\vec{k}_{\bar\nu}$ in the $\bar\nu p$ CM frame. In Eq.~() \begin{equation} |\vec{k}_l|_{\bar{\nu}_l p} = \frac{\lambda^{1/2}(s, m_l^2, M_\mathrm{inv}^2)}{2\sqrt{s}}, \quad |\vec{k}'|_{\phi B} = \frac{\lambda^{1/2}(M_\mathrm{inv}^2, m_\phi^2, M_B^2)}{2M_\mathrm{inv}} \end{equation} are the charged-lepton momentum in the $\bar\nu p$ CM frame and the meson momentum in the $\phi B$ CM frame, respectively. \subsection{Invariant amplitude} In the $(k_l-k_{\bar\nu})^2 \equiv q^2 \ll M_W^2$ limit, the amplitude can be cast as \begin{equation} -it = 2G_F V_{us} L^{\mu} H_{\mu} \,, \end{equation} where the leptonic current is \begin{equation} L^{\mu}=\bar{v}(k_{\bar\nu})\gamma^{\mu}(1-\gamma_5)v(k_l) \,, \end{equation} while the hadronic current \begin{equation} H_{\mu} = \bar{u}(p')\Gamma_\mu u(p) \end{equation} is determined by the sum of the following contributions \begin{itemize} \item KP (vector) \begin{equation} \Gamma_{\mu}^\mathrm{KP} = -\frac{1}{2}F_{\phi}\frac{q_{\mu}}{q^2-m_{K^-}^2+i\epsilon} T_{K^-p\rightarrow \phi B}\,. \end{equation} Note that in Fig.~~(a), the sum over the intermediate states $\phi' B'$ produces the $K^- p \rightarrow \phi B$ $t$-matrix element by virtue of Eq.~. \item CT (vector plus axial) \begin{eqnarray} \Gamma_{\mu}^{\mathrm{CT}(V)} &=& -\frac{1}{4F_{\phi}}\left[C_{\phi B}^{(V)} \gamma_\mu + \sum\limits_{\phi' B'} C_{\phi' B'}^{(V)} \gamma_{\mu} G_{\phi'B'} T_{\phi'B'\rightarrow \phi B}\right],\\ \Gamma_{\mu}^{\mathrm{CT}(A)} &=& -\frac{1}{4F_{\phi}}\left[C_{\phi B}^{(A)} \gamma_\mu\gamma^5 + \sum\limits_{\phi' B'} C_{\phi' B'}^{(A)} \gamma_{\mu} \gamma^5 G_{\phi'B'} T_{\phi'B'\rightarrow \phi B}\right]. \end{eqnarray} The coefficients $C_{\phi B}^{(V)}$ and $C_{\phi B}^{(A)}$ are tabulated in Table~ and Table~, respectively. The loop function is given by \begin{equation} G_{\phi'B'} = i\int\frac{d^4l}{(2\pi)^4}\frac{1}{l^2-m_{\phi'}^2+ i\epsilon} \frac{1}{\slashed p + \slashed q-\slashed l - M_{B'}+i\epsilon}. \end{equation} \item MF (axial) \begin{equation} \Gamma_{\mu}^\mathrm{MF} = \frac{1}{4\sqrt{2} F_\phi}\left[\sum\limits_{\phi''} C_{\phi''\phi}C_{\phi''B}\frac{(2k'-q)_{\mu}(k'-q)_{\nu}\gamma^{\nu}\gamma^5} {(k'-q)^2-m_{\phi''}^2+i\epsilon} + \sum\limits_{\phi'\phi''B'} C_{\phi''\phi'}C_{\phi''B'} G_{\phi'\phi''B'}^\mu T_{\phi'B'\rightarrow \phi B}\right], \end{equation} where $\phi''$ denotes the internal meson in the tree level diagram (c) of Fig.~. In most cases, only one type of meson can be exchanged but it happens that both $\pi^0$ and $\eta$ are allowed intermediate states. The $G^{\mu}_{\phi'\phi''B'}$ function is given by \begin{equation} G_{\phi'\phi''B'}^\mu = i\int\frac{d^4 l}{(2\pi)^4}~ (2l-q)^\mu ~(l-q)^\nu ~\gamma_{\nu}\gamma^5 \frac{1}{l^2-m_{\phi'}^2+i\epsilon}\frac{1}{(l-q)^2-m_{\phi''}^2+i\epsilon} \frac{1}{\slashed p-\slashed l+ \slashed q - M_{B'} + i\epsilon}. \end{equation} Finally, coefficients $C_{\phi_1\phi_2}$ and $C_{\phi B}$ are tabulated in Table~ and Table~, respectively. \end{itemize} \begin{table}[h!] \centering \setlength{\tabcolsep}{20pt} \begin{tabular}{cccccc} \hline\hline $C_{\phi B}^{(V)}$ & $p$ & $n$ & $\Lambda$ & $\Sigma^0$ & $\Sigma^+$ \\ \hline $K^-$ & $2$ & 0 & 0 & 0 & 0 \\ $\bar{K}^0$ & 0 & $1$ & 0 & 0 & 0 \\ $\pi^0$ & 0 & 0 & $\frac{\sqrt{3}}{2}$ & $\frac{1}{2}$ & 0 \\ $\eta$ & 0 & 0 & $\frac{3}{2}$ & $\frac{\sqrt{3}}{2}$ & 0 \\ $\pi^-$ & 0 & 0 & 0 & 0 & $1$ \\ \hline\hline \end{tabular} \caption{Coefficients $C_{\phi B}^{(V)}$ appearing in the CT contribution to the hadronic current [Eq.~()].} \end{table} \begin{table}[h!] \centering \setlength{\tabcolsep}{6pt} \begin{tabular}{cccccc} \hline\hline $C_{\phi B}^{(A)}$ & $p$ & $n$ & $\Lambda$ & $\Sigma^0$ & $\Sigma^+$ \\ \hline $K^-$ & $-2F$ & 0 & 0 & 0 & 0 \\ $\bar{K}^0$ & 0 & $-(D+F)$ & 0 & 0 & 0 \\ $\pi^0$ & 0 & 0 & $-\frac{1}{2\sqrt{3}}(D+3F)$ & $\frac{1}{2}(D-F)$ & 0 \\ $\eta$ & 0 & 0 & $-\frac{1}{2}(D+3F)$ & $\frac{\sqrt{3}}{2}(D-F)$ & 0 \\ $\pi^-$ & 0 & 0 & 0 & 0 & $D-F$ \\ \hline\hline \end{tabular} \caption{Coefficients $C_{\phi B}^{(A)}$ appearing in the CT contribution to the hadronic current [Eq.~()].} \end{table} \begin{table}[h!] \centering \setlength{\tabcolsep}{20pt} \begin{tabular}{*{6}{c}} \hline\hline $C_{\phi_1\phi_2}$ & $K^-$ & $\bar{K}^0$ & $\pi^0$ & $\eta$ & $\pi^-$ \\ \hline $\pi^0$ & $-\frac{1}{\sqrt{2}}$ & 0 & 0 & 0 & 0 \\ $\eta$ & $-\sqrt{\frac{3}{2}}$ & 0 & 0 & 0 & 0 \\ $\pi^+$ & 0 & $-1$ & 0 & 0 & 0 \\ $K^+$ & 0 & 0 & $\frac{1}{\sqrt{2}}$ & $\sqrt{\frac{3}{2}}$ & 0 \\ $K^0$ & 0 & 0 & 0 & 0 & $1$ \\ \hline\hline \end{tabular} \caption{Coefficients $C_{\phi_1\phi_2}$ appearing in the MF contribution to the hadronic current [Eq.~()].} \end{table} \begin{table}[h!] \centering \setlength{\tabcolsep}{5pt} \begin{tabular}{cccccc} \hline\hline $C_{\phi B}$ & $p$ & $n$ & $\Lambda$ & $\Sigma^0$ & $\Sigma^+$ \\ \hline $\pi^0$ & $D+F$ & 0 & 0 & 0 & 0 \\ $\eta$ & -$\frac{1}{\sqrt{3}}(D-3F)$ & 0 & 0 & 0 & 0 \\ $\pi^+$ & 0 & $\sqrt{2}(D+F)$ & 0 & 0 & 0 \\ $K^+$ & 0 & 0 & $-\frac{1}{\sqrt{3}}(D+3F)$ & $D-F$ & 0 \\ $K^0$ & 0 & 0 & 0 & 0 & $\sqrt{2}(D-F)$ \\ \hline\hline \end{tabular} \caption{Coefficients $C_{\phi B}$ appearing in the MF contribution to the hadronic current [Eq.~()].} \end{table} The hadronic current presented above does not take into account the $q^2$ dependence of the weak interaction vertices, which is poorly known. Following Ref.~, we have parametrized this dependence with a global dipole form factor \begin{equation} F(q^2) = \left(1 - \frac{q^2}{M_F^2} \right)^{-2} \end{equation} that multiplies all the terms in $H_\mu$. Up to SU(3) breaking effects, the value of the axial mass $M_F$ should be similar to the one in electromagnetic and axial nucleon form factors. Therefore, as in Refs.~ we have adopted $M_F \simeq 1$~GeV, accepting an uncertainty of around 10~\ \subsection{Non-relativistic reduction of the invariant amplitude} Because we only focus on the small momenta of the $\phi B$ components creating the $\Lambda(1405)$, we can perform a non relativistic reduction, which was also used in the description of the $\phi B$ amplitude in coupled channels of Ref.~. For the CT we get \begin{eqnarray} -it^{\mathrm{CT}(V)} &=& -\frac{1}{4F_\phi}(2G_FV_{us})L^{0} \left[ C_{\phi B}^{(V)} + \sum\limits_{\phi' B'} C_{\phi' B'}^{(V)} G'_{\phi'B'} T_{\phi'B'\rightarrow \phi B}\right],\nonumber\\ -it^{\mathrm{CT}(A)} &=& +\frac{1}{4F_{\phi}}(2G_FV_{us}) (\vec{L}\cdot \vec\sigma) \left[C_{\phi B}^{(A)} + \sum\limits_{\phi' B'} C_{\phi' B'}^{(A)} G'_{\phi'B'} T_{\phi'B'\rightarrow \phi B}\right], \end{eqnarray} where the loop function, after removing the baryon negative energy part, becomes \begin{eqnarray} G'_{\phi' B'} &=& \int\frac{d^3 l}{(2\pi)^3} \frac{1}{2\omega_{\phi'}(\vec{l})}\frac{M_{B'}}{E_{B'}(\vec{l})} \frac{1}{M_{\mathrm{inv}}-\omega_{\phi'}(\vec{l})-E_{B'}(\vec{l}) + i\epsilon} \,. \end{eqnarray} After the non relativistic reduction, the MF contributions can be written as \begin{eqnarray} -it^\mathrm{MF} &=& \frac{1}{4\sqrt{2}F_{\phi}}(2G_FV_{us})\left\{\sum\limits_{\phi''}C_{\phi''\phi}~C_{\phi''B}~\vec{\sigma}\cdot\vec{q}~ \frac{L^0(2k'-q)^0 +\vec{L}\cdot\vec{q}}{(k'-q)^2-m_{\phi''}^2+i\epsilon} \right.\nonumber\\ && \left. + \sum\limits_{\phi'\phi''B'}C_{\phi''\phi'}C_{\phi''B'} \left[\vec{L}\cdot\vec{\sigma} G^{(1)}_{\phi'\phi''B'} + (\vec{L}\cdot\vec{q}) (\vec{\sigma}\cdot\vec{q}) G_{\phi'\phi''B'}^{(2)}\right]\right\} \,, \end{eqnarray} where the loop functions are \begin{eqnarray} G_{\phi\phi' B'}^{(1)} &=& \int\frac{d^3l}{(2\pi)^3} \frac{1}{\omega_{\phi'}(\vec{l})\omega_\phi(\vec{l}-\tilde{\vec{q}})}\frac{M_{B'}}{E_{B'}(\vec{l})} \frac{\vec{l}^2}{3} \left\{\left[\omega_\phi(\vec{l}-\tilde{\vec{q}}) + \omega_{\phi'}(\vec{l})\right]^2 \right.\nonumber\\ && \left.+ \left[\omega_\phi(\vec{l}-\tilde{\vec{q}}) + \omega_{\phi'}(\vec{l})\right]\left[E_{B'}(\vec{l})-\tilde{p}^0\right] -\tilde{q}^0\omega_{\phi'}(\vec{l}) \right\}\nonumber\\ &&\times \frac{1}{M_\mathrm{inv}-E_{B'}(\vec{l}) - \omega_{\phi'}(\vec{l}) + i\epsilon} \frac{1}{\tilde{p}^0-E_{B'}(\vec{l})-\omega_\phi(\vec{l}-\tilde{\vec{q}}) + i\epsilon } \nonumber\\ && \times\frac{1}{\tilde{q}^0+\omega_\phi(\vec{l}-\tilde{\vec{q}})+\omega_{\phi'}(\vec{l})-i\epsilon} \frac{1}{\omega_{\phi'}(\vec{l})-\tilde{q}^0 + \omega_\phi(\vec{l}-\tilde{\vec{q}}) - i\epsilon}\,, \end{eqnarray} and \begin{eqnarray} G_{\phi\phi' B'}^{(2)} &=& \int\frac{d^3l}{(2\pi)^3} \frac{1}{2\omega_{\phi'}(\vec{l})\omega_{\phi}(\vec{l}-\tilde{\vec{q}})}\frac{M_{B'}}{E_{B'}(\vec{l})} \left\{\left[\omega_{\phi}(\vec{l}-\tilde{\vec{q}}) + \omega_{\phi'}(\vec{l})\right]^2 \right.\nonumber\\ && \left. + \left[\omega_{\phi}(\vec{l}-\tilde{\vec{q}}) + \omega_{\phi'}(\vec{l})\right]\left[E_{B'}(\vec{l})-\tilde{p}^0\right] -\tilde{q}^0\omega_{\phi'}(\vec{l}) \right\}\nonumber\\ &&\times \frac{1}{M_\mathrm{inv}-E_{B'}(\vec{l}) - \omega_{\phi'}(\vec{l}) + i\epsilon} \frac{1}{\tilde{p}^0-E_{B'}(\vec{l})-\omega_{\phi}(\vec{l}-\tilde{\vec{q}}) + i\epsilon } \nonumber\\ &&\times\frac{1}{\tilde{q}^0+\omega_{\phi}(\vec{l}-\tilde{\vec{q}})+\omega_{\phi'}(\vec{l})-i\epsilon} \frac{1}{\omega_{\phi'}(\vec{l})-\tilde{q}^0 + \omega_\phi(\vec{l}-\tilde{\vec{q}}) - i\epsilon} \,. \end{eqnarray} The quantities with tilde are defined in the $\phi B$ CM frame. \section{Results}~ Throughout this section, the results are presented for the muon flavor $l = \mu$. The $\Lambda(1405)$ can be observed in the invariant mass distribution of $\pi\Sigma$ pairs that has its threshold below the peak of the $\Lambda(1405)$ states. The cleanest signal for $I=0$ $\Lambda(1405)$ production appears in the $\pi^0\Sigma^0$ channel because $I=1$ is not allowed. In Fig.~, we show $d\sigma/dM_\mathrm{inv}$ for $\pi^0\Sigma^0$ production at three different laboratory energies, $E_{\bar\nu}=900$, $1100$, and $1300$~MeV. We can clearly see the resonant shape of the $\Lambda(1405)$ at all the energies. Note that, in spite of the two poles, there is a single peak. This is common to all the reactions, with the exception of electroproduction~, where the data are still relatively poor. Only the different weight of the two poles makes the peak appear at different energies in different processes. In the present case the distribution peaks around $1420$ MeV indicating that there is more weight from the pole at $1420$ MeV or, in other words, that the $\Lambda(1405)$ production induced by the $K^-p$ is dominant. To gain further insight into the interplay of the two poles of the $\Lambda(1405)$ resonance in this reaction, we have looked at the line shapes of the double differential cross section $d^2\sigma/\left(dM_\mathrm{inv} d\cos{\theta}\right)$ for different values of the $\theta$ angle between the initial $\bar{\nu}_\mu$ and the final $\mu^+$ in the $\bar{\nu} p$ CM frame (Fig.~). When $\theta$ increases, so does $|q^2|$, and the form factor causes a reduction in the cross section. To compare the shapes we have normalized all curves to the same area by multiplying the cross section at $\cos{\theta} = 0 (-1)$ by 3.4(14). In the backward direction, the distribution clearly resembles a single Breit-Wigner with a mass and a width remarkably close to the values of the heavier pole of the $\Lambda(1405)$. It is this pole that appears dominant at this kinematics. As $\theta$ decreases, the presence of the lighter state becomes more evident with larger strength accumulating below the peak, which is shifted towards smaller invariant masses. The line shape becomes asymmetric but the second state never shows up as a peak in the cross section. It is also very interesting to consider $d\sigma/dM_\mathrm{inv}$ for the three charged channels $\pi^0\Sigma^0$, $\pi^+\Sigma^-$ and $\pi^-\Sigma^+$. This is shown in Fig.~. The peak position for the different reactions is slightly shifted, but the largest differences are present below the maxima. This is due to the contribution of an $I=1$ amplitude which adds constructively or destructively depending on the channel~. It was also shown in Ref.~ that $\Lambda(1405)$ photoproduction data hint to a possible $I=1$ state around $1400$ MeV, which appears in some approaches~ but is at a border line in others~. In the work of Refs.~, the existence of such $I=1$ state is claimed from the study of the $K^-p\rightarrow \Lambda\pi^-\pi^+$ reaction. The large differences seen in the cross sections for the three $\pi\Sigma$ channels in the present reaction indicate that they are indeed rather sensitive to the $I=1$ amplitude and, thus, there is a potential for the extraction of information on the possible $I=1$ state. In Fig.~, we show now the integrated cross sections for $\pi^0\Sigma^0$, $\pi^-\Sigma^+$, and $\pi^+\Sigma^-$ production. We observe a steady growth of the cross sections with the antineutrino energy. These cross sections are largely driven by the $\Lambda(1405)$ resonance. Indeed, in Fig.~, both tree level and full model cross sections are shown. We observe that the contribution of the meson-baryon rescattering has a drastic effect in the results. The case of the $\pi^+\Sigma^-$ channel is the most spectacular because the tree level contribution is exactly zero. We have also investigated the $\bar K$-nucleon production reactions. Note that in this case the threshold energies, $\sqrt{s}=m_{K^-} + M_p=1430$~MeV and $m_{\bar K^0} + M_n =1437$~MeV, are already above the $\Lambda(1405)$ peak. Thus, we do not plot $d\sigma/dM_\mathrm{inv}$ in this case and show only the integrated cross section as a function of energy. These are shown in Fig.~ for $K^-p$ and in Fig.~ for $\bar{K}^0n$. As can be seen in the right panels of Figs.~,, unlike the $\pi \Sigma$ production case, the cross section is not increased by the resonance. On the contrary, the fast fall down of $d\sigma/dM_\mathrm{inv}$ close to the $K^-p$ threshold, seen in Fig.~ for $\pi\Sigma$, reflects the similar trend of the $t$ matrix which is common to all the channels. This affects the $\bar K$-nucleon production cross sections, most noticeably for $\bar{K}^0n$, the channel with a larger threshold. These unitarization effects were absent in the calculations reported in Ref.~. There are other differences between the present study and the one of Ref.~. First, here we have used the average $F_\phi = 1.15 f_\pi$, for consistency with the value taken in the study of $\phi B$ scattering~ (see Sec.~), instead of $F_\phi = f_\pi$ in Ref.~. This leads to little smaller cross section with respect to those of Ref.~. Furthermore, the $p$-wave contributions considered in Ref.~ but not here make the cross sections bigger as one departs from threshold. Finally, the non relativistic approximation becomes poorer for the higher energy and momentum transfers that can be probed as the reaction energy increases. As an example, the CT contribution here is about 30\ In the $K^-p$ channel, the largest contribution arises from the CT mechanism (left panel of Fig.~), in line with Fig.~3 of Ref.~. In the $\bar{K}^0 n$ channel, instead, the MF contribution becomes increasingly larger than the CT above $E_{\bar{\nu}}=1200$~MeV (left panel of Fig.~), in variance with Fig.~5 of Ref.~. Nevertheless, it should be mentioned that our predictions for KP, CT and MF terms converge to those of Ref.~ in the heavy-nucleon limit. \subsection{$\mathbf{\Lambda(1405)}$ production at MINER$\mathbf{\nu}$A} One of the goals of the MINER$\nu$A experiment is to study weak strangeness production~. It is therefore important to obtain the number of events in which the $\Lambda(1405)$ resonance is primarily produced during the antineutrino run. Let us consider the process $\bar{\nu}_\mu p \rightarrow \mu^+ \pi \Sigma$. The number of events for a given invariant mas of the $\pi \Sigma$ pair is \begin{equation} \frac{d N}{d M_{\mathrm{inv}}} = N_{\mathrm{POT}} f M N_A \int dE_{\bar\nu} \phi(E_{\bar\nu}) \frac{d \sigma_{\pi \Sigma}}{d M_{\mathrm{inv}}}(E_{\bar\nu}) \,. \end{equation} The differential cross section is averaged over the antineutrino flux $\phi(E_{\bar\nu})$. The flux prediction, in units of $\bar\nu/cm^2/\mathrm{POT}$, for the low-energy configuration is taken from Table V of Ref.~. The present estimate corresponds to a number of protons on target of $N_{\mathrm{POT}} = 2.01 \times 10^{20}$ in $\bar\nu$ mode, neglecting the small $\bar\nu_e$ component in the beam of muon antineutrinos. Although the MINER$\nu$A detector is made of different materials, here we consider only the scintillator (CH). In this case the proton fraction $f=(1+6)/(1+12)$. One should recall that $\pi \Sigma$ pairs can also be produced on neutrons but, in this case, the pair has negative charge, not leading to $\Lambda(1405)$ excitation. The scintillator mass is $M = 0.45 M_1 + 0.55 M_2$, with $M_1 =2.84 \times 10^6$ and $M_2 =5.47 \times 10^6$ grams, to take into account that 45\ The event distributions for $\pi^0 \Sigma^0$, $\pi^- \Sigma^+$ and $\pi^+ \Sigma^-$ pairs and their sum, in the region of the $\Lambda(1405)$ resonance, are shown in Fig.~. At $q^2=0$, the largest invariant mass shown in Fig.~, corresponds to a still moderate $\tilde{q}^0 = 456$~MeV, regardless of the antineutrino energy which can be high at MINER$\nu$A ($\langle E_{\bar\nu} \rangle \sim 3.5$~GeV). For negative values of $q^2$, the largest $\tilde{q}^0$ can be larger, and even more so $|\tilde{\vec{q}}|$. On the other hand, the cross section for these $q^2$ is suppressed by poorly known vector and axial form factors, which have been accounted here with the global form factor of Eq.~. The uncertainty in the number of events at non-zero $q^2$, accounted by a 10\ By integrating the distributions in Fig.~, one finds the following numbers of events: $N_{\pi^0 \Sigma^0} = 612^{+120}_{-112}$, $N_{\pi^+ \Sigma^-} = 517^{+100}_{-94}$, $N_{\pi^- \Sigma^+} = 838^{+163}_{-153}$. All in all, we predict about 2000 $\pi \Sigma$ pairs coming predominantly from $\Lambda(1405)$ decay. Modern neutrino experiments, including MINER$\nu$A, have detectors with nuclear targets. Nuclear effects, not considered in the present study, play an important role. It has been shown that strangeness can be abundantly produced in secondary collisions~. The events predicted above correspond to $\Lambda(1405)$ excitation in primary $\bar\nu N$ collisions but the actual signal will be different. The invariant mass of the outgoing $\pi \Sigma$ gets distorted by final state interactions with other nucleons in the nucleus~; the composition of the final state can change because of pion absorption and other inelastic processes like $\pi \, N \rightarrow K \, Y$, $\Sigma \, N \rightarrow N \, N \, \bar K$ and others. In the same way, the $\Lambda(1405)$ can be produced in secondary $\bar K N$ scattering. This dynamics requires a more detailed investigation to find specific indications of $\Lambda(1405)$ production in $\bar\nu$-nucleus collisions. Yet, as it happens in photonuclear reactions in nuclei, even if secondary collisions distort the resonance signal, there is still a sizeable fraction of events not affected by them. These events mostly come from primary interactions taking place in the back of the nucleus with respect to the direction of the three-momentum transfer $\vec{q}$ in the Laboratory frame. Therefore, a signal from the primary collisions can be observed in these reactions. This is the case in $\Delta(1232)$~ and $\omega$~ photoproduction. \section{Conclusions}~ We have studied $\Lambda(1405)$ production induced by antineutrinos, the first calculation of this sort. For this purpose we have combined elements of chiral perturbation theory in the presence of weak external fields with unitarization techniques in coupled channels. The $\Lambda(1405)$, consisting actually of two states, is generated through the multiple scattering of meson-baryon coupled channels with a kernel provided by the chiral Lagrangians. It can only be observed in the $\pi\Sigma$ final state, most cleanly in the $\pi^0\Sigma^0$ channel which has only $I=0$. As in most reactions, the $\Lambda(1405)$ appears as a single highly asymmetric peak in the $\pi \Sigma$ invariant mass distribution. The line shapes at different angles between the incoming $\bar\nu$ and the outgoing lepton in the reaction CM frame indicate that the process at backward angles is dominated by a state with mass and width of around 1420 and 40~MeV, respectively. As the angle decreases, the lighter states becomes increasingly more important. The $\pi^+\Sigma^-$ and $\pi^-\Sigma^+$ channels also contain an $I=1$ amplitude, where a possible resonance might be present according to some studies. This amplitude is responsible for large differences in the shapes of the $\pi \Sigma$ invariant mass distributions below the maximum for the three charge channels. Therefore, a combined study of $\pi^0\Sigma^0$, $\pi^+\Sigma^-$ and $\pi^-\Sigma^+$ production induced by antineutrinos could provide useful information about this hypothetical $I=1$ state. We have also evaluated the integrated cross sections for $\bar{\nu}_\mu p \rightarrow \mu^+ \pi \Sigma$ as a function of the antineutrino energy. These are much larger than the corresponding tree level results due to the $\Lambda(1405)$ excitation. We should note that the tree level is relatively more important for the $\bar{K} N$ final state because the latter is above the $\Lambda(1405)$. In this case, unitarization does not cause an enhancement of the cross section. One rather observes a reduction in the $\bar{K}^0n$ channel, which has the largest threshold. We have obtained that the number of events in which the $\Lambda(1405)$ is excited in primary $\bar\nu_\mu p$ collisions at the scintillator detector of the MINER$\nu$A experiment, in the antineutrino run, is of the order of 2000. It is large enough to conclude that $\Lambda(1405)$ production has a sizable impact in the scattering dynamics leading to antineutrino detection, and should be taken into account in future evolutions of neutrino event generators. Several open questions in the physics of (anti)neutrino interactions with matter call for new measurements of (anti)neutrino cross sections on proton and hydrogen targets~. Such experiments with antineutrinos would also provide a more complete understanding of the $\Lambda(1405)$ properties. \begin{acknowledgments} LAR wishes to thank J. Morfin for a useful discussion about the flux and event rate at the MINER$\nu$A experiment. The work of LAR has been partially supported by the U.S. Department of Energy, Office of Science, Office of High Energy Physics, through the Fermilab Intensity Frontier Fellows Program. He gratefully acknowledges the hospitality during his stay at Fermilab. X.-L.R thanks Prof. Jie Meng and Prof. Li-Sheng Geng for useful discussions. He acknowledges support from the Innovation Foundation of Beihang University for Ph.D. Graduates, the National Natural Science Foundation of China under Grant No. 11375024 and a fellowship from the China Scholarship Council. This work has been partly supported by the Spanish Ministerio de Econom\'ia y Competitividad and European FEDER funds under the contract number FIS2011-28853-C02-01 and FIS2011-28853-C02-02, and the Generalitat Valenciana in the program Prometeo II, 2014/068. We acknowledge the support of European Community-Research Infrastructure Integrating Activity Study of Strongly Interacting Matter (acronym HadronPhysics3, Grant Agreement n. 283286) under the Seventh Framework Program of EU. \end{acknowledgments} \begin{thebibliography}{99} \bibitem{dalitz} R.~H.~Dalitz, S.~F.~Tuan, Annals Phys.\ {\bf 10}, 307-351 (1960). \bibitem{dalitzdos} R.~H.~Dalitz, T.~C.~Wong, G.~Rajasekaran, Phys.\ Rev.\ {\bf 153}, 1617-1623 (1967). \bibitem{tonythomas} E.~A.~Veit, B.~K.~Jennings, R.~C.~Barrett, A.~W.~Thomas, Phys.\ Lett.\ {\bf B137}, 415 (1984). \bibitem{weise} N.~Kaiser, P.~B.~Siegel and W.~Weise, Nucl.\ Phys.\ A {\bf 594}, 325 (1995). \bibitem{angels} E.~Oset and A.~Ramos, Nucl.\ Phys.\ A {\bf 635}, 99 (1998). \bibitem{ollerulf} J.~A.~Oller and U.~G.~Meissner, Phys.\ Lett.\ B {\bf 500}, 263 (2001). \bibitem{hyodo} T.~Hyodo, S.~I.~Nam, D.~Jido and A.~Hosaka, Phys.\ Rev.\ C {\bf 68}, 018201 (2003). \bibitem{cola} D.~Jido, J.~A.~Oller, E.~Oset, A.~Ramos, U.~G.~Meissner, Nucl.\ Phys.\ {\bf A725}, 181-200 (2003). \bibitem{carmen} C.~Garcia-Recio, J.~Nieves, E.~Ruiz Arriola and M.~J.~Vicente Vacas, Phys.\ Rev.\ D {\bf 67}, 076009 (2003). \bibitem{review} J.~A.~Oller, E.~Oset and A.~Ramos, Prog.\ Part.\ Nucl.\ Phys.\ {\bf 45}, 157 (2000) [hep-ph/0002193]. \bibitem{borasoyweise} B.~Borasoy, R.~Nissler and W.~Weise, Eur.\ Phys.\ J.\ A {\bf 25}, 79 (2005). \bibitem{borasoymeissner} B.~Borasoy, U.~G.~Meissner and R.~Nissler, Phys.\ Rev.\ C {\bf 74}, 055201 (2006). \bibitem{hyodoweise} Y.~Ikeda, T.~Hyodo and W.~Weise, Nucl.\ Phys.\ A {\bf 881}, 98 (2012). \bibitem{Feijoo:2015yja} A.~Feijoo, V.~K.~Magas and A.~Ramos, arXiv:1502.07956 [nucl-th]. \bibitem{kanchan} K.~P.~Khemchandani, A.~Martinez Torres, H.~Kaneko, H.~Nagahiro and A.~Hosaka, Phys.\ Rev.\ D {\bf 84}, 094018 (2011). \bibitem{hyodorev} T.~Hyodo and D.~Jido, Prog.\ Part.\ Nucl.\ Phys.\ {\bf 67}, 55 (2012). \bibitem{ollerguo} Z.~-H.~Guo and J.~A.~Oller, Phys.\ Rev.\ C {\bf 87}, 035202 (2013). \bibitem{thomas} D.~W.~Thomas, A.~Engler, H.~E.~Fisk and R.~W.~Kraemer, Nucl.\ Phys.\ B {\bf 56}, 15 (1973). \bibitem{hemingway} R.~J.~Hemingway, Nucl.\ Phys.\ B {\bf 253}, 742 (1985). \bibitem{osethyodo} T.~Hyodo, A.~Hosaka, E.~Oset, A.~Ramos and M.~J.~Vicente Vacas, Phys.\ Rev.\ C {\bf 68}, 065203 (2003). \bibitem{osetramosjc} J.~C.~Nacher, E.~Oset, H.~Toki and A.~Ramos, Phys.\ Lett.\ B {\bf 461}, 299 (1999). \bibitem{nefkens} S.~Prakhov {\it et al.} [Crystal Ball Collaboration], Phys.\ Rev.\ C {\bf 70}, 034605 (2004). \bibitem{angelsmagas} V.~K.~Magas, E.~Oset and A.~Ramos, Phys.\ Rev.\ Lett.\ {\bf 95}, 052301 (2005). \bibitem{braun} O.~Braun, H.~J.~Grimm, V.~Hepp, H.~Strobele, C.~Thol, T.~J.~Thouw, D.~Capps and F.~Gandini {\it et al.}, Nucl.\ Phys.\ B {\bf 129}, 1 (1977). \bibitem{sekihara} D.~Jido, E.~Oset and T.~Sekihara, Eur.\ Phys.\ J.\ A {\bf 42}, 257 (2009). \bibitem{miyagawa} K.~Miyagawa and J.~Haidenbauer, Phys.\ Rev.\ C {\bf 85}, 065201 (2012). \bibitem{sekidos} D.~Jido, E.~Oset and T.~Sekihara, Eur.\ Phys.\ J.\ A {\bf 49}, 95 (2013). \bibitem{sidharta} M.~Bazzi, G.~Beer, L.~Bombelli, A.~M.~Bragadireanu, M.~Cargnelli, G.~Corradi, C.~Curceanu (Petrascu) and A.~d'Uffizi {\it et al.}, Phys.\ Lett.\ B {\bf 704}, 113 (2011). \bibitem{moriya} K.~Moriya {\it et al.} [CLAS Collaboration], Phys.\ Rev.\ C {\bf 87}, 035206 (2013). \bibitem{niiyama} M.~Niiyama, H.~Fujimura, D.~S.~Ahn, J.~K.~Ahn, S.~Ajimura, H.~C.~Bhang, T.~H.~Chang and W.~C.~Chang {\it et al.}, Phys.\ Rev.\ C {\bf 78}, 035202 (2008). \bibitem{roca} L.~Roca and E.~Oset, Phys.\ Rev.\ C {\bf 87}, 055201 (2013). \bibitem{rocados} L.~Roca and E.~Oset, Phys.\ Rev.\ C {\bf 88}, 055206 (2013). \bibitem{mai} M.~Mai and U.~G.~Meißner, Eur.\ Phys.\ J.\ A {\bf 51}, no. 3, 30 (2015) [arXiv:1411.7884 [hep-ph]]. \bibitem{anke} I.~Zychor, M.~Buscher, M.~Hartmann, A.~Kacharava, I.~Keshelashvili, A.~Khoukaz, V.~Kleber and V.~Koptev {\it et al.}, Phys.\ Lett.\ B {\bf 660}, 167 (2008). \bibitem{genganke} L.~S.~Geng and E.~Oset, Eur.\ Phys.\ J.\ A {\bf 34}, 405 (2007). \bibitem{fabbietti} G.~Agakishiev {\it et al.} [HADES Collaboration], Phys.\ Rev.\ C {\bf 87}, 025201 (2013). \bibitem{fabbidos} J.~Siebenson and L.~Fabbietti, Phys.\ Rev.\ C {\bf 88}, 055201 (2013). \bibitem{electro} H.~Y.~Lu {\it et al.} [CLAS Collaboration], Phys.\ Rev.\ C {\bf 88}, 045202 (2013). \bibitem{leinweber} B.~J.~Menadue, W.~Kamleh, D.~B.~Leinweber and M.~S.~Mahbub, Phys.\ Rev.\ Lett.\ {\bf 108}, 112001 (2012) [arXiv:1109.6716 [hep-lat]]. \bibitem{lang} G.~P.~Engel {\it et al.} [BGR [Bern-Graz-Regensburg] Collaboration], Phys.\ Rev.\ D {\bf 87}, no. 3, 034502 (2013) [arXiv:1212.2032 [hep-lat]]. \bibitem{Hall:2014uca} J.~M.~M.~Hall, W.~Kamleh, D.~B.~Leinweber, B.~J.~Menadue, B.~J.~Owen, A.~W.~Thomas and R.~D.~Young, Phys.\ Rev.\ Lett., in print [arXiv:1411.3402 [hep-lat]]. \bibitem{ross} J.~M.~M.~Hall, W.~Kamleh, D.~B.~Leinweber, B.~J.~Menadue, B.~J.~Owen, A.~W.~Thomas and R.~D.~Young, arXiv:1411.3781 [hep-lat]. \bibitem{albermela} A.~Martinez Torres, M.~Bayar, D.~Jido and E.~Oset, Phys.\ Rev.\ C {\bf 86}, 055201 (2012) [arXiv:1202.4297 [hep-lat]]. \bibitem{Formaggio:2013kya} J.~A.~Formaggio and G.~P.~Zeller, Rev.\ Mod.\ Phys.\ {\bf 84}, 1307 (2012) [arXiv:1305.7513 [hep-ex]]. \bibitem{Minerva} The MINER$\nu$A Collaboration, {\tt http://minerva.fnal.gov}. \bibitem{Drakoulakos:2004gn} D.~Drakoulakos {\it et al.} [Minerva Collaboration], hep-ex/0405002. \bibitem{Fields:2013zhk} L.~Fields {\it et al.} [MINERvA Collaboration], Phys.\ Rev.\ Lett.\ {\bf 111}, no. 2, 022501 (2013) [arXiv:1305.2234 [hep-ex]]. \bibitem{Fiorentini:2013ezn} G.~A.~Fiorentini {\it et al.} [MINERvA Collaboration], Phys.\ Rev.\ Lett.\ {\bf 111}, no. 2, 022502 (2013) [arXiv:1305.2243 [hep-ex]]. \bibitem{Tice:2014pgu} B.~G.~Tice {\it et al.} [MINERvA Collaboration], Phys.\ Rev.\ Lett.\ {\bf 112}, no. 23, 231801 (2014) [arXiv:1403.2103 [hep-ex]]. \bibitem{Higuera:2014azj} A.~Higuera {\it et al.} [MINERvA Collaboration], Phys.\ Rev.\ Lett.\ {\bf 113}, 261802 (2014) [arXiv:1409.3835 [hep-ex]]. \bibitem{Singh:2006xp} S.~K.~Singh and M.~J.~Vicente Vacas, Phys.\ Rev.\ D {\bf 74}, 053009 (2006) [hep-ph/0606235]. \bibitem{Mintz:2007zz} S.~L.~Mintz and L.~Wen, Eur.\ Phys.\ J.\ A {\bf 33}, 299 (2007). \bibitem{Kuzmin:2008zz} K.~S.~Kuzmin and V.~A.~Naumov, Phys.\ Atom.\ Nucl.\ {\bf 72}, 1501 (2009) [Yad.\ Fiz.\ {\bf 72}, 1555 (2009)]. \bibitem{Alam:2013cra} M.~R.~Alam, S.~Chauhan, M.~S.~Athar and S.~K.~Singh, Phys.\ Rev.\ D {\bf 88}, no. 7, 077301 (2013) [arXiv:1310.7704 [nucl-th]]. \bibitem{RafiAlam:2010kf} M.~Rafi Alam, I.~Ruiz Simo, M.~Sajjad Athar and M.~J.~Vicente Vacas, Phys.\ Rev.\ D {\bf 82}, 033001 (2010) [arXiv:1004.5484 [hep-ph]]. \bibitem{manolok} M.~R.~Alam, I.~R.~Simo, M.~S.~Athar and M.~J.~Vicente Vacas, Phys.\ Rev.\ D {\bf 85}, 013014 (2012). \bibitem{Alvarez-Ruso:2014bla} L.~Alvarez-Ruso, Y.~Hayato and J.~Nieves, New J.\ Phys.\ {\bf 16}, 075015 (2014) [arXiv:1403.2673 [hep-ph]]. \bibitem{gasser} J.~Gasser and H.~Leutwyler, Annals Phys.\ {\bf 158}, 142 (1984). \bibitem{pich} A.~Pich, Rept.\ Prog.\ Phys.\ {\bf 58} (1995) 563. \bibitem{Scherer:2002tk} S.~Scherer, Adv.\ Nucl.\ Phys.\ {\bf 27}, 277 (2003) [hep-ph/0210398]. \bibitem{pdg} J.~Beringer {\it et al.} [Particle Data Group Collaboration], Phys.\ Rev.\ D {\bf 86}, 010001 (2012). \bibitem{Cabibbo:2003cu} N.~Cabibbo, E.~C.~Swallow and R.~Winston, Ann.\ Rev.\ Nucl.\ Part.\ Sci.\ {\bf 53}, 39 (2003) [hep-ph/0307298]. \bibitem{Wu:2009nw} J.~-J.~Wu, S.~Dulat and B.~S.~Zou, Phys.\ Rev.\ C {\bf 81} (2010) 045210. \bibitem{Gao:2010hy} P.~Gao, J.~-J.~Wu and B.~S.~Zou, Phys.\ Rev.\ C {\bf 81}, 055203 (2010). \bibitem{Lalakulich:2012gm} O.~Lalakulich, K.~Gallmeister and U.~Mosel, Phys.\ Rev.\ C {\bf 86}, 014607 (2012) [arXiv:1205.1061 [nucl-th]]. \bibitem{carrasco} R.~C.~Carrasco and E.~Oset, Nucl.\ Phys.\ A {\bf 536}, 445 (1992). \bibitem{carrasco1} R.~C.~Carrasco, E.~Oset and L.~L.~Salcedo, Nucl.\ Phys.\ A {\bf 541}, 585 (1992). \bibitem{metag} D.~Trnka {\it et al.} [CBELSA/TAPS Collaboration], Phys.\ Rev.\ Lett.\ {\bf 94}, 192303 (2005) [nucl-ex/0504010]. \end{thebibliography} |
1501.04074 | Title: On the Density of Transitive Tournaments
Abstract: We prove that for every fixed $k$, the number of occurrences of the
transitive tournament $Tr_k$ of order $k$ in a tournament $T_n$ on $n$ vertices
is asymptotically minimized when $T_n$ is random. In the opposite direction, we
show that any sequence of tournaments $\{T_n\}$ achieving this minimum for any
fixed $k\geq 4$ is necessarily quasi-random. We present several other
characterizations of quasi-random tournaments nicely complementing previously
known results and relatively easily following from our proof techniques.
Body: \newcommand{\tendsto}{\mathop{\longrightarrow}\limits} \newcommand{\tendstodown}{\mathop{\longsearrow}\limits} \newcommand{\tendstoup}{\mathop{\longnearrow}\limits} \def\({\left(} \def\){\right)} \def\[{\left[} \def\]{\right]} \def\<{\left\langle} \def\>{\right\rangle} \def\lv{\left\lvert} \def\mv{\middle\vert} \def\rv{\right\rvert} \def\lV{\left\lVert} \def\mV{\middle\Vert} \def\rV{\right\rVert} \def\llb{\left\llbracket} \def\rrb{\right\rrbracket} \def\iff{\Longleftrightarrow} \def\implies{\Longrightarrow} \def\isimplied{\Longleftarrow} \newcommand{\floor}[1]{\ensuremath{\left\lfloor#1\right\rfloor}} \newcommand{\ceil}[1]{\ensuremath{\left\lceil#1\right\rceil}} \def\divides{\mathbin{\mid}} \newcommand{\symdiff}{\mathbin{\triangle}} \newcommand{\comp}{\mathbin{\circ}} \newcommand{\cond}{\mathbin{\vert}} \newcommand{\rest}{\vert} \let\geq\geqslant \let\leq\leqslant \let\:\colon \let\oldepsilon\epsilon \let\epsilon\varepsilon \let\oldphi\phi \let\phi\varphi \let\oldemptyset\emptyset \let\emptyset\varnothing \def\AA{\ensuremath{\mathbb{A}}} \def\BB{\ensuremath{\mathbb{B}}} \def\CC{\ensuremath{\mathbb{C}}} \def\DD{\ensuremath{\mathbb{D}}} \def\EE{\ensuremath{\mathbb{E}}} \def\FF{\ensuremath{\mathbb{F}}} \def\GG{\ensuremath{\mathbb{G}}} \def\HH{\ensuremath{\mathbb{H}}} \def\II{\ensuremath{\mathbb{I}}} \def\JJ{\ensuremath{\mathbb{J}}} \def\KK{\ensuremath{\mathbb{K}}} \def\LL{\ensuremath{\mathbb{L}}} \def\MM{\ensuremath{\mathbb{M}}} \def\NN{\ensuremath{\mathbb{N}}} \def\OO{\ensuremath{\mathbb{O}}} \def\PP{\ensuremath{\mathbb{P}}} \def\QQ{\ensuremath{\mathbb{Q}}} \def\RR{\ensuremath{\mathbb{R}}} \def\SS{\ensuremath{\mathbb{S}}} \def\TT{\ensuremath{\mathbb{T}}} \def\UU{\ensuremath{\mathbb{U}}} \def\VV{\ensuremath{\mathbb{V}}} \def\WW{\ensuremath{\mathbb{W}}} \def\XX{\ensuremath{\mathbb{X}}} \def\YY{\ensuremath{\mathbb{Y}}} \def\ZZ{\ensuremath{\mathbb{Z}}} \def\One{\ensuremath{\mathbbm{1}}} \def\cA{\ensuremath{\mathcal{A}}} \def\cB{\ensuremath{\mathcal{B}}} \def\cC{\ensuremath{\mathcal{C}}} \def\cD{\ensuremath{\mathcal{D}}} \def\cE{\ensuremath{\mathcal{E}}} \def\cF{\ensuremath{\mathcal{F}}} \def\cG{\ensuremath{\mathcal{G}}} \def\cH{\ensuremath{\mathcal{H}}} \def\cI{\ensuremath{\mathcal{I}}} \def\cJ{\ensuremath{\mathcal{J}}} \def\cK{\ensuremath{\mathcal{K}}} \def\cL{\ensuremath{\mathcal{L}}} \def\cM{\ensuremath{\mathcal{M}}} \def\cN{\ensuremath{\mathcal{N}}} \def\cO{\ensuremath{\mathcal{O}}} \def\cP{\ensuremath{\mathcal{P}}} \def\cQ{\ensuremath{\mathcal{Q}}} \def\cR{\ensuremath{\mathcal{R}}} \def\cS{\ensuremath{\mathcal{S}}} \def\cT{\ensuremath{\mathcal{T}}} \def\cU{\ensuremath{\mathcal{U}}} \def\cV{\ensuremath{\mathcal{V}}} \def\cW{\ensuremath{\mathcal{W}}} \def\cX{\ensuremath{\mathcal{X}}} \def\cY{\ensuremath{\mathcal{Y}}} \def\cZ{\ensuremath{\mathcal{Z}}} \def\fA{\ensuremath{\mathfrak{A}}} \def\fB{\ensuremath{\mathfrak{B}}} \def\fC{\ensuremath{\mathfrak{C}}} \def\fD{\ensuremath{\mathfrak{D}}} \def\fE{\ensuremath{\mathfrak{E}}} \def\fF{\ensuremath{\mathfrak{F}}} \def\fG{\ensuremath{\mathfrak{G}}} \def\fH{\ensuremath{\mathfrak{H}}} \def\fI{\ensuremath{\mathfrak{I}}} \def\fJ{\ensuremath{\mathfrak{J}}} \def\fK{\ensuremath{\mathfrak{K}}} \def\fL{\ensuremath{\mathfrak{L}}} \def\fM{\ensuremath{\mathfrak{M}}} \def\fN{\ensuremath{\mathfrak{N}}} \def\fO{\ensuremath{\mathfrak{O}}} \def\fP{\ensuremath{\mathfrak{P}}} \def\fQ{\ensuremath{\mathfrak{Q}}} \def\fR{\ensuremath{\mathfrak{R}}} \def\fS{\ensuremath{\mathfrak{S}}} \def\fT{\ensuremath{\mathfrak{T}}} \def\fU{\ensuremath{\mathfrak{U}}} \def\fV{\ensuremath{\mathfrak{V}}} \def\fW{\ensuremath{\mathfrak{W}}} \def\fX{\ensuremath{\mathfrak{X}}} \def\fY{\ensuremath{\mathfrak{Y}}} \def\fZ{\ensuremath{\mathfrak{Z}}} \newenvironment{functiondef}{ \begin{tabular}{ *{2}{>{$\displaystyle}r<{$}} >{$\displaystyle}c<{$} >{$\displaystyle}l<{$} }}{ \end{tabular} } \maketitle \begin{abstract} We prove that for every fixed~$k$, the number of occurrences of the transitive tournament~$\Tr_k$ of order~$k$ in a tournament~$T_n$ on~$n$ vertices is asymptotically minimized when~$T_n$ is random. In the opposite direction, we show that any sequence of tournaments~$\{T_n\}$ achieving this minimum for any fixed~$k\geq 4$ is necessarily quasi-random. We present several other characterizations of quasi-random tournaments nicely complementing previously known results and relatively easily following from our proof techniques. \end{abstract} For a fixed combinatorial object~$M$, e.g.~a graph, a hypergraph or a tournament, the number of its occurrences in a large random object~$N$ of the same signature is always easy to compute asymptotically. Reverse problems have become the subject of active research in combinatorics. In the first direction along these lines, we are interested in the questions of the kind ``for which templates~$M$ the number of occurrences is asymptotically optimized by the random object'' or, in other terms, when the latter forms an extremal configuration for the corresponding extremal problem. Perhaps, one of the most prominent settings in which this question has been studied systematically is that of~\emph{$k$-common graphs}~. In this setting, $M$ is an ordinary graph, $N$ is a~$k$-edge-coloring of a complete graph, and an ``occurrence'' means a monochromatic copy. The~$k$-common graphs are those for which the random edge coloring asymptotically minimizes the number of occurrences, and they were studied in many papers, see e.g.~. The second direction focusses on the complementary question: ``When all extremal configurations for this extremal problem are necessarily quasi-random?'' We do not attempt to review here the theory of quasi-randomness: beginning from the seminal papers by Thomason~ and Chung, Graham, Wilson~, it has developed into a vast field surveyed for example in~. But the main thrust of this theory is that many a priori different properties eventually lead to the same class of combinatorial objects, hereafter called {\em quasi-random}. Perhaps, the most intuitive of these properties says that quasi-random objects are precisely those~$N$ in which {\em all} fixed templates~$M$ have the right density of occurrences, understood asymptotically. \bigskip In this note we study questions of both kinds for tournaments. By analogy with graphs, it might be tempting to call a tournament~$T$ {\em common} if for any increasing sequence~$\{T_n\}$ of tournaments we have \begin{equation} \liminf_{n\to\infty} p(T,T_n)\geq \frac{k!}{|\Aut(T)|\cdot 2^{k\choose 2}}, \end{equation} where~$k\df |V(T)|$, $p(T,T_n)$ is the density with which~$T$ appears in~$T_n$ as an unlabelled sub-tournament and~$\Aut(T)$ denotes the group of automorphisms of~$T$. In the language of flag algebras~ it can be stated more cleanly and concisely as \begin{equation} \forall \phi\in\Hom^+(\mathcal A^0[T_{\text{Tournaments}}], \mathbb R) \of{\phi(T) \geq \frac{k!}{|\Aut(T)|\cdot 2^{k\choose 2}}}, \end{equation} and in what follows we will mostly use this language. The reader who feels uncomfortable with limit objects should have little difficulties translating our statements to the finite world by analogy with~\eqref{naive}. Upon a moment's reflection it becomes clear that only transitive tournaments may possibly be common since all others occur with zero density in the increasing sequence~$\{\Tr_n\}$ of transitive tournaments. This gives rise to the natural question if the converse is true, that is if \begin{equation} \phi(\Tr_k) \geq \frac{k!}{2^{k\choose 2}}, \end{equation} where again~$\phi$ is any algebra homomorphism from~$\Hom^+(\mathcal A^0[T_{\text{Tournaments}}], \mathbb R)$. To the best of our knowledge, this question was not explicitly asked before. The cases~$k=0,1,2$ are trivial, and the case~$k=3$ was implicitly answered by Chung and~Graham in~\cite[Fact~1]{ChGr2}. With a bit of effort, the bound~\eqref{main_inequality} for~$k=4$ can be extracted from Griffith's paper~\cite[Proposition~3.1~(i)]{Grif}, but his proof does not seem to be generalizable to larger values of~$k$. Our first main result confirms~\eqref{main_inequality} for all values of~$k$. In other words, transitive tournaments are common and vice versa. This result also has an interesting ``sharp threshold'' flavor to it. Namely, if we take an arbitrarily large transitive tournament and flip even a {\em single} arc in it, the extremal (minimizing) configuration jumps from the random tournament to the opposite side of the spectrum, i.e.~transitive tournaments. \medskip The theory of quasi-random tournaments was inaugurated by Chung and~Graham in~ and followed up in~. A significant portion of all these papers is devoted to the question formulated above in more general context: for which tournaments~$T$ the equality in~\eqref{flagmatic} implies that~$\phi$ is quasi-random? We show this for any transitive tournament~$\Tr_k$ with.}~$k\geq 5$. Many characterizations in~ are given in terms of ``flag concentration''. Given a tournament~$T$ and an edge~$\edge uv\in E(T)$, all other vertices~$w\in V(T)\setminus \{u,v\}$ can be classified into four classes (``flags''): \begin{enumerate} \item $\edge uw, \edge vw\in E(T)$, \item $\edge wu, \edge wv\in E(T)$, \item $\edge uw, \edge wv\in E(T)$. \item $\edge vw, \edge wu\in E(T)$, \end{enumerate} Following and expanding a bit the notation in~, we let~$O^A(u,v)$, $I^A(u,v)$, $\Tr_3^A(u,v)$ and~$\vec C_3^A(u,v)$ denote the numbers of vertices in the four classes (taken in this order, see also Figure~ below) divided by~$|V(T)|-2$. One of the keystone characterizations in~ ($P_4$, to be exact) says that a sequence of tournaments is quasi-random if and only if (which we do in Section~) or digraphons~\cite[Example~9.2]{DiJa}.}~$O^A(u,v)$ is ``nearly''~1/4 for ``almost all'' edges~$\edge uv$. A similar statement for~$I^A(u,v)$ follows by duality. Our methods allow us to do, in exactly the same manner, the two remaining cases: $\Tr_3^A(u,v)$ and~$\vec C_3^A(u,v)$. Stated in the finite language and combined with the previously known results, we now have that a sequence of tournaments~$\{T_n\}$ is quasi-random if and only if \begin{equation} \sum_{\edge uv\in E(T_n)} |F(u,v)-1/4| \leq o(n^2), \end{equation} where~$F$ corresponds to {\em any} of the four cases above. The note is organized as follows. In Section~ we remind some rudimentary concepts from the theory of flag algebras and show how to treat quasi-randomness in that context. Again, the reader who feels uncomfortable with this language (and is willing to tolerate a bit of coping with low-order error terms instead) should have no difficulties replacing~$\sigma$-extensions with averaging over vertices or arcs in large but finite tournaments, see e.g.~. In Section~ we prove the bound~\eqref{flagmatic}, and also that the equality is attained here if and only if~$\phi$ is quasi-random (Theorem~). Finally, in Section~ we prove additional characterizations of quasi-random tournaments in terms of ``flag concentrations'' (Theorems~ and~). \section{Quasi-Randomness in Flag Algebras} In this section, we translate the results of quasi-randomness that we are going to use to the language of flag algebras. We assume the reader has some familiarity with not only the basic setting of flag algebra, but also with extensions of homomorphisms~\cite[\S 3.2]{flag}. Following the notation of~, we consider the theory of tournaments~$T_{\text{Tournaments}}$ and we let~$1$ denote the (unique) type of size~$1$ and~$A$ denote the type of size~$2$ such that the vertex labelled with~$1$ beats the other (labelled) vertex (see Figure~). For a type~$\sigma$, we denote the unity of the algebra~$\cA^\sigma$ by~$\One_\sigma$, and, as always, $\One_0$ is abbreviated to~$\One$. We have already introduced the notation~$\Tr_k$ to denote the transitive tournament of size~$k$. We let~$\vec C_3$ be the~$3$-directed cycle (that is, the only other tournament of size~$3$) and we define the following tournaments of size~$4$. \begin{enumerate} \item The tournament~$W_4$, which is the (unique) \emph{non-transitive} tournament of size~$4$ that has a vertex with outdegree~$3$ (that is, there is a ``winner'' in~$W_4$); \item The tournament~$L_4$, which is the (unique) \emph{non-transitive} tournament of size~$4$ that has a vertex with indegree~$3$ (that is, there is a ``loser'' in~$L_4$); \item The tournament~$R_4$, which is the (unique) tournament of size~$4$ that has outdegree sequence~$(1,1,2,2)$. \end{enumerate} Note that along with~$\Tr_4$, this list covers all tournaments of size~$4$. We define the~$1$-flag~$\alpha$ as the (unique)~$1$-flag of size~$2$ in which the labelled vertex beats the unlabelled vertex. We also define the following~$A$-flags of size~$3$. \begin{enumerate} \item The flag~$O^A$, in which the only unlabelled vertex is beaten by both labelled vertices; \item The flag~$I^A$, in which the only unlabelled vertex beats both labelled vertices; \item The flag~$\Tr_3^A$, which is the only remaining~$A$-flag whose underlying model is~$\Tr_3$; \item The flag~$\vec C_3^A$, which is the only~$A$-flag whose underlying model is~$\vec C_3$. \end{enumerate} Again, this is the complete list of~$A$-flags of size~$3$. We also follow the original notation of flag algebras when using the downward operator~$\llb{}\cdot{}\rrb_\sigma$ to the~$0$-algebra or when using~$\sigma$-extensions of homomorphisms~$\phi\in\Hom^+(\cA^0,\RR)$ (which are denoted by~$\bm{\phi^\sigma}$). We remind that~$\bm{\phi^\sigma}$ can be conveniently viewed~\cite[Definition~10]{flag} as the unique~$\Hom^+(\cA^\sigma,\RR)$-valued random variables satisfying the identities \begin{equation} \expect{\bm{\phi^\sigma}(F)} = \frac{\phi(\llb F\rrb_\sigma)}{\phi(\llbracket\One_\sigma\rrbracket_\sigma)} \end{equation} for every~$F\in\cF^\sigma$. Finally, for a flag~$F_0$ of type~$\sigma$ with~$|\sigma|+1$ vertices we have the algebra homomorphism~$\pi^{F_0}\function{\mathcal A^0}{\mathcal A^\sigma_{F_0}}$, where~$\mathcal A^\sigma_{F_0}$ is the localization of the algebra~$\mathcal A^\sigma$ with respect to the multiplicative system~$\set{F_0^\ell}{\ell\in \mathbb N}$~\cite[\S 2.3.2]{flag}. Intuitively, it corresponds to taking the sub-model induced by the flag~$F_0$, and the localization is needed for proper normalization resulting from decreasing the set of vertices. In this note we will only need the operator~$\pi^{O^A}$. We will not need these concepts in the more complicated scenario when the smaller type is also non-trivial. \smallskip Let us denote the homomorphism of~$\Hom^+(\cA^0,\RR)$ corresponding to the random tournament by~$\phirdm$, that is, it is the almost sure limit of the sequence of random tournaments (where each arc orientation is picked independently with probability~$1/2$) when the number of vertices goes to infinity. Note that for every tournament~$T$, we have \begin{align*} \phirdm(T) & = \frac{\lv V(T)\rv!}{\lv\Aut(T)\rv\cdot 2^{\binom{|T|}{2}}}. \end{align*} In particular, since transitive tournaments possess only trivial automorphisms, we have \begin{align*} \phirdm(\Tr_k) & = \frac{k!}{2^{\binom{k}{2}}}, \end{align*} for every~$k\in\NN$. As for the quasi-randomness properties (the~$P$ properties of Chung--Graham~), we are interested in the following ones. \begin{itemize} \item $P_2$: $\phi(\Tr_4 + R_4) = 3/4$; \item $P_4$: $\bm{\phi^A}(O^A) = 1/4$ a.s. \end{itemize} That is, the above are equivalent to each other, and are equivalent to the fact~$\phi=\phirdm$ (which, in the terminology of~, is precisely~$\forall s\in\NN, P_1(s)$). The translation of property~$P_2$ follows from simple arithmetics relating homomorphism densities to flag algebra densities, and the translation of~$P_4$ follows from the definition of extensions of homomorphisms. \section{The Minimum Density of Transitive Tournaments} We start by defining for every~$k\geq 2$ the~$1$-flag~$\Tr_k^W$ as the flag obtained from~$\Tr_k$ by labelling its winner (i.e.~the unique vertex with outdegree~$k-1$) and the~$A$-flag~$\Tr_k^{W2}$ as the flag obtained from~$\Tr_k$ by labelling its winner with the label~$1$ and its runner-up (i.e.~the unique vertex with outdegree~$k-2$) with the label~$2$. In particular, with this definition we have~$\alpha=\Tr_2^W$ and~$O^A=\Tr_3^{W2}$. We further note that~$\alpha^2 = \Tr_3^W$, $\llb\alpha\rrb_1 = \Tr_2/2 = \One/2$ and for every~$k\geq 2$, we have \begin{align*} \Tr_k & = k\llbracket\Tr_k^W\rrbracket_1 = k(k-1)\llbracket\Tr_k^{W2}\rrbracket_A;\\ \Tr_k^{W2} & = \pi^{O^A}(\Tr_{k-2})\cdot(O^A)^{k-2}. \end{align*} We are now ready to prove the main result. We remind the reader that, although the proof is presented for all~$k\in\NN$, the result for~$k\leq 4$ was known before. \begin{theorem} In the theory of tournaments, for every~$k\in\NN$ and every~$\phi\in\Hom^+(\cA^0,\RR)$, we have \begin{align*} \phi(\Tr_k) & \geq \frac{k!}{2^{\binom{k}{2}}}. \end{align*} Furthermore, for any fixed~$k\geq 4$, the equality holds if and only if~$\phi$ is the quasi-random homomorphism~$\phirdm$. \end{theorem} \begin{proof} For~$k=0,1,2$, the result is trivial, because as flag algebra elements, we have~$\Tr_0=\Tr_1=\Tr_2=\One$ and the right hand side of the formula evaluates to~$1$ in these cases. For~$k=3$, note that since~$\llbracket\One_1\rrbracket_1 = \One$, we have $$ \phi(\Tr_3) = 3\phi(\llbracket\alpha^2\rrbracket_1) = 3\expect{\bm{\phi^1}(\alpha)^2} \geq 3\expect{\bm{\phi^1}(\alpha)}^2 = 3\phi(\llb\alpha\rrb_ 1)^2 = \frac{3}{4} = \frac{3!}{2^{\binom{3}{2}}}, $$ Now we proceed by induction. Suppose that~$k\geq 4$ and that the result is valid for~$k-2$. Since~$\llbracket\One_A\rrbracket_A = \One/2$, we have \begin{align*} \phi(\Tr_k) & = k(k-1)\phi(\llbracket\pi^{O^A}(\Tr_{k-2})\cdot(O^A)^{k-2}\rrbracket_A) \\ & = \frac{k(k-1)}{2} \expect{ \bm{\phi^A}(\pi^{O^A}(\Tr_{k-2})) \bm{\phi^A}\of{(O^A)^{k-2}}}. \end{align*} Since~$\bm{\phi^A}(O^A)> 0$ implies that~$\phi^A\circ \pi^{O^A}\in\Hom^+(\cA^0,\RR)$, by inductive hypothesis we have $$ \expect{ \bm{\phi^A}(\pi^{O^A}(\Tr_{k-2})) \bm{\phi^A}\of{(O^A)^{k-2}}} \geq \frac{(k-2)!}{2^{\binom{k-2}{2}}} \expect{\bm{\phi^A}((O^A)^{k-2})}. $$ By Jensen's inequality, we have \begin{equation} \expect{\bm{\phi^A}(O^A)^{k-2}} \geq \expect{\bm{\phi^A}(O^A)}^{k-2} = (2\phi(\llbracket O^A\rrbracket_A))^{k-2} = \(\frac{\phi(\Tr_3)}{3}\)^{k-2}. \end{equation} Since we already proved the case~$k=3$, we have~$\phi(\Tr_3)\geq 3/4$, so putting things together, we get \begin{equation} \phi(\Tr_k) \geq \frac{k(k-1)}{2}\cdot\frac{(k-2)!}{2^{\binom{k-2}{2}}} \(\frac{\phi(\Tr_3)}{3}\)^{k-2} \geq \frac{k!}{2^{\binom{k}{2}}}. \end{equation} \medskip For the ``furthermore'' part, note that if equality holds for~$\phi$, then along the proof we have equalities instead of inequalities. In particular, since Jensen's inequality was used in~\eqref{jensen} with the function~$x^{k-2}$, which for~$k\geq 4$ is strictly convex does not hold for~$k=3$.} on~$[0,+\infty)$, we have that~$\bm{\phi^A}(O^A)$ is a.s.~a constant, denote it by~$C$. Furthermore, we also have~$\phi(\Tr_3) = 3/4$ as it is used in the chain of inequalities~\eqref{chain}. This allows us to compute~$C$ as follows: $$ C=\expect{\bm{\phi^A}(O^A)} = 2\phi(\llbracket O^A\rrbracket_A) = \frac{\phi(\Tr_3)}{3} = \frac{1}{4}. $$ Therefore~$\phi$ satisfies~$P_4$, hence~$\phi=\phirdm$. \end{proof} \begin{remark*} Note that if we wanted to prove just the inequality statement (i.e.~without the ``furthermore'' part), we could have done the induction with a simpler argument involving~$\pi^\alpha(\Tr_{k-1})\alpha^{k-1}$ instead of~$\pi^{O^A}(\Tr_{k-2})\cdot(O^A)^{k-2}$. Moreover, modulo the following (straightforward) generalization of Cauchy--Schwarz inequality~\cite[(22)]{flag}: $$ \eval{f^k}{\sigma} \cdot \eval{\One_\sigma}{\sigma}^{k-1} \geq \eval{f}{\sigma}^k $$ we could have also avoided the extension~$\bm{\phi^A}$ at all and reduced the whole argument to a several-lines calculation. \end{remark*} \section{New Characterizations with~$A$-Flags and~$A$-Extensions} As we already said in the introduction, Chung and~Graham presented in~ quasi-random characterizations involving the~$A$-flag~$O^A$. Reverting all arcs of the tournament, we get the dual characterization involving~$I^A$ (which is stated below for completeness). \begin{itemize} \item $P_4'$: $\bm{\phi^A}(I^A) = 1/4$ a.s. \end{itemize} In this section, we prove that the two remaining~$A$-flags~$\Tr_3^A$ and~$\vec C_3^A$ of size~$3$ also characterize quasi-randomness of tournaments. Note, however, that these flags are self-dual, so they must be treated separately. \begin{theorem} In the theory of tournaments, for every~$\phi\in\Hom^+(\cA^0,\RR)$, we have that \begin{align*} \bm{\phi^A}(\Tr_3^A) = \frac{1}{4} \as \end{align*} if and only if~$\phi$ is the quasi-random homomorphism~$\phirdm$. \end{theorem} \begin{proof} This is similar to the second part of the proof of Theorem~, except that now we use the flag~$\Tr_3^A$ instead of~$O^A$. Let~$\Tr_4^A$ denote the~$A$-flag obtained from~$\Tr_4$ by labelling its winner with label~$1$ and labelling its loser with label~$2$ (see Figure~). Note that~$(\Tr_3^A)^2 = \Tr_4^A$ and that~$\Tr_4 = 12\llb \Tr_4^A\rrb_A$. Since~$\llbracket\One_A\rrbracket_A = \One/2$, we have $$ \phi(\Tr_4) = 12\phi(\llbracket (\Tr_3^A)^2\rrbracket_A) = 6\expect{\bm{\phi^A}(\Tr_3^A)^2}\geq 6\expect{\bm{\phi^A}(\Tr_3^A)}^2 = 6(2\phi(\llbracket\Tr_3^A\rrbracket_A))^2 = \frac{2}{3}\phi(\Tr_3)^2. $$ By Theorem~, we know that~$\phi(\Tr_3)\geq 3/4$, hence we have \begin{align*} \phi(\Tr_4) & \geq \frac{3}{8} = \frac{4!}{2^{\binom{4}{2}}}. \end{align*} But, since~$x^2$ is both strictly convex and strictly increasing on~$[0,+\infty)$, we have that equality holds if and only if we have almost surely \begin{align*} \bm{\phi^A}(\Tr_3^A) & = \expect{\phi^A(\Tr_3^A)} = \frac{\phi(\Tr_3)}{3} = \frac{1}{4}. \end{align*} By the ``furthermore'' part of Theorem~, we have \begin{align*} \bm{\phi^A}(\Tr_3^A) & = \frac{1}{4}\as \iff \phi(\Tr_4) = \frac{4!}{2^{\binom{4}{2}}} \iff \phi = \phirdm. \qedhere \end{align*} \end{proof} For the final flag~$\vec C_3^A$ the proof cannot be done analogously because the underlying tournament is now~$\vec C_3$ instead of~$\Tr_3$. Nevertheless, we are able to obtain the result by other means. \begin{theorem} In the theory of tournaments, for every~$\phi\in\Hom^+(\cA^0,\RR)$, we have that \begin{align*} \bm{\phi^A}(\vec C_3^A) = \frac{1}{4} \as \end{align*} if and only if~$\phi$ is the quasi-random homomorphism~$\phirdm$. \end{theorem} \begin{proof} In one direction it is simple. By Chung and Graham's characterizations~$P_4$ and~$P_4'$ (that is, after reverting the arcs) and by Theorem~, we have \begin{align*} \bm{\phirdm^A}(O^A) & = \bm{\phirdm^A}(I^A) = \bm{\phirdm^A}(\Tr_3^A) = \frac{1}{4} \as \end{align*} Hence, since~$\vec C_3^A = \One_A - O^A - I^A - \Tr_3^A$, we have \begin{align*} \bm{\phirdm^A}(\vec C_3^A) = \frac{1}{4} \as \end{align*} \medskip For the converse, note first that if~$\phi\in\Hom^+(\cA^0,\RR)$ is such that~$\bm{\phi^A}(\vec C_3^A) = 1/4$ a.s., then \begin{align*} \phi(\vec C_3) & = \frac{\phi(\llbracket\vec C_3^A\rrbracket_A)}{\phi(\llbracket\One_A\rrbracket_A)} = \expect{\bm{\phi^A}(\vec C_3^A)} = \frac{1}{4}. \end{align*} Now, it is straightforward to check the following flag algebra identities. \begin{align*} R_4 & = 12\llbracket (\vec C_3^A)^2 \rrbracket_A; & \vec C_3 & = \frac{1}{2}R_4 + \frac{1}{4}W_4 + \frac{1}{4}L_4. \end{align*} Using the first identity along with~$\bm{\phi^A}(\vec C_3^A)=1/4$ a.s., we get \begin{align*} \phi(R_4) & = 12\phi(\llbracket (\vec C_3^A)^2 \rrbracket_A) = 6\expect{\bm{\phi^A}(\vec C_3^A)^2} = \frac{3}{8}. \end{align*} Using the second identity, we get \begin{align*} \phi(W_4 + L_4) & = 4\phi(\vec C_3) - 2\phi(R_4) = \frac{1}{4}, \end{align*} hence~$\phi(\Tr_4+R_4) = 3/4$ (since~$\Tr_4+R_4+W_4+L_4=\One$), which is property~$P_2$ of Chung--Graham. Therefore~$\phi=\phirdm$. \end{proof} \bibliographystyle{alpha} \bibliography{razb} \begin{tikzpicture}[scale=1.1] \foreach \i in {0,...,3}{ \pgfmathsetmacro{\base}{2*\i} \pgfmathsetmacro{\next}{\base+1} \coordinate (T3A\i) at (\base cm, 1cm); \coordinate (T3B\i) at (\next cm, 1cm); \coordinate (T3C\i) at ($1/2*(T3A\i) + 1/2*(T3B\i) + (0cm,0.707106cm)$); \coordinate (T4A\i) at (\base cm,-1.5cm); \coordinate (T4B\i) at (\next cm,-1.5cm); \coordinate (T4C\i) at (\next cm,-0.5cm); \coordinate (T4D\i) at (\base cm,-0.5cm); \coordinate (TpA\i) at (\base cm, -3cm); \coordinate (TpC\i) at (\next cm, -3cm); \coordinate (TpB\i) at ($1/2*(TpA\i) + 1/2*(TpC\i)$); \coordinate (F2A\i) at (\base cm,-4.5cm); \coordinate (F2B\i) at (\next cm,-4.5cm); \coordinate (F3A\i) at (\base cm, -6cm); \coordinate (F3B\i) at (\next cm, -6cm); \coordinate (F3C\i) at ($1/2*(F3A\i) + 1/2*(F3B\i) + (0cm,0.707106cm)$); \coordinate (F4A\i) at (\base cm,-8.5cm); \coordinate (F4B\i) at (\next cm,-8.5cm); \coordinate (F4C\i) at (\next cm,-7.5cm); \coordinate (F4D\i) at (\base cm,-7.5cm); } \foreach \i in {1,2}{ \filldraw (T3A\i) circle (1pt); \filldraw (T3B\i) circle (1pt); \filldraw (T3C\i) circle (1pt); \draw[thick, arrows={-latex}] (T3A\i) -- (T3B\i); \draw[thick, arrows={-latex}] (T3B\i) -- (T3C\i); } \draw[thick, arrows={-latex}] (T3A1) -- (T3C1); \draw[thick, arrows={-latex}] (T3C2) -- (T3A2); \node[below] at ($1/2*(T3A1) + 1/2*(T3B1)$) {$\Tr_3$}; \node[below] at ($1/2*(T3A2) + 1/2*(T3B2)$) {$\vec C_3$}; \foreach \i in {0,...,3}{ \filldraw (T4A\i) circle (1pt); \filldraw (T4B\i) circle (1pt); \filldraw (T4C\i) circle (1pt); \filldraw (T4D\i) circle (1pt); \draw[thick, arrows={-latex}] (T4A\i) -- (T4B\i); \draw[thick, arrows={-latex}] (T4B\i) -- (T4C\i); } \draw[thick, arrows={-latex}] (T4A0) -- (T4C0); \draw[thick, arrows={latex-}] (T4A1) -- (T4C1); \draw[thick, arrows={latex-}] (T4A2) -- (T4C2); \draw[thick, arrows={latex-}] (T4A3) -- (T4C3); \draw[thick, arrows={-latex}] (T4A0) -- (T4D0); \draw[thick, arrows={-latex}] (T4B0) -- (T4D0); \draw[thick, arrows={-latex}] (T4C0) -- (T4D0); \draw[thick, arrows={latex-}] (T4A1) -- (T4D1); \draw[thick, arrows={latex-}] (T4B1) -- (T4D1); \draw[thick, arrows={latex-}] (T4C1) -- (T4D1); \draw[thick, arrows={-latex}] (T4A2) -- (T4D2); \draw[thick, arrows={-latex}] (T4B2) -- (T4D2); \draw[thick, arrows={-latex}] (T4C2) -- (T4D2); \draw[thick, arrows={-latex}] (T4A3) -- (T4D3); \draw[thick, arrows={-latex}] (T4B3) -- (T4D3); \draw[thick, arrows={latex-}] (T4C3) -- (T4D3); \node[below] at ($1/2*(T4A0) + 1/2*(T4B0)$) {$\Tr_4$}; \node[below] at ($1/2*(T4A1) + 1/2*(T4B1)$) {$W_4$}; \node[below] at ($1/2*(T4A2) + 1/2*(T4B2)$) {$L_4$}; \node[below] at ($1/2*(T4A3) + 1/2*(T4B3)$) {$R_4$}; \filldraw (TpB1) circle (1pt); \filldraw (TpA2) circle (1pt); \filldraw (TpC2) circle (1pt); \draw[thick, arrows={-latex}] (TpA2) -- (TpC2); \node[left] at (TpB1) {$1$}; \node[left] at (TpA2) {$1$}; \node[right] at (TpC2) {$2$}; \node[below] at (TpB1) {$1$}; \node[below] at (TpB2) {$A$}; \filldraw (F2B1) circle (1pt); \filldraw (F2A2) circle (1pt); \draw[thick, arrows={-latex}] (F2B1) -- (F2A2); \node[left] at (F2B1) {$1$}; \node[below] at ($1/2*(F2B1) + 1/2*(F2A2)$) {$\alpha$}; \foreach \i in {0,...,3}{ \filldraw (F3A\i) circle (1pt); \filldraw (F3B\i) circle (1pt); \filldraw (F3C\i) circle (1pt); \draw[thick, arrows={-latex}] (F3A\i) -- (F3B\i); \node[left] at (F3A\i) {$1$}; \node[right] at (F3B\i) {$2$}; } \draw[thick, arrows={-latex}] (F3A0) -- (F3C0); \draw[thick, arrows={-latex}] (F3B0) -- (F3C0); \draw[thick, arrows={latex-}] (F3A1) -- (F3C1); \draw[thick, arrows={latex-}] (F3B1) -- (F3C1); \draw[thick, arrows={-latex}] (F3A2) -- (F3C2); \draw[thick, arrows={latex-}] (F3B2) -- (F3C2); \draw[thick, arrows={latex-}] (F3A3) -- (F3C3); \draw[thick, arrows={-latex}] (F3B3) -- (F3C3); \node[below] at ($1/2*(F3A0) + 1/2*(F3B0)$) {$O^A$}; \node[below] at ($1/2*(F3A1) + 1/2*(F3B1)$) {$I^A$}; \node[below] at ($1/2*(F3A2) + 1/2*(F3B2)$) {$\Tr_3^A$}; \node[below] at ($1/2*(F3A3) + 1/2*(F3B3)$) {$\vec C_3^A$}; \filldraw (F4B1) circle (1pt); \filldraw (F4C1) circle (1pt); \filldraw (F4A2) circle (1pt); \filldraw (F4D2) circle (1pt); \draw[thick, arrows={-latex}] (F4B1) -- (F4A2); \draw[thick, arrows={-latex}] (F4D2) -- (F4A2); \draw[thick, arrows={-latex}] (F4B1) -- (F4C1); \draw[thick, arrows={-latex}] (F4C1) -- (F4D2); \draw[thick, arrows={-latex}] (F4B1) -- (F4D2); \draw[thick, arrows={-latex}] (F4C1) -- (F4A2); \node[left] at (F4B1) {$1$}; \node[right] at (F4A2) {$2$}; \node[below] at ($1/2*(F4B1) + 1/2*(F4A2)$) {$\Tr_4^A$}; \end{tikzpicture} |
1501.04075 | Title: Quenched Voronoi percolation
Abstract: We prove that the probability of crossing a large square in quenched Voronoi
percolation converges to 1/2 at criticality, confirming a conjecture of
Benjamini, Kalai and Schramm from 1999. The main new tools are a quenched
version of the box-crossing property for Voronoi percolation at criticality,
and an Efron-Stein type bound on the variance of the probability of the
crossing event in terms of the sum of the squares of the influences. As a
corollary of the proof, we moreover obtain that the quenched crossing event at
criticality is almost surely noise sensitive.
Body: \begin{abstract} We prove that the probability of crossing a large square in quenched Voronoi percolation converges to $1/2$ at criticality, confirming a conjecture of Benjamini, Kalai and Schramm from 1999. The main new tools are a quenched version of the box-crossing property for Voronoi percolation at criticality, and an Efron-Stein type bound on the variance of the probability of the crossing event in terms of the sum of the squares of the influences. As a corollary of the proof, we moreover obtain that the quenched crossing event at criticality is almost surely noise sensitive. \end{abstract} \maketitle \section{Introduction} The \emph{noise sensitivity} of a Boolean function was introduced in 1999 in a seminal paper of Benjamini, Kalai and Schramm~, and has since developed into an important area of probability theory (see, e.g.,~), linking discrete Fourier analysis with percolation theory and combinatorics. One of the main results of~ gave a sufficient condition for a sequence of functions $f_n \colon \{0,1\}^n \to \{0,1\}$ to be sensitive to small amounts of random noise in the following precise sense: if $\omega \in \{0,1\}^n$ is chosen uniformly at random, and $\omega^\eps$ is obtained from $\omega$ by resampling each variable with some fixed probability $\eps > 0$, then $f_n(\omega)$ and $f_n(\omega^\eps)$ are asymptotically independent. They used this theorem to show that the sequence of functions which encodes crossings of $n \times n$ squares in bond percolation on $\Z^2$ is noise sensitive. Thus, even if one knows all but a \emph{random} $o(1)$-proportion of the edges, one still (with high probability) has very little information about the crossing event. The authors of~ furthermore made a number of conjectures regarding more precise notions of sensitivity and sensitivity to different types of noise. Several of these conjectures have since played an important role in the subsequent development of the area, most spectacularly in~ and~, where extremely precise results were obtained about the Fourier spectrum of the crossing event, and about the `dynamical percolation' process introduced by H\"aggstr\"om, Peres and Steif~ and (independently) by Benjamini, see~. To give another example, they made the following conjecture for Bernoulli bond percolation on the square lattice: even if you are told the status of \emph{all} the vertical edges, you still have very little information about the crossing event. This conjecture was proved by Garban, Pete and Schramm~\cite[Theorem~1.3]{GPS}, as a consequence of their very precise bounds on the Fourier spectrum. Note that in this theorem we are given a \emph{deterministic} set of edges (of density $1/2$), rather than a \emph{random} set of edges (of density $1 - o(1)$) as in the result stated above. In this paper, we will prove a similar result (also conjectured in~) in the setting of \emph{Voronoi percolation}: that knowing the point set (but not the colours of the cells) gives asymptotically no information about the crossing event. In order to state our main result precisely, we will need a few basic definitions. Consider a set $\eta$ of $n$ points in the square $S = [0,1]^2$, each chosen independently and uniformly at random. For each $u \in \eta$, define the Voronoi (or Dirichlet) cell in 1850, who used them in his work on quadratic forms, although they appear to have been introduced even earlier, by Kepler and (independently) Descartes, see~. The natural generalisation to $d$ dimensions was first studied by Voronoi~ in 1908.} of $u$ to be $$C(u) \, = \, \big\{ x \in [0,1]^2 \,:\, \| u - x \|_2 \le \| v - x \|_2 \textup{ for every } v \in \eta \big\},$$ and let $\omega \colon \eta \to \{-1,1\}$ be a uniformly random two-colouring of the points of $\eta$; we will call the points $u$ (and the associated cells $C(u)$) with $\omega(u) = 1$ `red' and those with $\omega(u) = -1$ `blue'. We say that there is a \emph{red horizontal crossing} of $S$ if there is a path from the left- to the right-hand side of $S$ that only intersects red cells, and write $H_S$ for the event that there exists such a red horizontal crossing of $S$. Note that $\Pr( H_S ) = 1/2$, by symmetry. We refer the reader who is unfamiliar with Voronoi percolation to~ for a more extensive introduction. The following theorem confirms (in a strong form) a conjecture of Benjamini, Kalai and Schramm~. \begin{thm} There exists $c > 0$ such that $$\Pr\left( \frac{1}{2} - n^{-c} \,\le\, \Pr\big( H_S \,|\, \eta \big) \,\le\, \frac{1}{2} + n^{-c} \right) \, \ge \, 1 - n^{-c}$$ for all sufficiently large $n \in \N$. \end{thm} Let $f^\eta \colon \{-1,1\}^\eta \to \{0,1\}$ be the function such that $f^\eta(\omega) = 1$ if and only if $H_S$ holds. The key new idea of the proof of Theorem~ is the following Efron-Steif type bound (see Theorem~, below) on the variance of the probability of the crossing event in terms of the influences of $f^\eta$, which can be viewed as a random Boolean function: \begin{equation} \Var\Big( \Pr\big( H_S \,|\, \eta \big) \Big) \, \le \, \sum_{m=1}^n \Ex\big[ \Inf_m( f^\eta )^2 \big]. \end{equation} Recall that the influence $\Inf_m(f_n)$ of the $m^{th}$ variable of a Boolean function $f_n \colon \{-1,1\}^n \to \{0,1\}$ is defined to be the expected absolute change in $f_n$ when the sign of the $m^{th}$ variable is flipped, i.e., $$\Inf_m(f_n) \, = \, \Pr\big( f_n(\omega) \ne f_n(\omega') \big),$$ where $\omega$ is chosen uniformly, and $\omega'$ is obtained from $\omega$ by flipping the $m^{th}$ variable. Benjamini, Kalai and Schramm~ proved that \begin{equation} \sum_{m = 1}^n \Inf_m(f_n)^2 \to 0 \; \textup{ as } n \to \infty \qquad \Rightarrow \qquad (f_n)_{n \in \N} \textup{ is noise sensitive,} \end{equation} and moreover introduced a technique (the `algorithm method', see below) which can often be used to bound $\sum_{m = 1}^n \Inf_m(f_n)^2$ when $f_n$ encodes crossing events in percolation models. We will use this method (or, more precisely, the `randomized' version of it developed by Schramm and Steif~), together with a new `box-crossing property' for quenched Voronoi percolation (see below), to bound^n \Inf_m(f^\eta)^2$ holds with high probability as $n \to \infty$.} $\sum_{m = 1}^n \Inf_m(f^\eta)^2$, and hence deduce Theorem~. As an immediate consequence of the proof outlined above, together with~\eqref{eq:BKSthm}, we also obtain the following theorem. Let us say that \emph{quenched Voronoi percolation is almost surely noise sensitive at criticality} if \begin{equation} \Ex\big[ f^\eta(\omega) f^\eta(\omega^\eps) \,|\, \eta \big] - \Ex\big[ f^\eta(\omega) \,|\, \eta \big] \Ex\big[ f^\eta(\omega^\eps) \,|\, \eta \big] \to 0 \end{equation} as $n \to \infty$ with probability 1 for every $\eps \in (0,1)$, where $\omega$ and $\omega^\eps$ are as defined above. \begin{thm} Quenched Voronoi percolation is almost surely noise sensitive at criticality. \end{thm} In fact, as a consequence of the Schramm-Steif method, we obtain a stronger result: that the noise sensitivity exponent for quenched Voronoi percolation is positive. This means that there exists a constant $c > 0$ such that~\eqref{eq:NSdef} holds even if $\eps = n^{-c}$. \begin{rmk} The word `quenched' refers to the fact that we are proving a statement which holds for almost all choices of $\eta$. The phrase `at criticality' refers to the fact that $\omega$ is chosen uniformly at random. We remind the reader that the critical probability of Voronoi percolation in the plane is $1/2$, as was proved by Bollob\'as and Riordan~. \end{rmk} We remark that Theorem~ is not the first result of this type for a continuum percolation model. Indeed, a similar theorem for the Poisson Boolean model was proved by the first three authors with Broman~, and the techniques introduced in that paper have recently been extended by the first three authors with Balister and Bollob\'as~ to the settings of (annealed) Voronoi percolation and the Poisson Boolean model with random radii. (In each case the point set $\eta$ is perturbed, together with the colours/radii.) We emphasize, however, that the techniques introduced in this paper are completely different from those used in~, where the method involved choosing the point set in two stages, and applying the algorithm method in the non-uniform setting. Indeed, none of the previously-introduced techniques seem to have any chance of working in the setting of quenched Voronoi percolation. As mentioned above, in order to use the algorithm method we will need to prove a 1-arm estimate that will follow from a quenched version of the box-crossing property for Voronoi percolation at criticality. This result gives bounds on the probability that a rectangle (of fixed aspect ratio) is crossed at criticality, and is an analogue of the celebrated results for bond percolation on $\Z^2$ of Russo~ and Seymour and Welsh~. Corresponding results have been obtained in various related settings, and obtaining such bounds is frequently a key step in the proof of various important applications, see e.g.~. In particular, an important breakthrough was made by Bollob\'as and Riordan~, who proved an RSW-type theorem for (annealed) Voronoi percolation, and used it to deduce that the critical probability for percolation is $1/2$. The full box-crossing property in the annealed setting was obtained only very recently, by the fourth author~. We remark that this result will play an important role in our proof of Theorem~, below. As above, we write $H_R$ for the event that there is a red horizontal crossing of $R$. \begin{thm}[The quenched box-crossing property for Voronoi percolation] For every rectangle $R \subset \RR^2$, there exists a constant $c > 0$ such that the following holds. Let $n \in \N$, let $\eta \subset R$ be a set of $n$ points, each chosen uniformly at random, and let $\omega \colon \eta \to \{-1,1\}$ be a uniform colouring. Then $$\Pr\Big( c < \Pr\big( H_R \,|\, \eta \big) < 1 - c \Big) \to \, 1$$ as $n \to \infty$. \end{thm} We remark that, moreover, for every $\gamma > 0$ there exists $c = c(\gamma,R) > 0$ such that $\Pr\big( H_R \,|\, \eta \big) \not\in (c,1-c)$ has probability at most $n^{-\gamma}$. An analogous theorem if $\eta$ is a Poisson point process in the plane (or in the half-plane) follows by exactly the same proof. We will prove Theorem~ in three steps. First, we will prove a weaker result for Voronoi percolation in the plane (see Theorem~): this says that there exists a constant $c > 0$ such that \begin{equation} \Pr\bigg( \Pr\big( H_R \,|\, \eta \big) \le \frac{1}{2^k} \bigg) \le \, (1 - c)^k \end{equation} for all sufficiently large $k$. We will then deduce an analogous statement for Voronoi percolation in a half-plane; somewhat surprisingly, the deduction is not trivial, and we will have to do some work to deal with the boundary effects (see Section~). Finally, we will use these results, together with the algorithm method (see Section~) and our Efron-Stein type inequality~\eqref{eq:varinf}, proved in Section~, to show (see Theorem~) that $$\Pr\big( H_R \,|\, \eta \big) \, \to \, \Ex\big[ \Pr\big( H_R \,|\, \eta \big) \big]$$ in probability, as $n \to \infty$. This result will imply both Theorem~ and Theorem~, using the box-crossing property for annealed Voronoi proved in~. The organisation of the rest of the paper is as follows. First, in Section~, we will bound the variance of the probability of the crossing event by the expected sum of the squares of the influences of $f^\eta$. We will do so by introducing a martingale, whose steps correspond to choosing the points of $\eta$ one-by-one, and bounding the variance of step $m$ in terms of the expectation of the square of the influence of the $m^{th}$ element of $\eta$, see Lemma~. Armed with this lemma, the claimed bound~\eqref{eq:varinf} follows easily. Second, in Section~, we will prove weak bounds for the crossing probabilities in quenched Voronoi percolation~\eqref{eq:weakRSW} in both the plane, and the half-plane. The key tools in our (surprisingly simple) proof will be the `box-crossing property' for annealed Voronoi percolation, proved in~, together with colour-switching. In particular, we would like to highlight Lemma~, which states that $$\Pr\big( H_R \,|\, \eta \big) \, = \, \Ex\big[ 2^{-X} \,|\, \eta \big],$$ where $X$ is the random variable which counts the number of vertex-disjoint vertical monochromatic crossings of $R$. Although this lemma, once stated, is easy to prove, we have found it to be extremely useful, and expect it to have many other applications. Finally, in Section~, we will complete the proof of the main theorems, by using the algorithm method of Benjamini, Kalai and Schramm~ and Schramm and Steif~ to bound the sum of the squares of the influences of $f^\eta$. (Indeed, once we are armed with the results of Section~, the required bound follows by simply repeating the method of~.) Combining this bound with the results of Section~, we obtain Theorem~. By~\eqref{eq:BKSthm}, we obtain Theorem~, and by the box-crossing property for annealed Voronoi percolation~, we obtain Theorem~. \section{Variance and influence} In this section, we will prove a somewhat surprising bound on the (typical) dependence of the crossing event on the point set $\eta$ in terms of the (expected) influences of the colours. Since we will need to use the results of this section in the proof of Theorem~, as well as that of Theorem~, we will work in the more general set-up of an arbitrary rectangle $R \subset \RR^2$, so let $\eta$ be a set of $n$ points in $R$, each of which is chosen independently and uniformly at random. We will write $f_R^\eta \colon \{-1,1\}^\eta \to \{0,1\}$ for the function that encodes whether or not there is a red horizontal crossing of $R$ in the corresponding Voronoi tiling. Recall that $$\Inf_m(f_R^\eta) \, := \, \Pr\left( f_R^\eta(\omega) \neq f_R^\eta(\omega') \,\big|\, \eta \right),$$ where $\omega'$ equals $\omega$ except on the $m^{th}$ element of $\eta$. \pagebreak The main result of this section is the following inequality, which is highly reminiscent of the well-known inequality of Efron and Stein~. \begin{thm} For every rectangle $R \subset \RR^2$, $$\Var\Big( \Pr\big( H_R \,|\, \eta \big) \Big) \, \le \, \sum_{m=1}^n \Ex\big[ \Inf_m( f_R^\eta )^2 \big].$$ \end{thm} Note that the following corollary is an immediate consequence of the above theorem and Chebychev's inequality. \begin{cor} Let $a(n) = \Ex\Big[ \sum_{m = 1}^n \Inf_m(f_R^\eta)^2 \Big]$. Then $$\Pr\Big( \big| \Pr( H_R \,|\, \eta ) - \Pr( H_R ) \big| \ge a(n)^{1/3} \Big) \, \le \, a(n)^{1/3}.$$ \end{cor} The proof of Theorem~ uses the following simple martingale $(q_m)_{m=0}^{n}$. Let us choose the elements of $\eta$ one-by-one, and let $\eta_m$ denote the $m^{th}$ element. Now write $$q^\eta \, = \, \Pr\big( H_R \,\big|\, \eta \big)$$ for the probability of such a crossing given $\eta$, and define $$q_m \, := \, \Ex\big[ q^\eta \,|\, \F_m \big],$$ where $\F_m$ denotes the $\sigma$-algebra generated by $\eta_1,\ldots,\eta_m$. \begin{obs} $\Var\big( q^\eta \big) = \, \ds\sum_{m=1}^n \Var\big( q_m - q_{m-1} \big).$ \end{obs} \begin{proof} Since $q^\eta - \Ex[q^\eta] = \sum_{m=1}^n (q_m - q_{m-1})$, it will suffice to show that $\Cov(q_i - q_{i-1},q_j - q_{j-1}) = 0$ for every $1 \le i < j \le n$. To see this, we simply condition on $\F_i$, which gives $$\Cov(q_i - q_{i-1},q_j - q_{j-1}) \, = \, \Ex\Big[ \Ex\big[ \big( q_i - q_{i-1} \big)\big( q_j - q_{j-1} \big) \,\big|\, \F_i \big] \Big] \, = \, 0,$$ since $\Ex\big[ q_j - q_{j-1} \,|\, \F_i \big] = 0$, and $q_i - q_{i-1}$ is determined by $\eta_1,\ldots,\eta_i$. \end{proof} By Observation~, the following lemma completes the proof of Theorem~. \begin{lemma} For every $1 \le m \le n$, $$\Var\big( q_m - q_{m-1} \big) \, \le \, \Ex\big[ \Inf_m\big( f_R^\eta \big)^2 \big]$$ almost surely. \end{lemma} \begin{proof} First observe that, since $\Ex\big[ q_m - q_{m-1} \,|\, \F_{m-1} \big] = 0$ almost surely, by the conditional variance formula we have \begin{equation} \Var\big( q_m - q_{m-1} \big) \, = \, \Ex\big[ \Var\big( q_m \,\big|\, \F_{m-1} \big) \big]. \end{equation} Now, let $\eta^-$ be obtained from $\eta$ by deleting $\eta_m$. Since $q^{\eta^-}$ does not depend on the $m^{th}$ element of $\eta$, it follows that \begin{align*} \Var\big( q_m \,\big|\, \F_{m-1} \big) & \, = \, \Var\Big( \Ex\big[ q^\eta \,|\, \F_m \big] - \Ex\big[ q^{\eta^-} \,|\, \F_{m-1} \big] \,\big|\, \F_{m-1} \Big)\\ & \, = \, \Var\Big( \Ex\big[ q^\eta - q^{\eta^-} \,|\, \F_m \big] \,\big|\, \F_{m-1} \Big). \end{align*} Now, since $\Var(X) \le \Ex[X^2]$ for every random variable $X$, this is at most $$\Ex\Big[ \Ex\big[ q^\eta - q^{\eta^-} \,|\, \F_m \big]^2 \,\big|\, \F_{m-1} \Big],$$ which is in turn at most $$\Ex\Big[ \big( q^\eta - q^{\eta^-} \big)^2 \,\big|\, \F_{m-1} \Big],$$ by Jensen's inequality. We make the following claim, which completes the proof. \medskip \noindent \textbf{Claim:} $|q^\eta - q^{\eta^-}| \le \Inf_m(f_R^\eta)$, almost surely. \begin{proof}[Proof of claim] Let us write $\omega^{+}$ (resp. $\omega^{-}$) for the element of $\{-1,1\}^n$ obtained from $\omega$ by setting the $m^{th}$ coordinate equal to $1$ (resp. $-1$). Also define $f_R^{\eta^-}(\omega)$ for $\omega \in \{-1,1\}^n$ in the obvious way, by ignoring the $m^{th}$ coordinate of $\omega$. We have \begin{align*} q^\eta - q^{\eta^-} & \, \le \, \Pr\big( f_R^\eta(\omega) > f_R^{\eta^-}(\omega) \,|\, \eta \big)\\ & \, \le \, \Pr\big( f_R^\eta(\omega^{+}) > f_R^{\eta}(\omega^{-}) \,|\, \eta \big) \, = \, \Inf_m(f_R^\eta), \end{align*} since $f_R^\eta$ is monotone (as a function on $\{-1,0,1\}^\eta$).^\eta \to \{0,1\}$ by setting $f_R^\eta(\omega) = 1$ if there is a red horizontal crossing in the Voronoi tiling defined by the set $\eta' = \{u \in \eta : \omega(u) \ne 0 \}$ with colouring $\omega' = \omega |_{\eta'}$. The function $f_R^\eta$ is monotone increasing since (from the point of view of the crossing event) a red point is better than no point, and no point is better than a blue point.} An identical calculation shows that $q^{\eta^-} - q^{\eta} \le \Inf_m(f_R^\eta)$, and so the claim follows. \end{proof} The lemma now follows since, by~\eqref{eq:var:stepone}, we have \begin{align*} \Var\big( q_m - q_{m-1} \big) & \, = \, \Ex\big[ \Var\big( q_m \,\big|\, \F_{m-1} \big) \big] \, \le \, \Ex\Big[ \Ex\big[ \big( q^\eta - q^{\eta^-} \big)^2 \,\big|\, \F_{m-1} \big] \Big] \\ & \, \le \, \Ex\Big[ \Ex\big[ \Inf_m\big( f_R^\eta \big)^2 \,\big|\, \F_{m-1} \big] \Big] \, = \, \Ex\big[ \Inf_m\big( f_R^\eta \big)^2 \big], \end{align*} as required. \end{proof} As noted above, Theorem~ follows immediately from Observation~ and Lemma~. \pagebreak \section{Crossing probabilities for quenched Voronoi percolation: weak bounds} In this section we will prove a weaker version of Theorem~ for quenched Voronoi percolation in the plane, and deduce the corresponding theorem when $\eta$ is a subset of the half-plane. Crucially, these results will be sufficient to deduce a bound on the probability of the one-arm event that is strong enough for our application of the Schramm-Steif method in Section~. \subsection{Crossing probabilities in the plane} In this section, $\eta$ will denote a Poisson process in the plane of intensity $1$, and $\omega \colon \eta \to \{-1,1\}$ will be a (uniform) random two-colouring of $\eta$. Recall that, given a rectangle $R$ with sides parallel to the axes, $H_R$ denotes the event that there is a red horizontal crossing of $R$ in the Voronoi tiling given by $\eta$, coloured by $\omega$. \begin{thm} For every rectangle $R \subset \RR^2$, there exists a constant $c > 0$, depending only on the aspect ratio of $R$, such that $$\Pr\bigg( \Pr\big( H_R \,|\, \eta \big) \le \frac{1}{2^k} \bigg) \le \, (1-c)^k$$ for all sufficiently large $k \in \N$. \end{thm} We begin by defining the following random variable, whose value depends on both $\eta$ and $\omega$, and which counts the maximum number of vertex-disjoint vertical crossings of $R$: \begin{multline*} X \, = \, X(\eta,\omega) \, := \, \max\big\{ m \in \N_0 \,: \, \textup{there exist $m$ vertex-disjoint, monochromatic}\\ \textup{vertical crossings $\{\gamma_1,\ldots,\gamma_m\}$ of $R$} \big\}. \end{multline*} The following lemma will be a key tool in our proof of Theorem~. \begin{lemma} For almost every $\eta$, $$\Pr\big( H_R \,|\, \eta \big) \, = \, \Ex\big[ 2^{-X} \,|\, \eta \big].$$ \end{lemma} As noted in the Introduction, we have found this lemma to be surprisingly powerful, and expect it to have many other applications. Despite this, the lemma is very easy to prove; indeed, it follows almost immediately from a basic (and well-known) fact about `colour-switching', see~\eqref{eq:colourswitching} below. For those readers who are unfamiliar with colour-switching, we will begin by giving a brief introduction. Consider the event that $X = k$, i.e., that there exist $k$, but not $k+1$, vertex-disjoint, monochromatic vertical crossings of $R$. The following standard algorithm provides a method of finding such paths. First, we discover the left-most monochromatic path (in the coloured Voronoi tiling of $R$ given by the pair $(\eta,\omega)$) starting from the left-most cell which intersects the lower side of the rectangle. If it reaches the top of the rectangle, then we add this monochromatic path to our collection; otherwise we discover the whole monochromatic component of our starting-point. In either case, we then discover the left-most monochromatic path entirely to the right of the already-discovered cells, starting from the next available cell on the lower side of $R$ (if it exists). Repeating this process until we reach the right-side of $R$, we obtain a collection $(\gamma_1,\ldots,\gamma_k)$ of disjoint monochromatic crossings. An important feature of the algorithm above is that it allows us to use so-called `colour-switching' arguments, see e.g.~. The crucial observation is that, for a given $\eta$, and a given collection of paths $(\gamma_1,\ldots,\gamma_k)$ obtained via the algorithm, there is a bijection between the set of colourings in which $\gamma_j$ is red, and those in which it is blue. Indeed, if we swap the colours of all cells that are on or to the right of $\gamma_j$, then the algorithm produces exactly the same set of paths. More generally, writing $\Pi \in \{-1,1\}^{k}$ for the sequence of colours of the paths $(\gamma_1,\ldots,\gamma_k)$, we have the following simple fact. For every $\sigma \in \{-1,1\}^k$, we have \begin{equation} \Pr\big( \Pi = \sigma \,\big|\, X = k, \, \eta \big) = \frac{1}{2^k} \end{equation} almost surely. Lemma~ is an almost immediate consequence of~\eqref{eq:colourswitching}. \begin{proof}[Proof of Lemma~] Observe that the event $H_R$ holds if and only if all monochromatic vertical paths are red. By~\eqref{eq:colourswitching}, it follows that \begin{align*} \Pr\big( H_R \,|\, \eta \big) & \,=\, \sum_{k = 0}^\infty \Pr\Big( \Pi = (1,\ldots,1) \,\big|\, X = k, \, \eta \Big) \Pr\Big( X = k \,\big|\, \eta \Big)\\ & \,=\, \sum_{k = 0}^\infty \frac{ \Pr(X = k \, | \, \eta )}{2^k}\, = \, \Ex\big[ 2^{-X} \,|\, \eta \big], \end{align*} as required. \end{proof} In order to conclude to obtain the proof of Theorem~, we only need to show that $X$ cannot be too large. This will be a consequence of three properties of annealed Voronoi percolation: the FKG and BK inequalities, and the box crossing property, proved recently by the fourth author~. An event $A$ that depends on the pair $(\eta,\omega)$ is said to be \emph{red-increasing} if removing a blue point or adding a red point cannot cause the status of $A$ to change from true to false. For a proof of the following lemma, see~\cite[Lemma~8.14]{BRbook}. \begin{lemma}[The FKG inequality for annealed Voronoi percolation] Let $A$ and $B$ be red-increasing events. Then $$\Pr(A \cap B ) \, \ge \, \Pr(A) \cdot \Pr(B),$$ and moreover the reverse inequality holds if $B$ is replaced by its complement $B^c$. \end{lemma} The following lemma was proved by van den Berg~; his proof, sketched below, can also be found in the PhD thesis of Joosten~\cite[Section~4.3]{J} (who also refers to van den Berg). The corresponding inequality in the discrete setting is a celebrated result of van den Berg and Kesten~. \begin{lemma}[The BK inequality for annealed Voronoi percolation] Let $A$ and $B$ be red-increasing events. Then $$\Pr(A \circ B) \, \le \, \Pr(A) \cdot \Pr(B),$$ where $A \circ B$ denotes the event that $A$ and $B$ occur disjointly. \end{lemma} \begin{proof} Let $\ol{B}$ denote the event corresponding to $B$ with colours reversed. We have \begin{align*} \Pr(A \circ B) & \, = \, \Ex\big[ \Pr( A \circ B \,|\, \eta) \big] \, \le \, \Ex\big[ \Pr( A \cap \ol{B} \,|\, \eta) \big]\\ & \, = \, \Pr( A \cap \ol{B} ) \, \le \, \Pr(A) \Pr(\ol{B}) \, = \, \Pr(A) \Pr(B), \end{align*} where the first inequality follows from Reimer's inequality~\cite[Theorem~1.2]{Reimer}, and the second inequality follows from Lemma~. \end{proof} Our final tool is the following result of Tassion~. \begin{lemma}[The box-crossing property for annealed Voronoi percolation] There exists a constant $c_0 > 0$, depending only on the aspect ratio of $R$, such that $$\Pr\big( H_R \big) > c_0.$$ \end{lemma} We are now ready to prove the weak box-crossing property for quenched Voronoi percolation in the plane. \begin{proof}[Proof of Theorem~] Given the pair $(\eta,\omega)$, define $X^+$ (resp.\@ $X^-$) to be the maximal number of disjoint red (resp.\@ blue) paths from top to bottom in $R$, so in particular $X = X^+ + X^-$. By Lemma~ and the BK inequality for annealed Voronoi percolation, there exists a constant $c_0 > 0$, depending only on the aspect ratio of $R$, such that $$\Pr\big( X^+ \ge k \big) \, \le \, (1-c_0)^k$$ for all $k \ge 0$, and similarly $\Pr( X^- \ge k ) \le (1-c_0)^k$. Since the events $\big\{ X^+ \ge i \big\}$ and $\big\{ X^- < j \big\}$ are both red-increasing, it follows by the FKG inequality that \begin{align} \Pr\big(X\ge k\big) & \, \le \, \sum_{i+j\ge k} \Pr\Big( \big\{ X^+ \ge i \big\} \cap \big\{ X^-\ge j \big\} \Big) \nonumber \\ & \, \le \, \sum_{i+j\ge k} (1-c_0)^{i+j} \, \le \, \frac{(1 - c)^k}{2} \end{align} for some constant $c > 0$, if $k$ is sufficiently large. Now, by Lemma~ we have $$\Pr\big( H_R \,|\, \eta \big) \, = \, \Ex\big[ 2^{-X} \,|\, \eta \big] \, \ge \, \left(\frac12\right)^{k-1}\Pr\big( X < k \,|\, \eta \big),$$ almost surely. Thus, by Markov's inequality and~\eqref{eq:6}, we obtain $$\Pr\bigg( \Pr\big( H_R \,|\, \eta \big) \le \left( \frac12\right)^{k} \bigg) \, \le \, \Pr\bigg( \Pr\big( X \ge k \,|\, \eta \big) \ge \frac12 \bigg) \, \le \, (1 - c)^k,$$ as required. \end{proof} \subsection{Quenched crossing probabilities in the half-plane} In order to bound one-arm events starting at points near the boundary of $R$, we will require a result analogous to Theorem~ for a Poisson point process $\eta$ of intensity $1$ restricted to the half-plane $$\HH \, := \, \big\{ (x,y) \in \RR^2 : x \ge 0 \big\}.$$ For each rectangle $R \subset \HH$ with sides parallel to the axes, let $H^*_R$ denote the event that there is a red horizontal crossing of $R$ in the Voronoi tiling of $\HH$ given by $\HH \cap \eta$, coloured by $\omega$. \begin{thm} For every rectangle $R \subset \HH$, there exists a constant $c > 0$, depending only on the aspect ratio of $R$, such that $$\Pr\bigg( \Pr\big( H^*_R \,|\, \eta \big) \le \frac{1}{2^k} \bigg) \le \, (1-c)^k,$$ for all sufficiently large $k \in \N$. \end{thm} The proof of Theorem~ is identical to that of Theorem~, except we will need to replace Lemma~ with the following bound in the annealed setting. \begin{lemma} For every rectangle $R$, there exists a constant $c_1 > 0$, depending only on the aspect ratio of $R$, such that $$\Pr\big( H^*_R \big) > c_1.$$ \end{lemma} When the rectangle $R$ is sufficiently far from the boundary, Lemma~ follows from Lemma~, since the Voronoi tilings of $R$ is (with high probability) the same in both cases. We begin with an easy lemma that makes this statement precise. \begin{lemma} Given $\lambda > 0$, let $L > 0$ be sufficiently large. Let $\eta$ be a Poisson point process in the plane of intensity 1, and let $R$ be a $\lambda L \times L$ rectangle with $$\min\{ x : (x,y) \in R \} \ge (\log L)^{2/3}.$$ Then the Voronoi tilings of $R$ induced by $\eta$ and by $\HH \cap \eta$ are non-identical with probability at most $1/L^3$. \end{lemma} \begin{proof} If the Voronoi tilings of $R$ induced by $\eta$ and by $\HH \cap \eta$ are non-identical, then there must be a point $u \in R$ such that the closest point of $\eta$ to $u$ lies outside $\HH$. This implies that there is an empty ball of radius $(\log L)^{2/3}$ centred in $R$, the probability of which (by standard properties of Poisson processes) is super-polynomially small in $L$. \end{proof} We will assume from now on that $R$ is a $\lambda L \times L$ rectangle with $L$ sufficiently large, and such that the left-hand side of $R$ is on the line $x = 0$. (Note that, by choosing $c_1$ sufficiently small, this may be assumed without loss of generality.) The idea of the proof of Lemma~ is as follows. Set $\ell = (\log L)^{2/3}$, and partition\footnote{We can deal with rounding issues by increasing $\ell$ slightly if necessary.} the set $\{ (x,y) \in R : 0 \le x \le \ell \}$ into $L/\ell$ squares $S_1,\ldots,S_{L/\ell}$. Now observe that if there is a red horizontal crossing of the rectangle $R' = \{ (x,y) \in R : x \ge \ell \}$, but no red horizontal crossing of $R$, then there must exist a square $S_j$ such that the following `3-arm event' holds (see Figure~1, below). \begin{defn} For each $1 \le j \le L/\ell$, let $A^{(3)}(j)$ denote the event that the following hold: \begin{itemize} \item[$(a)$] there is a red path from $S_j$ to the right-hand side of $R$ that is contained in $R'$. \item[$(b)$] there are blue paths from $S_j$ to the top and the bottom of $R$ that are contained in~$R$. \end{itemize} \end{defn} By Lemma~ and the observations above, we have \begin{equation} \Pr\big(H^*_{R}\big) \,\ge\, \Pr\big(H_{R'}\big) - \frac{1}{L^3} -\, \sum_{j=1}^{L/\ell} \Pr\big( A^{(3)}(j) \big). \end{equation} Thus, by Lemma~, it will suffice to prove the following lemma. \begin{lemma} There exists $c > 0$ such that $$\sum_{j=1}^{L/\ell} \Pr\big( A^{(3)}(j) \big) \le L^{-c}$$ for all sufficiently large $L > 0$. \end{lemma} In the proof of Lemma~, we will use the following two events: \begin{itemize} \item $A^{(1)}(j)$ denotes the event that there is a red path from $S_j$ to the right-hand side of $R$ that is completely contained inside $R'$. \item $A^{(2)}(j)$ denotes the event that there is a red path from $S_j$ to the top of $R$, and a blue path from $S_j$ to the bottom of $R$, both of which are completely contained inside~$R$. \end{itemize} We are now ready to prove Lemma~. \begin{proof}[Proof of Lemma~] We first claim that, for each $j \in [L/\ell]$, we have \begin{equation} \Pr\big( A^{(3)}(j) \big) \, \le \, \Pr\big( A^{(1)}(j) \big) \cdot \Pr\big( A^{(2)}(j) \big).\end{equation} To prove~\eqref{eq:A3vsA1A2}, we apply colour-switching and the BK inequality for annealed Voronoi percolation (Lemma~). Indeed, the three paths in Definition~ are of alternating colours, so must be vertex-disjoint. Moreover, if such vertex-disjoint, monochromatic paths exist, then by colour-switching they are each equally likely to be red or blue. by switching the colours of all cells that are on or to the right of $\gamma_j$ for each $j \in \{1,2,3\}$. The left-most triple is obtained by choosing $\gamma_j$ to be the left-most monochromatic path to the top/right/bottom of $R$ that is entirely to the right of $\gamma_{j-1}$.} Thus, letting $B^{(2)}(j)$ denote the event that there are vertex-disjoint red paths from $S_j$ to the top and bottom of $R$, both of which are completely contained inside $R$, we have $$\Pr\big( A^{(3)}(j) \,|\, \eta \big) \, = \, \Pr\big( A^{(1)}(j) \circ B^{(2)}(j) \,|\, \eta \big) \quad \textup{and} \quad \Pr\big( A^{(2)}(j) \,|\, \eta \big) \, = \, \Pr\big( B^{(2)}(j) \,|\, \eta \big)$$ for almost all $\eta$. Taking expectation over $\eta$, and noting that $A^{(1)}(j)$ and $B^{(2)}(j)$ are both red-increasing events, we have $$\Pr\big( A^{(3)}(j) \big) \, \le \, \Pr\big( A^{(1)}(j) \big) \cdot \Pr\big( B^{(2)}(j) \big) \, = \, \Pr\big( A^{(1)}(j) \big) \cdot \Pr\big( A^{(2)}(j) \big)$$ by the BK inequality, as claimed. By~\eqref{eq:A3vsA1A2}, the lemma is an immediately consequence of the following two claims. \medskip \noindent \textbf{Claim 1:} $\Pr\big( A^{(1)}(j) \big) \le L^{-2c}$ for some constant $c > 0$. \begin{proof}[Proof of claim] By Lemma~, the probability that the Voronoi tiling of $R'$ by $\HH \cap \eta$ differs from that by $\eta$ is at most $1/L^3$. The claim therefore follows from the corresponding statement in the plane, which is a standard consequence of Lemma~, see~\cite[Theorem~3]{T}. \end{proof} \medskip \noindent \textbf{Claim 2:} $\ds\sum_{j = 1}^{L/\ell} \Pr\big( A^{(2)}(j) \big) \le 2^{C\ell}$ for some constant $C > 0$. \begin{proof}[Proof of claim] For each $j \in [L/\ell]$, choose a (distinct) cell $u(j) \in \eta$ which intersects both $S_j$ and the left-hand side of $R$. Consider the event $E(j)$, illustrated in Figure~2, that there is a red path from $u(j)$ to the top of $R$, and a blue path from the cell $u'(j)$ immediately below $u(j)$ to the bottom of $R$, both of which are completely contained inside $R$. We claim that \begin{equation} \Pr\big( A^{(2)}(j)\big) \le \, 2^{C\ell} \cdot \Pr\big( E(j)\big) \end{equation} for each $j \in [L/\ell]$. To prove this, we simply use brute force to tunnel from the boundary of $S_j$ to $u(j)$. Indeed, as long as, for every pair of points $\{v,w\}$ on the boundary of $S_j$, there are vertex-disjoint paths (in the Voronoi tiling of $S_j$) from $u(j)$ to $v$ and from $u'(j)$ to $w$, each of length at most $C\ell/2$, then we can connect $u(j)$ and $u'(j)$ to the endpoints of the paths guaranteed by the event $A^{(2)}(j)$ with probability at least $2^{-C\ell}$. But such paths exist unless the Poisson process is much denser in $S_j$ than one would expect, and this occurs with probability that is super-polynomially small in $L$. Finally, simply note that, since at most one of the events $E(j)$ can occur, we have $$\sum_{j = 1}^{L/\ell} \Pr\big( A^{(2)}(j)\big) \le \, 2^{C\ell} \sum_{j = 1}^{L/\ell} \Pr\big( E(j) \big) \le \, 2^{C\ell}$$ as claimed. \end{proof} Combining Claims~1 and~2 with~\eqref{eq:A3vsA1A2}, and recalling that $\ell = (\log L)^{2/3}$, we obtain $$\sum_{j=1}^{L/\ell} \Pr\big( A^{(3)}(j) \big) \, \le \, \sum_{j=1}^{L/\ell} \Pr\big( A^{(1)}(j) \big) \cdot \Pr\big( A^{(2)}(j) \big) \, \le \, L^{-2c} \cdot 2^{C\ell} \, \le \, L^{-c}$$ if $L$ is sufficiently large, as required. \end{proof} For completeness, let us spell out how to deduce Lemma~ from Lemmas~ and~. \begin{proof}[Proof of Lemma~] By~\eqref{eq:3}, and Lemmas~ and~, we have $$\Pr\big(H^*_{R}\big) \,\ge\, \Pr\big(H_{R'}\big) - \frac{1}{L^3} -\, \sum_{j=1}^{L/\ell} \Pr\big( A^{(3)}(j) \big) \,\ge\, c_0 - \frac{1}{L^3} - L^{-c} \, \ge \, c_1$$ if $L$ is sufficiently large, as required. \end{proof} As noted above, Theorem~ follows by repeating the proof of Theorem~, using Lemma~ in place of Lemma~. \subsection{A bound on the probability of the 1-arm event in a rectangle} To finish the section, let us deduce the following proposition from Theorems~ and~. Let $R \subset \RR^2$ be a rectangle of area $n$, and let $\eta \subset R$ be a set of $n$ points, each chosen uniformly at random, and let $\omega$ be a uniform colouring of $\eta$. (Thus, the distribution of $\eta$ inside $R$ is very close to that of a Poisson process of intensity 1.) Given $u \in R$ and $d > 0$, we write $M(u,d)$ for the event (depending on $\eta$ and $\omega$) that there is a monochromatic path from $u$ to some point of $R$ at $\ell_2$-distance $d$ from $u$. We will use the following result in Section~ in order to bound the `revealment' of our randomized algorithm, and hence to deduce our main theorems. Since the deduction of this result from the box-crossing property is standard, we will be fairly brief with the details. \begin{prop} For every $\gamma > 0$, there exists $\eps > 0$ such that the following holds. Suppose that $d = d(n) \to \infty$ as $n \to \infty$ and let $u \in R$. Then $$\Pr\Big( \Pr\big( M(u,d) \,|\, \eta \big) \ge d^{-\eps} \Big) \, \le \, d^{-\gamma}$$ for all sufficiently large $n \in \N$. \end{prop} \begin{proof} Fix a point $u \in R$, and for each $j \in \N$, let $\AA_j$ denote the square annulus, centred on $u$, with inner side-length $7^j$ and outer side-length $3 \cdot 7^j$. Let $O_j$ denote the event that there is a blue circuit around the (partial) annulus $\AA_j \cap R$. (This circuit must either be closed, or have both its endpoints on the boundary of $R$; in either case, it must separate $u$ from the exterior of $\AA_j$.) Let $$J = \big\{ j \in \N \,:\, \sqrt{d} \le 7^{j+1} \le d \big\},$$ and consider the collection of annuli $\C = \{ \AA_j : j \in J\}$. Roughly speaking, we will show that either at least half the annuli in $\C$ contain an `unusual' collection of points, or the probability that none of the events $O_j$ occurs is at most $d^{-\eps}$. Let $c$ be the constant in Theorems~ and~, and fix a large constant $k \in \N$ (depending on~$c$ and~$\gamma$). For each $j \in J$, let $D^{(1)}_j$ denote the event (depending on $\eta$) that $\Pr\big( O_j \,|\, \eta \big) > 2^{-4k}$, and let $D^{(2)}_j$ denote the event that for every $z \in \AA_j$, there exists some point $x \in \eta$ at distance at most $\log d$ from $z$. Define $$D_j \, := \, D^{(1)}_j \cap D^{(2)}_j,$$ and observe that the events $D_j$ are independent. We claim that \begin{equation} \Pr(D_j) \, \ge \, 1 - 5(1-c)^k \end{equation} if $d$ is sufficiently large. To prove~\eqref{eq:Djbound}, note first (cf. Lemma~) that $$\Pr\big( D^{(2)}_j \big) \, \to \, 1$$ as $d \to \infty$. Next, observe that, by the FKG inequality and Theorems~ and~, we have $$\Pr\Big( \Pr\big( O_j \,|\, \eta \big) \le 2^{-4k} \Big) \, \le \, 4(1-c)^k$$ for all sufficiently large $k \in \N$. Thus $\Pr(D_j^c) \le 4(1-c)^k + o(1) \le 5(1-c)^k$, as claimed. Now, let $D^*$ denote the event that $D_j$ holds for at least half of the elements $j \in J$, and observe that $$\Pr(D^*) \, \ge \, 1 - 2^{|J|} \left( 5(1-c)^k \right)^{|J|/2} \ge \, 1 - d^{-\gamma},$$ since $|J| = \Omega\big( \log d \big)$, and $k$ was chosen sufficiently large in terms of $\gamma$ and $c$. But for those $\eta$ such that $D^*$ holds, we have $$\Pr\big( M(u,d) \,|\, \eta \big) \, \le \, \Pr\bigg( \bigcap_{j \in J} O_j^c \,\big|\, \eta \bigg) \, = \, \prod_{j \in J} \Pr\big( O_j^c \,|\, \eta \big) \, \le \, \big( 1 - 2^{-4k} \big)^{|J|/2} \, \le \, d^{-\eps},$$ for some $\eps > 0$, as required. \end{proof} \section{The proof of the main theorems} In this section we will use the algorithm method of Benjamini, Kalai and Schramm~ and Schramm and Steif~, together with the results proved in the previous sections, in order to bound the sum of the squares of the influences, and hence deduce the following theorem. Throughout this section, let us fix a rectangle $R \subset \RR^2$. \begin{thm} There exists $c = c(R) > 0$ such that the following holds. Let $\eta \subset R$ be a set of~$n$ points, each chosen uniformly at random, and let $\omega \colon \eta \to \{-1,1\}$ be a uniform colouring. Then $$\Pr\Big( \big| \Pr( H_R \,|\, \eta ) - \Pr( H_R ) \big| \ge n^{-c} \Big) \le n^{-c}$$ for all sufficiently large $n \in \N$. \end{thm} Theorem~ follows immediately from Theorem~ by taking $R = [0,1]^2$, and recalling that in this case $\Pr(H_R) = 1/2$. Note that Theorem~ also follows from Theorem~, together with the RSW theorem for annealed Voronoi percolation, which implies that $\Pr(H_R)$ is bounded away from 0 and 1 uniformly in $n$. The randomised algorithm method was introduced in~ and used there to show that the are exceptional times in dynamical site percolation on the triangular lattice. Since this method is well-known, and our application is rather standard, we shall be somewhat less careful with the details than in earlier sections, focusing instead on conveying the main ideas. Given a randomized algorithm $\A$ that determines a function $f_n \colon \{-1,1\}^n \to \{0,1\}$, define the revealment of $\A$ to be $$\delta_\A(f_n) \, := \, \max_{i \in [n]} \Pr\big( i \textup{ is queried by } \A \big).$$ Schramm and Steif~ proved a very powerful bound on the Fourier coefficients of a real-valued function on the hypercube $\{-1,1\}^n$ in terms of the revealment of any randomized algorithm that determines $f$. For monotone Boolean functions their result easily implies the following theorem, which will be sufficient for our purposes. \begin{thm}[Schramm and Steif, 2010] Given a monotone function $f \colon \{-1,1\}^n \to \{0,1\}$, and a randomized algorithm $\A$ that determines $f$, we have $$\sum_{m = 1}^n \Inf_m(f)^2 \, \le \, \delta_\A(f).$$ \end{thm} By Theorems~ and~, it only remains to show that, with probability at least $1 - n^{-c}$, there exists an algorithm that determines $f_R^\eta$ with revealment that is polynomially small in $n$. (Recall that $f_R^\eta \colon \{-1,1\}^\eta \to \{0,1\}$ denotes the function such that $f_R^\eta(\omega) = 1$ if and only if $H_R$ occurs.) The algorithm we will use is essentially the same as that introduced in~, so we shall describe it in an intuitive (and therefore slightly non-rigorous) fashion, and refer the reader to the original paper for the details. \begin{SSalg} Let $\A$ be the algorithm that, given $\eta$, queries bits of $\omega$ as follows: \begin{itemize} \item[1.] Choose a point $x$ in the middle third of the left-hand side of $R$ uniformly at random. \item[2.] Explore the boundary between red and blue cells, with red on the left, starting from $x$. Here we place boundary conditions as follows: the left-hand side of $R$ is red above $x$, and blue below, and the bottom of $R$ is also blue. If this path: \begin{itemize} \item[$(a)$] reaches the right-hand side of $R$, then $f_R^\eta(\omega) = 1$. \item[$(b)$] reaches the bottom of $R$, and ends at the top, then $f_R^\eta(\omega) = 0$. \item[$(c)$] ends at the top of $R$ without reaching the bottom, then go to step 3. \end{itemize} \item[3.] Explore the boundary between red and blue cells, with red on the right, starting from $x$. Here we reverse the boundary conditions, i.e., the left-hand side of $R$ is blue above $x$, and red below, and the top of $R$ is also blue. If this path: \begin{itemize} \item[$(a)$] reaches the right-hand side of $R$, then $f_R^\eta(\omega) = 1$. \item[$(b)$] otherwise $f_R^\eta(\omega) = 0$. \end{itemize} \end{itemize} \end{SSalg} Note that we only query bits as needed, i.e., we query those vertices whose cell we meet along one of our paths. Still following~, this allows us to immediately bound the revealment of $\A$ as follows. Recall that, given $u \in R$ and $d > 0$, we write $M(u,d)$ for the event that there is a monochromatic path from $u$ to some point of $R$ at $\ell_2$-distance $d$ from $u$. \begin{lemma} Let $\A$ be the Schramm-Steif randomized algorithm. Then $$\delta_\A(f_R^\eta) \, \le \, \max_{u \in \eta} \Pr\Big( M\big( u, n^{-1/4} \big) \,\big|\, \eta \Big) + O\big( n^{-1/4} \big)$$ almost surely. \end{lemma} \begin{proof}[Sketch of the proof] Let $u \in \eta$, and consider the probability that $u$ is queried by $\A$. First, note that the probability that the random start-point $x$ is within distance $n^{-1/4}$ of $u$ is $O(n^{-1/4})$. But if the distance between $u$ and $x$ is greater than $n^{-1/4}$, and $u$ is nonetheless queried by $\A$, then the event $M\big( u, n^{-1/4} \big)$ holds. \end{proof} To bound the revealment of $\A$, it will therefore suffice to bound the probability of the event $M\big( u, n^{-1/4} \big)$. This is an immediate consequence of Proposition~. \begin{lemma} For every $\gamma > 0$, there exists $c > 0$ such that $$\Pr\Big( \max_{u \in \eta} \Pr\Big( M\big( u, n^{-1/4} \big) \,\big|\, \eta \Big) \ge n^{-c} \Big) \, \le \, \frac{1}{n^\gamma}$$ for all sufficiently large $n \in \N$. \end{lemma} \begin{proof} Renormalizing $R$ to have area $n$, we find that distances are multiplied by $\Theta\big( \sqrt{n} \big)$, and so we may apply Proposition~ with $d = \Theta\big( n^{1/4} \big)$. The claimed bound now follows from the proposition, using the union bound over points $u \in \eta$. \end{proof} We are now ready to deduce our bound on the sum of the squares of the influences, and hence (by Theorem~ and the results of~ and~) our main theorems. \begin{lemma} For every $\gamma > 0$, there exists $c > 0$ such that $$\Pr\bigg( \sum_{m = 1}^n \Inf_m\big( f_R^\eta \big)^2 \ge n^{-c} \bigg) \, \le \, \frac{1}{n^\gamma}$$ for all sufficiently large $n \in \N$. \end{lemma} \begin{proof} This follows immediately from Theorem~, together with Lemmas~ and~. Indeed, we have $$\sum_{m = 1}^n \Inf_m(f_R^\eta)^2 \, \le \, \delta_\A(f_R^\eta) \, \le \, n^{-c}$$ with probability at least $1 - n^{-\gamma}$, as required. \end{proof} We can now deduce Theorem~. \begin{proof}[Proof of Theorem~] By Corollary~ and Lemma~ (applied with $\gamma = 1$, say), we have $$\Pr\Big( \big| \Pr( H_R \,|\, \eta ) - \Pr( H_R ) \big| \ge n^{-c} \Big) \, \le \, n^{-c}$$ for some $c > 0$, and all sufficiently large $n \in \N$, as required. \end{proof} Finally, let us note that the theorems stated in the Introduction all follow easily. \begin{proof}[Proof of Theorem~] As noted above, this is an immediate corollary of Theorem~. Indeed, simply set $R = [0,1]^2$ and observe that $\Pr(H_R) = 1/2$. \end{proof} Next we deduce that quenched Voronoi percolation is noise sensitive at criticality. We remark that, although we do not give the details here, it is a standard consequence of the method of~ that a stronger statement holds: That~\eqref{eq:NSdef} holds even if $\eps = n^{-c}$ for some (sufficiently small) constant $c > 0$. \begin{proof}[Proof of Theorem~] The theorem follows immediately from Lemma~, together with the Benjamini-Kalai-Schramm Theorem (see~\eqref{eq:BKSthm}). \end{proof} Finally, we prove the quenched box-crossing property for Voronoi percolation. \begin{proof}[Proof of Theorem~] This follows from Theorem~, together with the RSW theorem for annealed Voronoi percolation, which was proved in~. Indeed, for every rectangle $R$, there exists $c_0 > 0$ such that $$c_0 \, < \, \Pr(H_R) \, < \, 1 - c_0,$$ see~\cite[Theorem~3]{T}. Hence, by Theorem~, we have, for some $c > 0$, \begin{equation} \Pr\Big( c < \Pr\big( H_R \,|\, \eta \big) < 1 - c \Big) \, \ge \, 1 - n^{-c} \end{equation} for all sufficiently large $n \in \N$. \end{proof} Note that, by partitioning $R$ into a bounded number of strips and taking $c$ sufficiently small, inequality~\eqref{eq:lastequation} implies that the probability of the event $\Pr\big( H_R \,|\, \eta \big) \not\in (c,1-c)$ can be made smaller than any given polynomial, as claimed. \section*{Acknowledgements} The authors would like to thank Rob van den Berg for allowing us to include his proof of the BK inequality for annealed Voronoi percolation. The first author would also like to thank Elchanan Mossel and G\'abor Pete for encouraging discussions, and the second and third authors would like to thank Paul Balister and B\'ela Bollob\'as for a number of very interesting conversations about quenched Voronoi percolation. \begin{thebibliography}{99} \bibitem{ABBGM} D. Ahlberg, P. Balister, B. Bollob\'as, S. Griffiths and R. Morris, Noise sensitivity in Voronoi percolation and the Gilbert model with random radii, in preparation. \bibitem{ABGM} D. Ahlberg, E. Broman, S. Griffiths and R. Morris, Noise sensitivity in continuum percolation, \emph{Israel J. Math.}, \textbf{201} (2014), 847--899. \bibitem{A} K.S.~Alexander, The RSW theorem for Continuum Percolation and the CLT for Euclidean minimal spanning trees, \emph{Ann. Appl. Probab.}, \textbf{6} (1996), 466--494. \bibitem{BN} V. Beffara and P. Nolin, On monochromatic arm exponents for critical $2$D percolation, \emph{Ann. Prob.}, \textbf{40} (2012), 1286--1304. \bibitem{BKS} I.~Benjamini, G.~Kalai, and O.~Schramm, Noise sensitivity of Boolean functions and applications to percolation, \emph{Inst. Hautes Etudes Sci. Publ. Math.}, \textbf{90} (1999), 5--43. \bibitem{B} J. van den Berg, personal communication. \bibitem{BK} J. van den Berg and H. Kesten, Inequalities with applications to percolation and reliability, \emph{J. Appl. Prob.}, \textbf{22} (1985), 556--569. \bibitem{BRbook} B.~Bollob\'as and O.~Riordan, Percolation, Cambridge University Press, 2006. \bibitem{BR} B.~Bollob\'as and O.~Riordan, The critical probability for random Voronoi percolation in the plane is $1/2$, \emph{Prob. Theory Rel. Fields}, \textbf{136} (2006), 417--468. \bibitem{D} G.L. Dirichlet, \"Uber die Reduktion der positiven quadratischen Formen mit drei unbestimmten ganzen Zahlen, \emph{J. Reine Angew. Math.}, \textbf{40} (1850), 209--227. \bibitem{DHN} H. Duminil-Copin, C. Hongler and P. Nolin, Connection probabilities and RSW-type bounds for the two-dimensional FK Ising model, \emph{Comm. Pure App. Math.}, \textbf{64} (2011), 1165--1198. \bibitem{ES} B. Efron and C. Stein, The Jackknife Estimate of Variance, \emph{Ann. Statist.}, \textbf{9} (1981), 586--596. \bibitem{GPS} C.~Garban, G.~Pete and O.~Schramm, The Fourier spectrum of critical percolation, \emph{Acta Math.}, \textbf{205} (2010), 19--104. \bibitem{GS} C.~Garban and J.E.~Steif, Noise Sensitivity of Boolean Functions and Percolation, Cambridge University Press, 2014. \bibitem{HPS} O. H\"aggstr\"om, Y. Peres and J. Steif, Dynamical Percolation, \emph{Ann. l'Inst. Henri Poincar\'e, Prob. Stat.}, \textbf{33} (1997), 497--528. \bibitem{J} M. Joosten, Random fractals and scaling limits in percolation, PhD thesis, Vrije Universiteit Amsterdam, 2012. Available at http://hdl.handle.net/1871/32792. \bibitem{LP} T.M. Liebling and L. Pournin, Voronoi Diagrams and Delaunay Triangulations: Ubiquitous Siamese Twins, \emph{Documenta Math.}, Extra Vol. ISMP, (2012) 419--431. \bibitem{LS} E.~Lubetzky and A.~Sly, Critical Ising on the Square Lattice Mixes in Polynomial Time, \emph{Comm. Math. Phys.}, \textbf{313} (2012), 815--836. \bibitem{Reimer} D. Reimer, Proof of the van den Berg--Kesten Conjecture, \emph{Combin. Prob. Computing}, \textbf{9} (2000), 27--32. \bibitem{Roy} R. Roy, The Russo-Seymour-Welsh theorem and the equality of critical densities and the `dual' critical densities for continuum percolation on $\RR^2$, \emph{Ann. Prob.}, \textbf{} (1990), 1563--1575. \bibitem{R} L.~Russo, A note on percolation, \emph{Z. Wahrscheinlichkeitstheorie Verw. Geb.}, \textbf{43} (1978), 39--48. \bibitem{SS} O.~Schramm and J.E.~Steif, Quantitative noise sensitivity and exceptional times for percolation, \emph{Ann. Math.}, \textbf{171} (2010), 619--672. \bibitem{SW} P.D.~Seymour and D.J.A.~Welsh, Percolation probabilities on the square lattice, \emph{Ann. Discrete Math.}, \textbf{3} (1978), 227--245. \bibitem{Sm} S. Smirnov, Critical percolation in the plane: conformal invariance, Cardy's formula, scaling limits, \emph{C.~R.~Acad. Sci., Ser. I, Math.}, \textbf{333} (2001), 239--244. \bibitem{S} J.E.~Steif, A survey on dynamical percolation, Fractal Geometry and Stochastics IV, Progress in Probability, \textbf{61}, pp. 145--174, Birkhauser, 2009. \bibitem{T} V.~Tassion, Crossing probabilities for Voronoi percolation, submitted. \bibitem{V} G.F.~Voronoi, Nouvelles applications des param\`etres continus \`a la th\'eorie des formes quadratiques, \emph{J. Reine Angew. Math.}, \textbf{133} (1908), 97--178. \end{thebibliography} |
1501.04079 | Title: Topology and convexity in the space of actions modulo weak equivalence
Abstract: We analyse the structure of the quotient $\mathrm{A}_\sim(\Gamma,X,\mu)$ of
the space of measure-preserving actions of a countable discrete group by the
relation of weak equivalence. This space carries a natural operation of convex
combination. We show that the convex structure of
$\mathrm{A}_\sim(\Gamma,X,\mu)$ is compatible with the topology, and as a
consequence deduce that $\mathrm{A}_\sim(\Gamma,X,\mu)$ is path connected.
Using ideas of Tucker-Drob we are able to give a complete description of the
topological and convex structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for
amenable $\Gamma$ by identifying it with the simplex of invariant random
subgroups. In particular we conclude that $\mathrm{A}_\sim(\Gamma,X,\mu)$ can
be represented as a compact convex subset of a Banach space if and only if
$\Gamma$ is amenable. We consider the space $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$
of stable weak equivalence classes and show that is always a compact convex
subset of a Banach space. For a free group $\mathbb{F}_N$, we show that if one
restricts to the compact convex set $\mathrm{FR}_{\sim_s}(\mathbb{F}_N,X,\mu)
\subseteq \mathrm{A}_{\sim_s}(\mathbb{F}_N,X,\mu)$ of the stable weak
equivalence classes of free actions, the extreme points are dense in
$\mathrm{FR}_{\sim_s}(\mathbb{F}_N,X,\mu)$.
Body: \title{Topology and convexity in the space of actions modulo weak equivalence} \author{Peter Burton} \date{\today} \maketitle \begin{abstract} We analyse the structure of the quotient $\mathrm{A}_\sim(\Gamma,X,\mu)$ of the space of measure-preserving actions of a countable discrete group by the relation of weak equivalence. This space carries a natural operation of convex combination. We introduce a variant of an abstract construction of Fritz which encapsulates the convex combination operation on $\mathrm{A}_\sim(\Gamma,X,\mu)$. This formalism allows us to define the geometric notion of an extreme point. We also discuss a topology on $\mathrm{A}_\sim(\Gamma,X,\mu)$ due to Abert and Elek in which it is Polish and compact, and show that this topology is equivalent others defined in the literature. We show that the convex structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ is compatible with the topology, and as a consequence deduce that $\mathrm{A}_\sim(\Gamma,X,\mu)$ is path connected. Using ideas of Tucker-Drob we are able to give a complete description of the topological and convex structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for amenable $\Gamma$ by identifying it with the simplex of invariant random subgroups. In particular we conclude that $\mathrm{A}_\sim(\Gamma,X,\mu)$ can be represented as a compact convex subset of a Banach space if and only if $\Gamma$ is amenable. In the case of general $\Gamma$ we prove a Krein-Milman type theorem asserting that finite convex combinations of the extreme points of $\mathrm{A}_\sim(\Gamma,X,\mu)$ are dense in this space. We also consider the space $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ of stable weak equivalence classes and show that it can always be represented as a compact convex subset of a Banach space. In the case of a free group $\mathbb{F}_N$, we show that if one restricts to the compact convex set $\mathrm{FR}_{\sim_s}(\mathbb{F}_N,X,\mu) \subseteq \mathrm{A}_{\sim_s}(\mathbb{F}_N,X,\mu)$ consisting of the stable weak equivalence classes of free actions, then the extreme points are dense in $\mathrm{FR}_{\sim_s}(\mathbb{F}_N,X,\mu)$. \end{abstract} \section{Introduction.} By a probability space we mean a standard Borel space $Y$ with a Borel probability measure $\nu$. If $\nu$ is nonatomic, we say the pair $(Y,\nu)$ is a standard probability space. If $\nu$ is nonatomic then $Y$ must be uncountable and thus by Theorem 17.41 in every standard probability space is isomorphic to the unit interval with Lebesgue measure. Let $\Gamma$ be a countable discrete group. By a measure-preserving action of $\Gamma$ on $(Y,\nu)$ we mean a Borel action $a: \Gamma \times Y \to Y$ which preserves the measure $\nu$. We write $\Gamma \curvearrowright^a (Y,\nu)$. In accordance with the standard conventions of ergodic theory, we identify two actions which agree almost everywhere. Thus a measure-preserving action of $\Gamma$ on $(Y,\nu)$ is equivalently a homomorphism from $\Gamma$ into the group $\mathrm{Aut}(Y,\nu)$ of measure-preserving automorphisms of $(Y,\nu)$, where again two such automorphisms are identified if they agree almost everywhere.\\ \\ We fix a standard probability space $(X,\mu)$ throughout the remainder of the paper. As in we can define the Polish space $\mathrm{A}(\Gamma,X,\mu)$ of measure-preserving actions of $\Gamma$. Kechris defines the following relation of weak containment among measure-preserving actions of $\Gamma$, by analogy with the standard notion of weak containment for representations. \begin{definition} If $\Gamma \curvearrowright^a (X,\mu)$ and $\Gamma \curvearrowright^b (Y,\nu)$ are measure-preserving actions of $\Gamma$ on probability spaces, we say $a$ is \textbf{\textbf{weakly contained}} in $b$ and write $a \prec b$ if for any finite sequence $A_1,\ldots,A_n$ of measurable subsets of $X$, finite $F \subseteq \Gamma$ and $\epsilon > 0$ there exist measurable subsets $B_1,\ldots,B_n$ of $Y$ such that for all $\gamma \in F$ and all $i,j \leq n$ we have \[ | \mu( \gamma^a A_i \cap A_j) - \nu( \gamma^b B_i \cap B_j)| < \epsilon. \] We say $a$ is \textbf{\textbf{weakly equivalent}} to $b$ and write $a \sim b$ if $a \prec b$ and $b \prec a$. \end{definition} We may assume in this definition that $A_1,\ldots,A_n$ form a partition of $X$. Note that we do not require $(X,\mu)$ and $(Y,\nu)$ to be standard, that is to say we include the case where they might be countable. The relation of weak containment is $G_\delta$, so the quotient $\mathrm{A}_\sim(\Gamma,X,\mu)$ of $\mathrm{A}(\Gamma,X,\mu)$ by weak equivalence is well-behaved. \\ \\ We also consider a generalization of weak containment, due to Tucker-Drob. For probability spaces $(Y_i,\nu_i), 1 \leq i \leq m$ and positive real numbers $\alpha_i, 1 \leq i \leq m$ with $\sum_{i=1}^m \alpha_i = 1$ we let $\bigsqcup_{i=1}^m \alpha_i Y_i$ be the probability space formed by endowing the disjoint union of the $Y_i$ with the measure $\sum_{i=1}^m \alpha_i \nu_i$ given by $\left( \sum_{i=1}^m \alpha_i \nu_i \right)(A) = \sum_{i=1}^m \alpha_i \nu_i(A \cap Y_i)$. If $\Gamma \curvearrowright^{a_i} (Y_i,\nu_i)$ are measure-preserving actions, then $\sum_{i=1}^m \alpha_i a_i$ is the action on $\bigsqcup_{i=1}^m \alpha_i Y_i$ given by letting $\Gamma$ act like $a_i$ on $Y_i$. \begin{definition} If $\Gamma \curvearrowright^a (X,\mu)$ and $\Gamma \curvearrowright^b (Y,\nu)$ are measure-preserving actions, we say $a$ is \textbf{\textbf{stably weakly contained}} in $b$ if for all $A_1,\ldots,A_k \in \mathrm{MALG}_\mu$, all finite $F \subseteq \Gamma$ and all $\epsilon > 0$ there exist $\alpha_1, \ldots,\alpha_m$ such that $\sum_{i=1}^m \alpha_1 = 1$ and sets $B_1,\ldots,B_k \subseteq \bigsqcup_{i=1}^m \alpha_i Y_i$ such that \[ \left \vert \mu(\gamma^a A_i \cap A_j) - \sum_{i=1}^m \alpha_i \nu \left( \gamma^{\sum_{i=1}^m \alpha_i b} B_i \cap B_j \right) \right \vert < \epsilon.\] We write $a \prec_s b$ if $a$ is stably weakly contained in $b$ and $a \sim_s b$ for $a \prec_s b$ and $b \prec_s a$. \end{definition} When we wish to distinguish between and action and its equivalence class, we write $[a]$ for the weak equivalence class of $a$ and $[a]_s$ for the stable weak equivalence class. The quotient of $\mathrm{A}(\Gamma,X,\mu)$ by the relation of stable weak containment is denoted $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$. The goal of this paper is to analyze the topological and geometric structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ and $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ .\\ \\ More specifically, unlike $\mathrm{A}(\Gamma,X,\mu)$, the spaces $\mathrm{A}_\sim(\Gamma,X,\mu)$ and $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ carry a well-defined operation of convex combination. This is inherited from the operation of endowing the disjoint union of two probability spaces with a convex combination of their respective measures. In Section 2 we introduce a variation of a construction of Fritz which abstracts the idea of convex combinations. Fritz's objects are referred to as `convex spaces'; we weaken the definition in order to encompass the convex structure on $\mathrm{A}_\sim(\Gamma,X,\mu)$ obtaining the notion of `weak convex space'. We show that this relates naturally to other ideas of convexity, define a notion of convex function and generalize the important geometric notions of `convex hull', `extreme point' and `face' from the classical situation of vector spaces to this abstract framework. We also define `topological weak convex spaces' as weak convex structures which are appropriately compatible with an underlying topology.\\ \\ In Section 3 we consider methods of topologizing $\mathrm{A}_\sim(\Gamma,X,\mu)$. The first topology defined on this space was in , and a second formulation was given in . These are equivalent, Polish, compact and finer than the quotient of the weak topology on $\mathrm{A}(\Gamma,X,\mu)$. We discuss a third topology, implicit in and pointed out to us by Kechris. This is shown to be equivalent to the previous two. We also consider a natural topology on $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$.\\ \\ In Section 4 we describe how to endow $\mathrm{A}_\sim(\Gamma,X,\mu)$ with the structure of a weak convex space and show that it is in fact a topological weak convex space. Furthermore, we show that the metric giving $\mathrm{A}_\sim(\Gamma,X,\mu)$ its Polish topology is compatible with the convex structure in the sense that the distance function to any compact convex set is a convex function. \\ \\ In Section 5 we analyze the structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for amenable $\Gamma$. The main tool is the following idea. Let $\mathrm{Sub}(\Gamma)$ be the space of subgroups of $\Gamma$, regarded as a subspace of $\{0,1\}^\Gamma$ with the product topology. $\mathrm{Sub}(\Gamma)$ is then a compact metric space on which $\Gamma$ acts by conjugation. \begin{definition} An \textbf{\textbf{invariant random subgroup}} of $\Gamma$ is a conjugation-invariant Borel probability measure on $\mathrm{Sub}(\Gamma)$. \end{definition} Invariant random subgroups have been studied in numerous recent papers, including , , and . If $\Gamma \curvearrowright^a (X,\mu)$ is a measure-preserving action, then the pushforward measure $(\mathrm{stab}_a)_* \mu$ is an invariant random subgroup of $\Gamma$ called the type of $a$. We extend ideas of Tucker-Drob from to show the following. \begin{theorem} If $\Gamma$ is amenable, then $\mathrm{A}_\sim(\Gamma,X,\mu)$ is isomorphic to $\mathrm{IRS}(\Gamma)$ as a topological convex space. In particular, if $\Gamma$ is amenable then $\mathrm{A}_\sim(\Gamma,X,\mu)$ is isomorphic to a compact convex subset of a Banach space. \end{theorem} In Section 6 we consider the structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for general $\Gamma$. If $\Gamma$ is nonamenable, the existence of strongly ergodic actions of $\Gamma$ implies that the convex structure on this space has the pathology that the convex combination of a point $x$ with itself might be different from $x$. This is why we need to consider weak convex spaces instead of just convex spaces. The main result of this section is the following Krein-Milman type theorem. \begin{theorem} $\mathrm{A}_\sim(\Gamma,X,\mu)$ is equal to the closed convex hull of its extreme points. In other words, finite convex combinations of the extreme points of $\mathrm{A}_\sim(\Gamma,X,\mu)$ are dense in $\mathrm{A}_\sim(\Gamma,X,\mu)$. \end{theorem} Given this result, it seems interesting to describe the extreme points of $\mathrm{A}_\sim(\Gamma,X,\mu)$. In the amenable case, the identification with $\mathrm{IRS}(\Gamma)$ provides a complete such description, since the extreme points of $\mathrm{IRS}(\Gamma)$ are known to be the ergodic measures and consequently the extreme points of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for amenable $\Gamma$ are exactly those actions with ergodic type. In the nonamenable case this description does not suffice. It is clear that any strongly ergodic action is an extreme point. We are able to show the following. \begin{theorem} Suppose $[a] \in \mathrm{A}_\sim(\Gamma,X,\mu)$ is an extreme point. Let $a = \int_Z a_z d \eta(z)$ be the ergodic decomposition of $a$. Then there is a measure-preserving action $b$ of $\Gamma$ such that for $\eta$-almost all $z \in Z$ we have $[a_z] = [b]$. \end{theorem} Let $\mathrm{FR}_{\sim}(\Gamma,X,\mu)$ denote the subspace of $\mathrm{A}_\sim(\Gamma,X,\mu)$ consisting of the weak equivalence classes of free actions. We prove: \begin{theorem} Let $\mathbb{F}_N$ be a free group of finite or countably infinite rank. Then the weak equivalence classes containing a free ergodic action are dense in $\mathrm{FR}_{\sim}(\mathbb{F}_N,X,\mu)$.\end{theorem} In Section 7 we use a characterization of convex subsets of Banach spaces from to show the following. \begin{theorem} For any $\Gamma$, the space $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ is isomorphic to a compact convex subset of a Banach space. \end{theorem} We characterize the extreme points of $A_{\sim_s}(\Gamma,X,\mu)$ as precisely those stable weak equivalence classes which contain an ergodic action. This result was obtained by Tucker-Drob and Bowen independently of the author. Tucker-Drob and Bowen have also shown that $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ is a simplex, and the set $\mathrm{FR}_{\sim_s}(\Gamma,X,\mu)$ of stable weak equivalence classes of free actions is a subsimplex. Recall that a Poulsen simplex is a simplex such that the extreme points are dense. Thus from Theorem we have: \begin{corollary} Let $\mathbb{F}_N$ be a free group of finite or countably infinite rank. Then $\mathrm{FR}_{\sim_s}(\mathbb{F}_N,X,\mu)$ is a Poulsen simplex. \end{corollary} \subsection*{Acknowledgements} We would like to thank Alexander Kechris for introducing us to this topic and for many helpful discussions. We also thank Robin Tucker-Drob for informing us of his result with Bowen that the space of stable weak equivalence classes forms a simplex, and for raising the question of when it forms a Poulsen simplex. \section{Weak convex spaces.} We first describe the formalism realized by $\mathrm{A}_\sim(\Gamma,X,\mu)$. \subsection{Convex spaces and weak convex spaces.} Convex spaces were introduced in and further developed in as an abstract setting to study the notion of convex combination. \begin{definition} A \textbf{\textbf{convex space}} is a set $X$ together with a family $\mathcal{V}$ of binary operations $cc_{t}$ for each $t \in [0,1]$ such that for all $x,y,z \in X$ and all $s,t \in [0,1]$ \begin{description} \item{(1)} $cc_0(x,y) = x$, \item{(2)} $cc_t(x,x) = x$, \item{(3)} $cc_t(x,y) = cc_{1-t}(y,x)$, \item{(4)} $cc_{t}(cc_s(x,y),z) = cc_{st}\left (x,cc_{\frac{t(1-s)}{1-st}}(y,z)\right)$. \end{description} \end{definition} We will usually write $tx +_{\mathcal{V}} (1-t)y$ for $cc_t(x,y)$, omitting the subscript $\mathcal{V}$ when the convex structure being considered is clear. Note that $\mathrm{(4)}$ allows us to unambiguously define $\sum_{i=1}^n \lambda_i x_i$ for $(x_i)_{i=1}^n \subseteq X$ and $(\lambda_i)_{i=1}^n \subseteq [0,1]$ such that $\sum_{i=1}^n \lambda_i = 1$. We will need to weaken the definition of a convex space to cover the situation where a convex combination of a point $x$ with itself could be different from $x$. \begin{definition} An \textbf{\textbf{weak convex space}} is a set $X$ with a family $cc_{t}$ of binary operations for $t \in [0,1]$ satisfying $\mathrm{(1)},\mathrm{(3)}$ and $\mathrm{(4)}$ of Definition . \end{definition} \begin{definition} A \textbf{\textbf{topological (weak) convex space}} is a topological space $X$ carrying a (weak) convex structure such that the ternary operation $cc: [0,1] \times X^2 \to X$ given by $cc(t,x,y) = cc_t(x,y)$ is continuous. \end{definition} \subsection{Extreme points and faces.} We can define extreme points in a weak convex space in exactly the same way as in a vector space. \begin{definition} If $A$ is a convex set in a weak convex space, we say $x \in A$ is an \textbf{\textbf{extreme point}} if $x = ty + (1-t)z$ for $0 < t<1$ and some $y,z \in A$ implies $y = z = x$. Write $\mathrm{ex}(A)$ for the set of extreme points of $A$. If $A$ is a compact convex subset of a topological weak convex space, we say a \textbf{\textbf{face}} of $A$ is a nonempty closed subset $F \subseteq A$ such that if $x,y \in A$, $0 <t <1$ and $tx + (1-t)y \in F$ then $x,y \in F$. \end{definition} \section{Topology on $\mathrm{A}_\sim(\Gamma,X,\mu)$.} Let $\Gamma$ be a countable group and $\mathrm{A}_\sim(\Gamma,X,\mu)$ be its space of actions modulo weak equivalence. We consider a metric on $\mathrm{A}_\sim(\Gamma,X,\mu)$ which is implicit in .\\ \\ Fix an enumeration $(\gamma_i)_{i=0}^\infty$ of $\Gamma$. If $\mathcal{A} = \{A_1,\ldots,A_k\}$ is a partition of $X$ into $k$ pieces, $a \in \mathrm{A}(\Gamma,X,\mu)$ and $n \in \mathbb{N}$, let $M^{\mathcal{A}}_{n,k}(a) \in [0,1]^{n \times k \times k}$ be the point whose $p,q,r$ coordinate is $\mu(\gamma_p^a A_q \cap A_r)$, where $p \leq n$ and $q,r \leq k$. Let $C_{n,k}(a) = \overline{ \{ M_{n,k}^\mathcal{A}(a): \mathcal{A} \mbox{ is a partition of } X \mbox{ into } k \mbox{ pieces.} \} }$ Then we can define a pseudometric $d$ on $\mathrm{A}(\Gamma,X,\mu)$ by the formula \[ d(a,b) = \sum_{n,k=1}^\infty \frac{1}{2^{n+k}} d_H( C_{n,k}(a), C_{n,k}(b) ) \] where $d_H$ is the Hausdorff distance in the hyperspace of compact subsets of $[0,1]^{n \times k \times k}$. It is easy to see that $a \sim b$ if and only if $d(a,b) = 0$, so $d$ descends to a metric on $\mathrm{A}_\sim(\Gamma,X,\mu)$, which we also denote by $d$. Let $\tau_1$ be the topology induced by $d$. We note that this definition extends to actions on countable spaces. We will write $A_\sim^*(\Gamma)$ for the space of all actions of $\Gamma$ on probability spaces. \\ \\ We now describe a different construction of the topology on $\mathrm{A}_\sim(\Gamma,X,\mu)$ due to Tucker-Drob in order to show it agrees with the one we have just introduced. (Tucker-Drob shows in that his formulation agrees with the one from ). \\ \\ Let $S$ be a compact Polish space, and consider $S^\Gamma$, which is also a compact Polish space. $\Gamma$ acts on $S^\Gamma$ by the shift action $s$ given by $(\gamma^s f)(\delta) = f(\gamma^{-1} \delta)$. Let $M_s(S^\Gamma)$ be the compact Polish space of shift-invariant probability measures on $S^\Gamma$ and let $\mathcal{K}_S = \mathcal{K}(M_s(S^\Gamma))$ be the hyperspace of compact subsets of $M_s(S^\Gamma)$ equipped with the Hausdorff topology. Then $\mathcal{K}_S$ is again compact and Polish. Now consider a $S$-valued random variable $\phi \in L(X,\mu,S)$ on $X$, that is to say a measurable map $\phi: X \to S$. For each a measure-preserving action $a \in \mathrm{A}(\Gamma,X,\mu)$ we get a map $\Phi_S^{\phi,a}: X \to S^\Gamma$ by letting $\Phi_S^{\phi,a}(x)(\gamma) = \phi( (\gamma^{-1})^a x)$ and consequently a shift-invariant measure $(\Phi_S^{\phi,a})_* \mu$ on $S^\Gamma$. Then define a subset $E(a,S)$ of $M_s(S^\Gamma)$ by \[E(a,S) = \{ (\Phi_S^{\phi,a})_* \mu: \phi: X \to S \mbox{ is measurable} \}.\] Let $\Phi_S: \mathrm{A}(\Gamma,X,\mu) \to \mathcal{K}_S$ be given by $\Phi_S(a) = \overline{E(a,S)}$. When $S = K$ is the Cantor set, we omit the subscript $S$ on the notations just introduced. By Proposition 3.5 in , we have $a \sim b$ if and only if $\Phi(a) = \Phi(b)$ so we can consider the initial topology on $\mathrm{A}_\sim(\Gamma,X,\mu)$ induced by $\Phi$. Call this $\tau_2$. We now work towards showing $\tau_1$ agrees with $\tau_2$. There will be a series of preliminary steps. This entire argument can be regarded as a `perturbed' version of Proposition 3.5 in .\\ \\ We first fix a compatible metric on $M_s(K^\Gamma)$. Let $\mathcal{A}_K$ be the collection of clopen subsets of $K^\Gamma$ of the form $\pi_F^{-1} \left ( \prod_{\gamma \in F} A_\gamma \right)$ where $A_\gamma \subseteq K$ an element of some fixed countable clopen basis for $K$, $F \subseteq \Gamma$ is finite and $\pi: K^\Gamma \to K^F$ is the projection onto the $F$-coordinates. Since the elements of $\mathcal{A}_K$ generate the Borel $\sigma$-algebra of $K^\Gamma$, for $(\nu_n)_{n=1}^\infty \subseteq M_s(K^\Gamma)$ we have $\nu_n \to \nu$ in $M_s(K^\Gamma)$ if and only if $\nu_n(A) \to \nu(A)$ for every $A \in \mathcal{A}_K$. So, enumerating the elements of $\mathcal{A}_K$ as $(A^K_i)_{i=1}^\infty$, $\delta_K$ given by \[ \delta_K(\nu,\rho) = \sum_{i=1}^\infty \frac{1}{2^i} | \nu(A^K_i) - \rho(A^K_i)| \] is a compatible metric on $M_s(K^\Gamma)$. \begin{lemma} For any $\epsilon > 0$ there exists $k \in \mathbb{N}$ such that every $a$ and every $\phi \in L(X,\mu,K)$ there is $\psi \in L(X,\mu,K)$ with $\delta_K((\Phi^{\phi,a})_* \mu, (\Phi^{\psi,a})_* \mu) < \epsilon$ such that the range of $\psi$ has size $\leq k$. Note that $k$ depends only on $\epsilon$, not on $a$ or $\phi$. \end{lemma} \begin{proof} Fix $\epsilon$. Choose $N$ large enough that $\sum_{i=N}^\infty \frac{1}{2^i} < \epsilon$. For each $i \leq N$, write $A_i = \pi_{F_i}^{-1} \left( \prod_{\gamma \in F_i} A^i_\gamma \right)$ for $A^i_\gamma \subseteq K$ clopen and $F_i \subseteq \Gamma$ finite. We have for all $a \in \mathrm{A}(\Gamma,X,\mu)$ and $\phi,\psi \in L(X,\mu,K)$, \begin{align} | \Phi^{\phi,a}(A_i) - \Phi^{\psi,a}(A_i)| & = \left \vert \Phi^{\phi,a}\left ( \pi_{F_i}^{-1} \left( \prod_{\gamma \in F_i} A^i_\gamma \right) \right)- \Phi^{\psi,a}\left( \pi_{F_i}^{-1} \left( \prod_{\gamma \in F_i} A^i_\gamma \right) \right) \right \vert \nonumber \\ & = |\mu( \{x: \Phi^{\phi,a}(x)(\gamma) \in A^i_\gamma \mbox{ for all } \gamma \in F_i \} ) - \mu( \{x: \Phi^{\psi,a}(x)(\gamma) \in A^i_\gamma \mbox{ for all } \gamma \in F_i \} ) | \nonumber \\ & = | \mu(\{x: \phi((\gamma^{-1})^a x) \in A^i_\gamma \mbox{ for all } \gamma \in F_i\}) -\mu(\{x: \psi((\gamma^{-1})^a x) \in A^i_\gamma \mbox{ for all } \gamma \in F_i \})| \nonumber \\ & =\left \vert \mu \left ( \bigcap_{\gamma \in F_i} \gamma^a \phi^{-1}(A^i_\gamma) \right) - \left( \bigcap_{\gamma \in F_i} \gamma^a \psi^{-1}(A^i_\gamma) \right) \right \vert. \end{align} Now, fix $\phi \in L(X,\mu,K)$. Let $(B_j)_{j=1}^k$ be the finite partition of $K$ given by the atoms of the Boolean algebra generated by $(A^i_\gamma)_{i \leq N, \gamma \in F_i}$. Note that $k$ depends only on $\epsilon$. For each $j \leq k$, let $y_j$ be any point in $B_j$. Define a map $\psi: X \to K$ by letting $\psi(x) = y_j$ for the unique $j$ such that $x \in \phi^{-1}(B_j)$. Then $\psi^{-1}(B_j) = \phi^{-1}(B_j)$ for each $j$, and hence $\phi^{-1}(A^i_\gamma) = \psi^{-1}(A^i_\gamma)$ for each $i \leq N$ and $\gamma \in F_i$. Therefore the value of the expression $(1) $ is $0$ and $\delta_K( (\Phi^{\phi,a})_* \mu, (\Phi^{\psi,a})_* \mu ) < \epsilon.$ \end{proof} \begin{lemma} If $\overline{ E(a_n,L)} \to \overline{ E(a,L)}$ in $\mathcal{K}(M_s(L^\Gamma))$ for every finite set $L$ then $\overline{ E(a_n,K)} \to \overline{E(a,K)}$ in $\mathcal{K}(M_s(K^\Gamma))$. \end{lemma} \begin{proof} Fix $\epsilon > 0$ in order to show that eventually $d_{\mathcal{K}}\left(\overline{E(a_n,K)},\overline{E(a,K)} \right) < \epsilon$, where $d_{\mathcal{K}}$ is the Hausdorff distance in $\mathcal{K}(M_s(K^\Gamma))$ constructed from $\delta_K$. For $k \in \mathbb{N}$ and $b \in \mathrm{A}(\Gamma,X,\mu)$ let \[E_k(b,K) = \{ (\Phi^{\phi,a})_* \mu: \phi: X \to K \mbox{ is measurable and the range of } \phi \mbox{ has size } \leq k \}. \] By Lemma we can choose $k \in \mathbb{N}$ such that $E(b,K) \subseteq B_{\frac{\epsilon}{4}}(E_k(b,K))$ for every $b \in \mathrm{A}(\Gamma,X,\mu)$ where $B_r(A) = \{ \nu \in M_s(K^\Gamma): \delta_K (\nu,\rho) < r$ for some $\rho \in A\}$. Notice that $E_k(b,K) = \bigcup_{\substack{L \subseteq K,\\ |L| = k}} E(b,L)$. Fix a set $L$ of size $k$ and choose $N$ large enough such that if $n\geq N$ then $d_{\mathcal{K}_L}\left(\overline{E(a_n,L)},\overline{E(a,L)} \right) < \frac{\epsilon}{4}$ where $d_{\mathcal{K}_L}$ is the Hausdorff distance in $\mathcal{K}(M_s(L^\Gamma))$. Since the construction is independent of the set chosen to realize $L$, we have in fact $d_{\mathcal{K}_L}\left(\overline{E(a_n,L)},\overline{E(a,L)} \right) < \frac{\epsilon}{4}$ for every finite set $L$ of size $k$. For a fixed finite $L \subseteq K$ let $E_L(b,K) = \{ (\Phi^{b,\phi})_* \mu: \phi: X \to K \mbox{ measurable, } \phi(X) \subseteq L \} $. Then for any $b,c \in \mathrm{A}(\Gamma,X,\mu)$ we have \[d_\mathcal{K}\left(\overline{E_L(b,K)}, \overline{E_L(c,K)} \right) = d_{\mathcal{K}_L} \left( \overline{E(b,L)}, \overline{E(c,L)} \right), \] so that when $n\geq N$, \begin{align*} d_{\mathcal{K}}\left(\overline{E_k(a_n,K)}, \overline{E_k(a,K)} \right) &= d_{\mathcal{K}} \left( \bigcup_{\substack{L \subseteq K\\ |L| = k}} \overline{E(a_n,L)},\bigcup_{\substack{L \subseteq K \\ |L| = k}} \overline{E(a,L)} \right) \\ &\leq \sup_{\substack{L \subseteq K \\ |L| = k}} d_{\mathcal{K}_L} \left( \overline{E(a_n,L)}, \overline{E(a,L)} \right) < \frac{\epsilon}{4}. \end{align*} Therefore when $n \geq N$, \begin{align*} d_{\mathcal{K}}\left(\overline{E(a_n,K)},\overline{E(a,K)}\right) &\leq d_{\mathcal{K}}\left(\overline{E(a_n,K)}, \overline{E_k(a_n,K)}\right) + d_{\mathcal{K}}\left( \overline{E_k(a_n,K)}, \overline{E_k(a,K)}\right) \\ & \hspace{1 in}+ d_{\mathcal{K}}\left(\overline{E_k(a,K)}, \overline{E(a,K)}\right) \\ &< \frac{3 \epsilon}{4}. \end{align*} \end{proof} \begin{lemma} Let $L$ be a finite set of size $k$. Then for each finite set $(A_p)_{p=1}^q$ of basic clopen sets $A_p \subseteq L^\Gamma$ and $\epsilon > 0$ there is $\delta > 0$ such that if $d(a,b) < \delta$ then for all $\phi \in L(X,\mu,L)$ there exists $\psi \in L(X,\mu,L)$ such that $|(\Phi_L^{a,\phi})_* \mu(A_p) - (\Phi_L^{b,\psi})_* \mu(A_p) | < \epsilon$ for all $p \leq q$. \end{lemma} \begin{proof} Write $A_p = \bigcap_{\gamma \in F_p} \pi_\gamma^{-1}(\jmath_p(\gamma))$ for some $F_p \subseteq \Gamma$ finite, $\jmath: F_p \to k$ and fix $\epsilon > 0$. Choose a finite $F \subseteq \Gamma$ with $(F_p)^2 \subseteq F$ for all $p \leq q$. We may assume the identity $e \in F$. Suppose $d(a,b) < \frac{\delta}{2^{|F| + k^{|F|}}}$; we will specify a value for $\delta$ later. Now fix $\phi:X \to k$ and let $B_i = \phi^{-1}(i)$. Given $\eta: F \to k$, let $B_\eta = \bigcap_{\gamma \in F} \gamma^{a} B_{\eta(\gamma)}$. We can then find a partition $\{D_\eta\}_{\eta \in k^F}$ such that \[ | \mu( \gamma^{a}B_{\eta_1} \cap B_{\eta_2}) - \mu( \gamma^b D_{\eta_1} \cap D_{\eta_2})| < \delta \] for all $\eta_1,\eta_2 \in k^F$ and $\gamma \in F$. Define $\psi: X \to k$, by $\psi(y) = l$ if $y \in D_{\eta}$ for some $\eta$ with $\eta(e) = l$. Furthermore, for each $l \leq k$ let $D_l = \bigsqcup \{D_\eta: \eta \in k^F$ and $\eta(e) = l \} = \psi^{-1}(l)$. For each $J \subseteq F$ and $\sigma \in k^J$ let $D_\sigma = \bigsqcup \{ D_\eta: \eta \in k^F$ and $\sigma \sqsubseteq \eta \}$, where $\sigma \sqsubseteq \eta$ means $\eta$ extends $\sigma$ and let $\tilde{D}_\sigma = \bigcap_{\gamma \in J} \gamma^b D_{\sigma(\gamma)}$. Furthermore if $\gamma \in \Gamma$, $J \subseteq \Gamma$ and $\sigma \in k^J$ let $\gamma \cdot \sigma \in k^{\gamma J}$ be given by $(\gamma \cdot \sigma)(\delta) = \sigma( \gamma^{-1} \delta)$. For $\sigma \in K^{F_p}$ and $\gamma \in F_p$ we have \begin{align*} | \mu(\gamma^b D_{\sigma} \cap D_{\gamma \cdot \sigma}) - \mu(\gamma^a B_{\sigma} \cap B_{\gamma \cdot \sigma})| & \leq \sum_{(\eta \in k^{F}: \sigma \sqsubseteq \eta)} \sum_{(\eta' \in k^{F}: \gamma \cdot \sigma \sqsubseteq \eta')} |\mu( \gamma^b D_\eta \cap D_{\eta'}) - \mu( \gamma^a B_{\eta} \cap B_{\eta'})| \\ & \leq \delta (k^{|F|})^2 \end{align*} In particular, setting $\gamma = e$ we see $|\mu(B_\sigma) - \mu(D_\sigma)| < \delta k^{2 |F|}$ for every $\sigma: F_p \to k$. Since $\gamma^a B_\sigma = B_{\gamma \cdot \sigma} = \gamma^a B_\sigma \cap B_{\gamma \cdot \sigma}$ we have \begin{align*} |\mu(D_\sigma) - \mu( \gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| & \leq | \mu(D_\sigma) - \mu(\gamma^a B_\sigma)| + |\mu(\gamma^a B_\sigma \cap B_{\gamma \cdot \sigma}) - \mu(\gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| \\ & = | \mu(D_\sigma) - \mu(B_\sigma)| + |\mu(\gamma^a B_\sigma \cap B_{\gamma \cdot \sigma}) - \mu(\gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| \\ &< 2 \delta k^{2 |F|} \end{align*} and also \begin{align*} |\mu(D_{\gamma \cdot \sigma}) - \mu( \gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| & \leq | \mu(D_{\gamma \cdot \sigma}) - \mu(B_{\gamma \cdot \sigma})| + |\mu(\gamma^a B_\sigma \cap B_{\gamma \cdot \sigma}) - \mu(\gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| \\ &< 2 \delta k^{2 |F|} \end{align*} Therefore \begin{align} \mu( (\gamma^b D_\sigma) \triangle (D_{\gamma \cdot \sigma})) & = \mu( \gamma^b D_\sigma) + \mu(D_{\gamma \cdot \sigma}) - 2 \mu( \gamma^b D_\sigma \cap D_{\gamma \cdot \sigma}) \nonumber \\ & \leq |\mu(D_{\gamma \cdot \sigma}) - \mu( \gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| + |\mu(D_{\gamma \cdot \sigma}) - \mu( \gamma^b D_\sigma \cap D_{\gamma \cdot \sigma})| \nonumber \\ & < 4 \delta k^{2 |F|} \end{align} Since $(D_\eta)_{\eta \in k^F}$ is a partition of $X$ and $(F_p)^2 \subseteq F$ we have \[ D_{\jmath_p} = \bigsqcup_{\substack{ \eta \in k^F \\ \jmath_p \sqsubseteq \eta }} D_\eta = \bigcap_{\gamma \in F_p} \bigsqcup_{\substack{\sigma \in k^{\gamma F_p}\\ \sigma(\gamma) = \jmath_p(\gamma)}} D_\sigma = \bigcap_{\gamma \in F_p} \bigsqcup_{\substack{\sigma \in k^{F_p} \\ \sigma(e) = \jmath_p(\gamma)}} D_{\gamma \cdot \sigma}.\] Now, by $(2)$, \begin{equation} \mu \left( \left( \bigcap_{\gamma \in F_p} \bigsqcup_{\substack{\sigma \in k^{F_p} \\ \sigma(e) = \jmath(\gamma)}} D_{\gamma \cdot \sigma} \right) \triangle \left( \bigcap_{\gamma \in F_p} \bigsqcup_{\substack{\sigma \in k^{F_p} \\ \sigma(e) = \jmath_p(\gamma)}} \gamma^b D_\sigma \right) \right) < (|F_p| k^{|F_p|}) ( 4 \delta k^{2 |F|}). \end{equation} Note that $ \bigcap_{\gamma \in F_p} \bigsqcup_{\substack{\sigma \in k^{F_p} \\ \sigma(e) = \jmath_p(\gamma)}} \gamma^b D_\sigma = \bigcap_{\gamma \in F_p} \gamma^b D_{\jmath_p(\gamma)} = \tilde{D}_{\jmath_p}$, so $(3)$ reads $|\mu(D_{\jmath_p}) - \mu(\tilde{D}_{\jmath_p})| < (|F_p| k^{|F_p|}) ( 4 \delta k^{2 |F|})$. Moreover, \begin{align*} (\Phi^{b,\psi}_L)_* \mu(A_p) & = \mu(\{x: \Phi^{b,\psi}_L(x) \in A_p \}) \\ & = \mu(\{x: \Phi^{b,\psi}_L(x)(\gamma) = \jmath_p(\gamma) \mbox{ for all } \gamma \in F_p \}) \\ & = \mu(\{x: \psi((\gamma^{-1})^b x) = \jmath_p(\gamma) \mbox{ for all } \gamma \in F_p \}) \\ & = \mu(\{x: x \in \gamma^b \psi^{-1}( \jmath_p(\gamma)) \mbox{ for all } \gamma \in F_p \}) \\ & = \mu \left( \bigcap_{\gamma \in F_p} \gamma^b D_{\jmath_p(\gamma)} \right) \\ & = \mu(\tilde{D}_{\jmath_p}). \end{align*} Similarly, $(\Phi^{a,\phi}_L)_*\mu(A_p) = \mu(B_{\jmath_p})$. So we finally have \begin{align*} | (\Phi^{b,\psi}_L)_* \mu(A_p) - (\Phi^{a,\phi}_L)_*\mu(A_p) | & = |\mu(\tilde{D}_{\jmath_p}) - \mu(B_{\jmath_p})| \\ & \leq |\mu(\tilde{D}_{\jmath_p}) - \mu(D_{\jmath_p})| + |\mu(D_{\jmath_p}) - \mu(B_{\jmath_p})| \\ &< (|F_p| k^{|F_p|}) ( 4 \delta k^{2 |F|}) + 2 \delta k^{2 |F|}.\end{align*} Since $k$ is fixed in advance, $|F_p| \leq |F|$ and $F$ depends only on $(A_p)_{p=1}^q$, it is clear that $\delta$ can be chosen so $(|F_p| k^{|F_p|}) ( 4 \delta k^{2 |F|}) + 2 \delta k^{2 |F|} < \epsilon$ for all $p \leq q$. \end{proof} We can now prove the main result of this section. \begin{theorem} $\tau_1 = \tau_2$. \end{theorem} \begin{proof} Suppose that $a_n \to a$ in $\tau_1$. We need to prove $\Phi(a_n) \to \Phi(a)$ in $\mathcal{K}(M_s(K^\Gamma))$. By Lemma it suffices to fix a finite set $L$ and show $\overline{E(a_n,L)} \to \overline{E(a,L)}$ in $\mathcal{K}(M_s(L^\Gamma))$. Let $k = |L|$. Write $E_n = E(a_n,L)$ and $E = E(a,L)$. As before, if we let $\mathcal{A}_L = (A^L_i)_{i=1}^\infty$ be the collection of clopen subsets of $L^\Gamma$ of the form $\bigcap_{\gamma \in F} \pi_\gamma^{-1}(j_\gamma)$ for a finite $F \subseteq \Gamma$ and $j_\gamma \leq k$, then \[ \delta_L(\nu,\rho) = \sum_{i=1}^\infty \frac{1}{2^i} |\nu(A^L_i) - \rho(A^L_i)| \] is a compatible metric on $M_s(L^\Gamma)$. Fix $\epsilon > 0$ in order to show that eventually $d_L(\overline{E_n},\overline{E}) < \epsilon$, where $d_L$ is the Hausdorff distance in $\mathcal{K}(M_s(L^\Gamma))$ constructed from $\delta_L$. Choose $N$ sufficiently large that $\sum_{i=N}^\infty \frac{1}{2^i} < \frac{\epsilon}{2}$. By Lemma there is $\delta > 0$ such that if $d(a,b) < \delta$ then for each $i \leq N$ and all $\phi \in L(X,\mu,L)$ there exists $\psi \in L(X,\mu,L)$ such that $|(\Phi_L^{a,\phi})_* \mu(A^L_i) - (\Phi_L^{b,\psi})_* \mu(A^L_i) | < \frac{\epsilon}{2}$. Thus if $M$ is large enough that $d(a_n,a) < \delta$ for $n \geq M$, we have $d_L( \overline{E_n}, \overline{E}) < \epsilon$.\\ \\ Now suppose $\Phi(a_n) \to \Phi(a)$ in $\mathcal{K}(M_s(K^\Gamma))$. Fix $r,q$ and $\epsilon > 0$ in order to show that eventually $d_H(C_{r,q}(a_n),C_{r,q}(a)) < \epsilon$. Choose $q$ distinct points $(x_p)_{p=1}^q \in K$ and let $(D_p)_{p=1}^q$ be a family of disjoint clopen subsets of $K$ with $x_p \in D_p$. Now let $M$ be large enough that all sets of the form $\pi_{\gamma_s}^{-1}(D_p) \cap \pi_e^{-1}(D_t)$ for $s \leq r$ and $p,t \leq q$ appear as some $A^K_i$ for $i \leq M$ in our previously chosen clopen basis $\mathcal{A}_K$. Then choose $N$ large enough that when $n \geq N$, $d_{\mathcal{K}}( \Phi(a_n), \Phi(a) ) < \frac{\epsilon}{2^{M}}$. Then for each $\phi \in L(X,\mu,K)$ we have $\psi \in L(X,\mu,K)$ such that $\delta_K((\Phi^{a_n,\phi})_* \mu, (\Phi^{a,\psi})_* \mu) < \frac{\epsilon}{2^{M}}$. So in particular, if $n \geq N$ then for each $\phi \in L(X,\mu,K)$ there exists $\psi \in L(X,\mu,K)$ such that \[ | (\Phi^{a_n,\phi})_* \mu(\pi_{\gamma_s}^{-1}(D_p) \cap \pi_e^{-1}(D_t)) - (\Phi^{a,\psi})_* \mu(\pi_{\gamma_s}^{-1}(D_p) \cap \pi_e^{-1}(D_t))| < \epsilon \] for all $p,t \leq q$ and $s \leq r$.\\ \\ Now suppose $n \geq N$ and let $(B_p)_{p=1}^q$ be a partition of $X$. Define $\phi: X \to K$ by taking $\phi(x) = x_p$ for the unique $p \leq q$ with $x \in B_p$ so by the previous paragraph we have a corresponding $\psi$. Observe that for all $\gamma \in \Gamma$ we have \begin{align*} \mu( \gamma^{a_n} B_p \cap B_t) &= \mu( \gamma^{a_n} \phi^{-1}(D_p) \cap \phi^{-1}(D_t) ) \\ & = \mu(\{x: \phi((\gamma^{a_n})^{-1} x) \in D_p \mbox{ and } \phi(x) \in D_t \}) \\ & = \mu( \{ x: \Phi^{\phi,a_n}(x)(\gamma) \in D_p \mbox{ and } \Phi^{\phi,a_n}(x)(e) \in D_t \}) \\ & = \mu( \{ x: \Phi^{\phi,a_n}(x) \in \pi_{\gamma}^{-1}(D_p) \mbox{ and } \Phi^{\phi,a_n}(x) \in \pi_e^{-1}(D_t) \}) \\ & = \mu( \{x : \Phi^{\phi,a_n}(x) \in \pi_{\gamma}^{-1}(D_p) \cap \pi_1^{-1}(D_t) \} ) \\ & = (\Phi^{\phi,a_n})_* \mu( \pi_{\gamma}^{-1}(D_p) \cap \pi_1^{-1}(D_t) ). \end{align*} Similarly letting $H_p = \psi^{-1}(D_p)$ we have $\mu( \gamma^a H_p \cap H_t) = (\Phi^{\psi,a_n})_* \mu( \pi_{\gamma}^{-1}(D_p) \cap \pi_1^{-1}(D_t) )$. Thus for all $p,t \leq q$ and $s \leq r$, \[ |\mu( \gamma_s^{a_n} B_p \cap B_t ) - \mu(\gamma_s^a H_p \cap H_t)| = |(\Phi^{\phi,a_n})_* \mu( \pi_{\gamma+s}^{-1}(D_p) \cap \pi_e^{-1}(D_t) ) - (\Phi^{\psi,a_n})_* \mu( \pi_{\gamma_s}^{-1}(D_p) \cap \pi_e^{-1}(D_t) )| < \epsilon.\] We have shown that when $n \geq N$, $C_{r.q}(a_n) \subseteq B_{\epsilon}(C_{r,q}(a))$. The argument that eventually $C_{r,q}(a) \subseteq B_\epsilon( C_{r,q}(a_n))$ is identical. \end{proof} \subsection{Topology on the space of stable weak equivalence classes.} Let $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ be the space of stable weak equivalence classes and let $\iota$ be the trivial action of $\Gamma$ on an standard probability space. By Lemma 3.7 in , we have $a \prec_s b$ if and only if $a \prec \iota \times b$. Moreover, Theorem 1.1 in says that $\overline{ E (a \times \iota,K)} = \cch(E(a,K))$, where $M_s(K^\Gamma)$ carries its natural topological convex structure as a compact convex subset of a Banach space. Letting $\Psi:\mathrm{A}(\Gamma,X,\mu) \to \mathcal{K}(M_s(K^\Gamma))$ be the map $a \mapsto \cch(E(a,K)) $ we have $\Psi(a) = \Psi(b)$ if and only if $a \sim_s b$. Tucker-Drob gives $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ the initial topology induced by $\Psi$, in which it is a compact Polish space. Thus we have $a_n \to a$ in the topology of $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ if and only if $a_n \times \iota \to a \times \iota$ in the topology of $\mathrm{A}_\sim(\Gamma,X,\mu)$. Therefore we can introduce a metric $d_s$ on $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ by setting $d_s(a,b) = d(a \times \iota,b \times \iota)$. \section{$\mathrm{A}_\sim(\Gamma,X,\mu)$ as a weak convex space.} We now describe how to give $\mathrm{A}_\sim(\Gamma,X,\mu)$ the structure of a weak convex space. Given $t \in [0,1]$ and $a,b \in \mathrm{A}_\sim(\Gamma,X,\mu)$ we let $c \in A\left(\Gamma,X_1 \sqcup X_2,t \mu_1 + (1-t) \mu_2 \right)$ be the disjoint sum of representative actions $a$ and $b$ on the disjoint union of two copies $X_1$ and $X_2$ of $X$ with the first copy carrying a copy of the measure $\mu$ weighted by $t$ and the second copy carrying a copy of $\mu$ weighted by $(1-t)$. To get an action in $\mathrm{A}(\Gamma,X,\mu)$ we need to choose an isomorphism of $(X,\mu)$ with $\left(X_1 \sqcup X_2, t \mu_1 + (1-t) \mu_2 \right)$, but the weak equivalence class of $c$ does not depend on this or on the representatives we chose. So we have a well-defined binary operation $\mathrm{A}_\sim(\Gamma,X,\mu)^2 \to \mathrm{A}_\sim(\Gamma,X,\mu)$. Call this $cc_t$. It is clear that $\mathrm{(1)}$, $\mathrm{(3)}$ and $\mathrm{(4)}$ of Definition are satisfied, so $\mathrm{A}_\sim(\Gamma,X,\mu)$ is a weak convex space. Moreover, we have the following. \begin{proposition} $\mathrm{A}_\sim(\Gamma,X,\mu)$ is a topological weak convex space. \end{proposition} \begin{proof} We must show that $cc$ is continuous. Suppose that $t_j \to t$ in $[0,1]$ and $a_j \to a$ and $b_j \to b$ in the topology of $A_\sim(\Gamma,Y,\mu)$. Write $c_j =t_j a_j + (1-t_j) b_j $ and $c = t a + (1-t)b$. Fixing $l,m \in \mathbb{N}$ write $C(d)$ for $C_{l,m}(d)$. We need to prove that for every $\epsilon >0$ there is $J$ so that if $j > J$ then we have $d_H(C(c_j),C(c)) < \epsilon$, where $d_H$ is the Hausdorff distance in $[0,1]^{l \times m^2}$.\\ \\ First we must show that for sufficiently large $j$, for every partition $B_1,\ldots,B_l$ of $Y$ there is a partition $D_1,\ldots,D_l$ of $Y$ depending on $j$ such that for all $s,t \leq l$ and $p \leq m$, \[ |\mu(\gamma_p^{c_j} D_s \cap D_t) - \mu(\gamma_p^{c} B_s \cap B_t)| < \epsilon. \] Choose $J_1$ so that if $j > J_1$ then $ |t_j - t| < \frac{\epsilon}{6}$. Choose $J_2 > J_1$ so if $j > J_2$ then $d_H(C_{a_j},C_a) < \frac{\epsilon}{6}$ and $d_H(C_{b_j},C_b) < \frac{\epsilon}{6}$. Fix $j > J_2$. Writing $\theta$ for the isomorphism from $(Y_1 \sqcup Y_2, t \mu + (1-t) \mu)$ to $(Y,\mu)$ and $\theta_j$ for the isomorphism from $(Y_1 \sqcup Y_2, t_j \mu + (1-t_j) \mu)$ to $(Y,\mu)$ we have a partition $(B_{s,i})_{s=1}^l$ of $Y_i$ given by $B_{s,i} = \theta^{-1}(B_s) \cap Y_i$. So we can find a partition $(D_{s,i})_{s=1}^l$ of $Y_i$ such that for all $p \leq m$ and all $s,t \leq l$ we have \[ |\mu(\gamma_p^{a_j} D_{s,1} \cap D_{t,1}) - \mu(\gamma_p^{a} B_{s,1} \cap B_{t,1}) | < \frac{\epsilon}{6}\] and \[ |\mu(\gamma_p^{b_j} D_{s,2} \cap D_{t,2}) - \mu(\gamma_p^{b} B_{s,2} \cap B_{t,2}) | < \frac{\epsilon}{6}\] Now, let $D_s = \theta_j(D_{s,1} \sqcup D_{s,2})$. Note that since each $\theta_j(Y_i)$ is $c_j$ invariant, \begin{align*} \mu(\gamma_p^{c_j} D_s \cap D_t) &= \mu( \gamma_p^{c_j} \theta_j(D_{s,1}) \cap \theta_j(D_{t,1})) + \mu( \gamma_p^{c_j} \theta_j(D_{s,2}) \cap \theta_j(D_{t,2})) \\ & = \mu( \theta_j( \gamma_p^{a_j} D_{s,1} \cap D_{t,1} ) ) + \mu( \theta_j( \gamma_p^{b_j} D_{s,2} \cap D_{t,2} ) ) \\ & = t_j \mu(\gamma_p^{a_j} D_{s,1} \cap D_{t,1} ) + (1-t_j) \mu(\gamma_p^{b_j} D_{s,2} \cap D_{t,2} ). \end{align*} Similarly since $\theta(Y_i)$ is $c$-invariant we have \begin{align*} \mu(\gamma_p^{c} B_s \cap B_t)&= \mu( \gamma_p^{c} \theta(B_{s,1}) \cap \theta(B_{t,1})) + \mu( \gamma_p^{c} \theta(B_{s,2}) \cap \theta(B_{t,2})) \\ & = \mu( \theta( \gamma_p^{a} B_{s,1} \cap B_{t,1} ) ) + \mu( \theta( \gamma_p^{b} B_{s,2} \cap B_{t,2} ) ) \\ & = t \mu(\gamma^{a}_p \cap B_{s,1} \cap B_{t,1}) + (1-t) \mu(\gamma^{b}_p \cap B_{s,2} \cap B_{t,2}) \end{align*} Note that if $|x_1 - x_2| < \delta$ and $|y_1 - y_2| < \delta$ then $|x_1y_1 - x_2 y_2| < 3 \delta$. So our assumptions guarantee that we have \[|t_j \mu(\gamma_p^{a_j} D_{s,1} \cap D_{t,1} ) - t \mu(\gamma^{a}_p \cap B_{s,1} \cap B_{t,1})| < \frac{\epsilon}{2} \] and \[|(1-t_j )\mu(\gamma_p^{b_j} D_{s,2} \cap D_{t,2} ) - (1-t) \mu(\gamma^{b}_p \cap B_{s,2} \cap B_{t,2})| < \frac{\epsilon}{2} \] hence \[ |\mu(\gamma_p^{c_j} D_s \cap D_t) - \mu(\gamma_p^{c} B_s \cap B_t)| < \epsilon \] as claimed.\\ \\ Now we must show that for sufficiently large $j$, every partition $B_1,\ldots,B_l$ of $Y$ there is a partition $D_1,\ldots,D_l$ of $Y$ depending on $j$ such that for all $s,t \leq l$ and $p \leq m$ we have \[|\mu(\gamma_p^{c} D_s \cap D_t) - \mu(\gamma_p^{c_j} B_s \cap B_t)| < \epsilon.\] The argument is similar to the previous step, so we omit it. \end{proof} \begin{corollary} $A_\sim(\Gamma,Y,\mu)$ is path connected. \end{corollary} \begin{corollary} $A_\sim(\Gamma,Y,\mu)$ is uncountable. \end{corollary} We now record a lemma which will be useful later, guaranteeing that the metric on $\mathrm{A}_\sim(\Gamma,X,\mu)$ behaves nicely with respect to the convex structure. \begin{lemma} For any convex set $K \subseteq \mathrm{A}_\sim(\Gamma,X,\mu)$ the function $d( \cdot,K) = \inf_{b \in K} d( \cdot,b)$ is convex. \end{lemma} \begin{proof} Let $x,y \in \mathrm{A}_\sim(\Gamma,X,\mu)$ and consider $tx + (1-t)y$. Fix $n,k$ and write $C(a)$ for $C_{n,k}(a)$. It suffices to show that \[ \inf_{b \in K} d_H( C(tx + (1-t)y), C(b)) \leq t( \inf_{b \in K} d_H( C(x), C(b) ) + (1-t)( \inf_{b \in K} d_H( C(y),C(b)))\] where $d_H$ is the Hausdorff distance in the space $[0,1]^{n \times k^2}$. Fix $\epsilon > 0$. It suffices to find $a \in K$ with \begin{equation} d_H( C(tx + (1-t)y), C(a)) \leq t( \inf_{b \in K} d_H( C(x), C(b) ) + \epsilon ) + (1-t)( \inf_{b \in K} d_H( C(y),C(b)) + \epsilon). \end{equation} Choose $c \in K$ with $d_H(C(x),C(c)) < \inf_{b \in K} d_H( C(x), C(b) ) + \epsilon$ and choose $d \in K$ with $d_H(C(x),C(d)) < \inf_{b \in K} d_H( C(y), C(b) ) + \epsilon$. Note that since $K$ is convex, $tc + (1-t)d \in K$. We claim \[ d_H( C(tx + (1-t) y), C(tc + (1-t) d)) \leq t d_H(C(x),C(c)) + (1-t) d_H(C(y),C(d)) \], which implies $(4)$. Let $\delta > 0$, it then suffices to show \begin{equation} d_H( C(tx + (1-t) y), C(tc + (1-t) d)) \leq t( d_H(C(x),C(c)) + \delta) + (1-t) (d_H(C(y),C(d)) + \delta). \end{equation} Let $X_1$ and $X_2$ be two copies of $X$ and $\nu$ be the measure on $X_1 \sqcup X_2$ given by $t (\mu \upharpoonright X_1) + (1-t) (\mu \upharpoonright X_2)$. Let $\mathcal{P} = (P_i)_{i=1}^k$ be a partition of $X_1 \sqcup X_2$. This induces a partition $\mathcal{P}_1 = (P^1_i)_{i=1}^k$ of $X_1$ given by $P^1_i = P_i \cap X_1$ and similarly we have a partition $\mathcal{P}_2 = (P^2_i)_{i=1}^k$ of $X_2$. We can find a partition $\mathcal{Q}_1 = (Q^1_i)_{i=1}^k$ of $X_1$ such that for $m \leq n$ and $i,j \leq k$ we have \[ | \mu(\gamma_m^x P^1_i \cap P^1_j) - \mu(\gamma_m^c Q^1_i \cap Q^1_j)| < d_H(C(x),C(c)) + \delta \] and similarly we can find a partition $\mathcal{Q}_2 = (Q^2_i)_{i=1}^k$ of $X_2$ such that for $m \leq n$ and $i,j \leq k$ we have \[ | \mu(\gamma_m^y P^2_i \cap P^2_j) - \mu(\gamma_m^d Q^2_i \cap Q^2_j)| < d_H(C(y),C(d)) + \delta. \] Let $\mathcal{Q} = (Q_i)_{i=1}^k$ be the partition of $X_1 \sqcup X_2$ given by $Q_i = Q^1_i \sqcup Q^2_i$. Write $t( d_H(C(x),C(c)) + \delta) + (1-t) (d_H(C(y),C(d)) + \delta) = r$. Then for all $m \leq n$ and $i,j \leq k$ we have \begin{align*} |\nu( \gamma_m^{tx + (1-t)y} P_i \cap P_j) - \nu( \gamma_m^{tc + (1-t)d} Q_i \cap Q_j)| & \leq |t\mu(\gamma_m^x P^1_i \cap P^1_j) - t\mu(\gamma_m^c Q^1_i \cap Q^1_j)| \\ & + | (1-t)\mu(\gamma_m^y P^2_i \cap P^2_j) - (1-t)\mu(\gamma_m^d Q^2_i \cap Q^2_j)| \\ &\leq r \end{align*} We have shown that $C(tx + (1-t) y) \subseteq B_r(C(tc + (1-t)d))$. The argument that $C(tc + (1-t)d) \subseteq B_r(C(tx + (1-t)y))$ is identical, so we omit it. Thus we conclude $d_H( C(tx + (1-t) y), C(tc + (1-t) d)) \leq r$ and $(5)$ holds. \end{proof} We note that $\mathrm{A}_\sim(\Gamma,X,\mu)$ in fact has additional structure in that it admits convex combinations of infinitely many elements. We first consider the case of a countable convex combination. If $\lambda_i \in [0,1]$ are such that $\sum_{i=1}^\infty \lambda_i = 1$ and $a_i \in \mathrm{A}_\sim(\Gamma,X,\mu)$ then we can naturally define an action $\sum_{i=1}^\infty \lambda_i a_i$ on the disjoint sum $\bigsqcup_{i=1}^\infty X_i$ with the $i$ copy of $X$ weighted by $\lambda_i$. It remains to check that this is independent of the choice of representatives $a_i$. \begin{proposition} If $a_i \prec b_i$ for all $i$, then $\sum_{i=1}^\infty \lambda_i a_i \prec \sum_{i=1}^\infty \lambda_i b_i$.\end{proposition} \begin{proof} Let $A_1,\ldots,A_k \subseteq \bigsqcup_{m=1}^\infty X_m$, $\epsilon > 0$ and $F \subseteq \Gamma$ finite be given. Choose $N$ such that $\sum_{m=N}^\infty \lambda_m < \frac{\epsilon}{2}$. For each $m < N$, consider the partition $A^m_1,\ldots,A^m_k$ of $X_m$ given by $A^m_i = A_i \cap X_m$. We can find for each $m<N$ a partition $B^m_1,\ldots,B^m_k$ such that for all $\gamma \in F$ and $i,j \leq k$ we have \[ |\mu(\gamma^{a_i} A^m_i \cap A^m_j) - \mu(\gamma^{b_i} B^m_i \cap B^m_j)| < \frac{\epsilon}{2}.\] Let $B_i = \bigsqcup_{m=1}^\infty B^m_i$. Then \begin{align*} \left \vert \mu\left(\gamma^{\sum_{m=1}^\infty \lambda_m a_m} A_i \cap A_j \right) - \mu\left(\gamma^{\sum_{m=1}^\infty \lambda_m b_m} B_i \cap B_j \right) \right \vert & \leq |\sum_{m=1}^N \lambda_m \mu(\gamma^{a_m} A^m_i \cap A^m_j) - \sum_{m=1}^N \lambda_m \mu(\gamma^{b_m} B^m_i \cap B^m_j)| \\ \hspace{1 in} &+ |\sum_{m=M}^\infty \lambda_m \mu(\gamma^{a_m} A^m_i \cap A^m_j) - \sum_{m=M}^\infty \lambda_m \mu(\gamma^{b_m} B^m_i \cap B^m_j)| \\ & \leq \sum_{m=1}^N \lambda_m |\mu(\gamma^{a_i} A^m_i \cap A^m_j) - \mu(\gamma^{b_i} B^m_i \cap B^m_j)| + \frac{\epsilon}{2} \\ & \leq \frac{\epsilon}{2} \left( \sum_{m=1}^N \lambda_m \right) + \frac{\epsilon}{2} \leq \epsilon. \end{align*} \end{proof} It is in fact possible to define integrals of weak equivalence classes of actions over a probability measure. Let $(Z,\eta)$ be a probability space and suppose that for each $z$ we have a probability space $(Y_z,\nu_z)$ and a measure-preserving action $\Gamma \curvearrowright^{a_z} (Y_z,\nu_z)$ such that the map $z \mapsto [a_z]$ from $(Z,\eta)$ to $A^*_\sim(\Gamma)$ is measurable, where $[a_z]$ is the weak equivalence class of $a_z$. Note that we do not require $(X_z,\nu_z)$ or $(Z,\eta)$ to be standard. Let $Y = \bigsqcup_{z \in Z} Y_z$ and put a measure $\nu$ on $Y$ by taking $\nu(A) = \int_Z \nu_z(A \cap Y_Z) d \eta(z)$. $Y$ will be a standard probability space isomorphic to $(X,\mu)$ if $(Z,\eta)$ is standard or $\eta$-almost all $(Y_z,\nu_z)$ are standard. Let $\Gamma \curvearrowright^a (Y,\nu)$ be given by letting $\Gamma$ act like $a_z$ on $Y_z$. We write $a = \int_Z a_z d \eta(z)$. We then have a map $\phi: Y \to Z$ given by letting $\phi(y)$ be the unique $z$ such that $y \in Y_z$. This is clearly a factor map from $a$ to $\iota_{Z,\eta}$ and $\nu = \int_Z \nu_z d \eta(z)$ is the disintegration of $\nu$ over $\eta$ via $\phi$. Thus Theorem 3.12 in guarantees that if $b_z$ are actions of $\Gamma$ on $(Y_z,\nu_z)$ with $b_z \sim a_z$ then if $b = \int_Z b_z d \eta(z)$ we have $a \sim b$. Therefore this construction gives a well-defined weak equivalence class of actions of $\gamma$. If we restrict $(Y_z,\nu_z)$ to be standard, then we in fact have a mapping from the space $M(\mathrm{A}_\sim(\Gamma,X,\mu))$ of probability measures on $\mathrm{A}_\sim(\Gamma,X,\mu)$ to $\mathrm{A}_\sim(\Gamma,X,\mu)$. \begin{lemma} For any $n,k$, and $(Z,\eta)$ and measurable assignment $z \mapsto a_z$, we have $C_{n,k}\left( \int_Z a_z d \eta(z) \right) \subseteq \cch \left( \bigcup_{z \in Z} C_{n,k}(a_z) \right)$. \end{lemma} \begin{proof} Fix $n,k$ and let $a = \int_Z a_z d \eta(z)$. Let $(X_z,\mu_z)$ by the underlying measure space of $a_z$. Let $\mathcal{L}$ be a countable dense subset of $\mathrm{MALG} \left ( \bigsqcup_{z \in Z} X_z, \int_Z \mu_z d \eta(z) \right )$, so that $\mathcal{L}^k$ is dense in the space of $k$-partitions of $\bigsqcup_{z \in Z} X_z$. Then $\{M_{\mathcal{A}}(a) \}_{\mathcal{A} \in \mathcal{L}^n }$ is dense in $C_{n,k} \left(\int_Z a_z d \eta(z) \right)$, so it suffices to show that each $M_{\mathcal{A}}(a) \in \cch \left( \bigcup_{z \in Z} C(a_z) \right)$. For each $\mathcal{A}$, the function $f_\mathcal{A}:Z \to \mathbb{R}^{n \times k \times k}$ given by $z \mapsto M_{\mathcal{A}_z}(a_z)$ is a Borel function, where $\mathcal{A}_z$ is the partition of $X_z$ given by $(A \cap X_z)_{A \in \mathcal{A}}$. Thus $M_{\mathcal{A}}(a) = \int_Z f_{\mathcal{A}}(a) d \eta(z)$. We may assume that $Z$ carries a Polish topology such that $f_{\mathcal{A}}$ is continuous for all $\mathcal{A} \in \mathcal{L}^n$. Choose a sequence of measures $(\nu_i)_{i=1}^\infty$ such that $\nu_i$ has finite support and $\nu_i \to \eta$ in the topology of $M(Z)$, the space of all Borel probability measures on $Z$. If we write $\nu_i = \sum_{j=1}^{j(i)} \alpha_j \delta_{z_j}$ then \[ \int_Z f_{\mathcal{A}}(z) d \nu_i(z) = \sum_{j=1}^{j(i)} \alpha_j f_{\mathcal{A}}(z_j) \in \ch \left( \bigcup_{z \in A} C(a_z) \right). \] Since $\nu_i \to \eta$, we have \[ \int_Z f_{\mathcal{A}}(z) d \nu_i(z) \to \int_Z f_{\mathcal{A}}(z) d \eta(z) \] which proves the lemma. \end{proof} \section{The structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for amenable $\Gamma$.} When $\Gamma$ is amenable, the structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ can be completely described using the notion of an invariant random subgroup. We begin with the following, the following extends Theorem 1.8 in . Recall that if $\Gamma \curvearrowright^a (X,\mu)$ is a measure-preserving action, we have a map $\mathrm{stab}_a: X \to \mathrm{Sub}(\Gamma)$ given by $x \mapsto \mathrm{stab}_a(x)$. The type of $a$ is the invariant random subgroup of $\Gamma$ given by $(\mathrm{stab}_a)_* \mu$. \begin{proposition} If $\Gamma$ is amenable and $a,b \in \mathrm{A}(\Gamma,X,\mu)$ then $\mathrm{type}(a) = \mathrm{type}(b)$ if and only if $a \sim b$. \end{proposition} \begin{proof} By type is an invariant of weak equivalence so suppose $\mathrm{type}(a) = \mathrm{type}(b)$. \\ \\ Let $X^a_\infty = \{x \in X: [\Gamma: \mathrm{stab}_a(x) = \infty] \}$ and $X^b_\infty = \{x \in X: [\Gamma: \mathrm{stab}_b(x)] = \infty \}$. Notice that $X^a_\infty$ is $a$-invariant and $X^b_\infty$ is $b$-invariant and since $\mathrm{type}(a) = \mathrm{type}(b)$, $\mu(X^a_\infty) = \mu(X^b_\infty)$. Suppose that $\mu(X^a_\infty) > 0$ and let $a_\infty = a \upharpoonright X^a_\infty$ with normalized measure $\frac{\mu \upharpoonright X^a_\infty}{\mu(X^a_\infty)}$ and define $b_\infty$ similarly. Then $\mathrm{type}(a_\infty) = \mathrm{type}(b_\infty)$ and these are concentrated on the infinite index subgroups of $\Gamma$, therefore $a_\infty \sim b_\infty$ by Theorem 1.8 (2) in . Thus to prove the proposition it suffices to show the following. Note that for this we do not require $\Gamma$ to be amenable. \begin{lemma} Suppose $a,b \in \mathrm{A}(\Gamma,X,\mu)$ are actions such that $\mathrm{type}(a) = \mathrm{type}(b)$ and these are concentrated on the finite-index subgroups of $\Gamma$. Then $a \sim b$. \end{lemma} \begin{proof} We may assume that $\theta = \mathrm{type}(a) = \mathrm{type}(b)$ is concentrated on the subgroups of index $n$ for some fixed $n$. Consider an $a$-orbit $C$. For each linear ordering $<_C^i$ of $C$, we get a homomorphism $\psi_C^i: \Gamma \to \mathrm{Sym}(n)$, where $\mathrm{Sym}(n)$ is the symmetric group on $n$ letters. Place a Borel linear order $\sqsubset$ on $\mathrm{Sym}(n)^\Gamma$. Let then $<^a_C = <_C^{i_0}$ be the linear order such that $\psi_C^{i_0}$ is $\sqsubset$-least among all the $\psi^i_C$. Write $\phi^a_C$ for $\psi^{i_0}_C$. Use this same construction to choose homomophisms $\phi^b_D$ for each $b$-orbit $D$. Write $\phi^a_x$ for $\phi^a_{[x]_{E_a}}$ and similarly $\phi^b_x$ for $\phi^b_{[x]_{E_b}}$.\\ \\ For a homomorphism $\phi: \Gamma \to \mathrm{Sym}(n)$ let $j_\phi$ be the corresponding action of $\Gamma$ on $\{1,\ldots,n\}$. Say $\phi$ is transitive if $j_{\phi}$ is transitive. Each transitive homomorphism $\phi: \Gamma \to \mathrm{Sym}(n)$ determines a conjugacy class $\mathcal{H}_\phi$ of index $n$ subgroups of $\Gamma$ as the stabilizers of $j_\phi$. For each $a$-orbit $[x]_{E_a}$ the stabilizers of the action of $\Gamma$ on $[x]_{E_a}$ also determine a conjugacy class $\mathcal{H}^a_x$ of index $n$ subgroups of $\Gamma$. Let $c$ be the action of $\mathrm{Sym}(n)$ on $\mathrm{Sym}(n)^\Gamma$ by $(f \cdot \phi)(\gamma)(k) = f \phi(\gamma) f^{-1}(k)$. Then $[\phi^a_x]_{E_c} = \left \{ \psi^i_{[x]_{E_a}}: <_{[x]_{E_a}}^i \mbox{ is a linear ordering of } [x]_{E_a} \right \}$. Let $\mathcal{L}$ be the set of all transitive homomorphisms $\phi: \Gamma \to \mathrm{Sym}(n)$ such that $\phi$ is $\sqsubset$-least in $[\phi]_{E_c}$. It is clear that for $\phi \in \mathcal{L}$, $\phi^a_x = \phi$ if and only if $\mathcal{H}^a_x = \mathcal{H}_\phi$. Similarly $\phi^b_x = \phi$ if and only if $\mathcal{H}^b_x = \mathcal{H}_\phi$. Thus for any $A \subseteq \mathcal{L}$, we have \begin{align*} \mu( \{x: \phi^a_x \in A \}) & = \mu( \{x: \mathcal{H}^a_x = \mathcal{H}_\phi \mbox{ for some } \phi \in A \}) \\ & = \mu( \{x: \mathrm{stab}_a(x) \mbox{ is conjugate to an element of } \mathcal{H}_\phi \mbox{ for some } \phi \in A \} ) \\ & = \theta(\{H \in \mathrm{Sub}(\Gamma): H \mbox{ is conjugate to an element of } \mathcal{H}_\phi \mbox{ for some }\phi \in A \}) \\ & = \mu( \{ x: \mathrm{stab}_b(x) \mbox{ is conjugate to an element of } \mathcal{H}_\phi \mbox{ for some } \phi \in A \} ) \\ & = \mu( \{x : \phi^b_x \in A\}). \end{align*} Now, fix a finite set $F \subseteq \Gamma$ and a partition $A_1,\ldots,A_m$ of $X$. For each map $\omega: F \to \mathrm{Sym}(n)$ let $X^a_\omega = \{x \in X: \phi^a_x \upharpoonright F = \omega \}$ and similarly $X^b_\omega = \{x \in X: \phi^b_x \upharpoonright F = \omega \}$. Then $(X^a_{\omega})_{\omega \in \mathrm{Sym}(n)^F}$ and $(X^b_\omega)_{\omega \in \mathrm{Sym}(n)^F}$ are finite decompositions of $X$ with $\mu(X^a_\omega) = \mu(X^b_\omega)$. For $k \leq n$ let \[ X^a_{\omega,k} = \left \{x \in X^a_{\omega}: x \mbox{ is in the } k \mbox{-position with respect to }<^a_{[x]_{E_a}} \right \} \] and define $X^b_{\omega,k}$ similarly. We claim that for each $k$ there is a measure-preserving bijection $S^a_{\omega,k}$ of $X^a_{\omega,k}$ with $X^a_{\omega,1}$. Let $\sqsubset_1$ be a wellordering of $\Gamma$. For each $\gamma \in \Gamma$ let \[X^a_{\omega,k,\gamma} = \left \{x \in X^a_{\omega,k}: \mbox{ the } \sqsubset_1\mbox{- least }\delta \in \Gamma \mbox{ with }\delta^a x \in X^a_{\omega,1} \mbox{ is equal to } \gamma \right \}. \] Put then $S^a_{\omega,k} \upharpoonright X^a_{\omega,k,\gamma} = \gamma^a$. In particular, this shows that $\mu(X^a_{\omega,k}) = \frac{\mu(X^a_\omega)}{n}$. We can perform the same construction for $b$ and we see that $\mu(X^b_{\omega,k}) = \frac{\mu(X^b_\omega)}{n}$. So $\mu(X^a_{\omega,1}) = \mu(X^b_{\omega,1})$ and hence there is a measure-preserving bijection $T_{\omega,1}$ of each $X^a_{\omega,1}$ with $X^b_{\omega,1}$. Define a measure-preserving bijection $T_{\omega}$ of $X^a_\omega$ with $X^b_\omega$ by letting $T_{\omega}(x) = (S^b_{\omega,k})^{-1} T S^a_{\omega,k}(x)$ for $x \in X^a_{\omega,k}$. Let then $T = \bigcup_{\omega \in \mathrm{Sym}(n)^F } T_\omega$ so $T \in \mathrm{Aut}(X,\mu)$.\\ \\ We claim that for all $\gamma \in F$ and all $x \in X$, we have $T(\gamma^a x) = \gamma^b T(x)$. Indeed, suppose $x \in X^a_{\omega,k}$ so that $x$ is in the $k$-position with respect to $<^a_{[x]_{E_a}}$. Then $\gamma^a x$ is in the $\phi^a_x(\gamma)(k) = \omega(k)$ position with respect to $<^a_{[x]_{E_a}}$ so $T(\gamma^a_x)$ is in the $\omega(k)$ position of the $E_b$-class $D$ such that $T_{\omega,1} S^a_{\omega,k}(x) \in D$, where $D$ has the canonical order $<^b_D$. On the other hand, $T(x) = T_{\omega}(x)$ is in the $k$-position of $D$ with respect to $<^b_D$. Hence $\gamma^b T(x)$ is in the $\phi^b_{T(x)}(\gamma)(k) = \omega(k)$ position of $D$ and we have the claim. Now, for $i \leq m$ putting $B_i = T(A_i)$ we have for any $\gamma$ in $F$ and $i,j \leq m$, \begin{align*} \mu(\gamma^b B_i \cap B_j) & = \mu(\gamma^b T(A_i) \cap T(A_j)) \\ & = \mu(T(\gamma^a A_i) \cap T(A_j) \\ & = \mu(T(\gamma^a A_i \cap A_j)) \\ & = \mu(\gamma^a A_i \cap A_j) \end{align*} and therefore $a \sim b$. \end{proof} \end{proof} In , Tucker-Drob shows that for amenable $\Gamma$, the space $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ of stable weak equivalence classes is homeomorphic to the space $\mathrm{IRS}(\Gamma)$ of invariant random subgroups of $\Gamma$. Indeed, $\mathrm{type}(a) = \mathrm{type}(b)$ if and only if $a \sim_s b$ and the map $A_{\sim_{s}}(\Gamma,X,\mu) \to \mathrm{IRS}(\Gamma)$ given by $a \mapsto \mathrm{type}(a)$ is a homeomorphism. So we have the following. \begin{corollary} For amenable $\Gamma$, $a \sim_s b$ if and only if $a \sim b$. \end{corollary} Moreover, let $x \in X$, $t \in [0,1]$ and $a,b \in \mathrm{A}(\Gamma,X,\mu)$ and consider the action $ta + (1-t)b$ on $t X_1 \sqcup (1-t) X_2$. We have $\mathrm{stab}_{ta + (1-t)b} = \mathrm{stab}_a(x)$ if $x \in X_1$ and $\mathrm{stab}_b(x)$ if $x \in X_2$. Thus for any $H \leq \Gamma$, $\{x: \mathrm{stab}_{ta + (1-t)b}(x) = H\} = \{x \in X_1: \mathrm{stab}_a(x) = H \} \sqcup \{x \in X_2: \mathrm{stab}_b(x) = H\}$ so for any $A \subseteq \mathrm{Sub}(\Gamma)$ we have \begin{align*} (t\mu_1+(1-t)\mu_2)( \{ x: \mathrm{stab}_{ta+(1-t)b}(x) \in A\}) &= (t\mu_1 + (1-t)\mu_2)( \{x \in X_1: \mathrm{stab}_a(x) \in A \} \\& \sqcup \{x \in X_2: \mathrm{stab}_b(x) \in A\}) \\ & = t \mu(\{x: \mathrm{stab}_a(x) \in H \}) + (1-t) \mu(\{x: \mathrm{stab}_b(x) \in A\}). \end{align*} Therefore $\mathrm{type}(ta+(1-t)b) = t( \mathrm{type}(a)) + (1-t)(\mathrm{type}(b))$ and Theorem follows. Note in particular that if $\Gamma$ is amenable then $ta + (1-t)a \sim a$, so for amenable groups $\mathrm{A}_\sim(\Gamma,X,\mu)$ is actually a convex space, not just a weak convex space.\\ \\ It is known (see for example ) that $\mathrm{IRS}(\Gamma)$ is a simplex in $C(\mathrm{Sub}(\Gamma))^*$, the dual of the Banach space $C(\mathrm{Sub}(\Gamma))$ of continuous functions on $\mathrm{Sub}(\Gamma)$. So by the classical Krein-Milman theorem we have that for amenable $\Gamma$, $\cch(\ex(\mathrm{A}_\sim(\Gamma,X,\mu))) = \mathrm{A}_\sim(\Gamma,X,\mu)$. We will prove an analogous result for general $\Gamma$ using other means. Moreover, $\ex(\mathrm{IRS}(\Gamma))$ is precisely the ergodic measures in $\mathrm{IRS}(\Gamma)$ so when $\Gamma$ is amenable, $\ex(\mathrm{A}_\sim(\Gamma,X,\mu))$ is the set of actions with ergodic type. \section{The structure of $\mathrm{A}_\sim(\Gamma,X,\mu)$ for general $\Gamma$.} Recall from that $E_0$ is the equivalence relation given by eventual equality on $2^{\mathbb{N}}$ and if $E$ is an equivalence relation on $X$ and $F$ is an equivalence relation on $Y$ then a Borel homomorphism from $E$ to $F$ is a Borel map $f: X \to Y$ such that $x_1 E x_2$ implies $f(x_1) F f(x_2)$. A equivalence relation $E$ on a measure space is said to be strongly ergodic (or $E_0$-ergodic) if for any homomorphism from $E$ to $E_0$, the preimage of some $E_0$-class is conull. By Proposition 5.6 in if $a$ is strongly ergodic then every $b$ with $b \sim a$ is ergodic. In particular, $\frac{1}{2}a + \frac{1}{2}a$ is not ergodic, so $\frac{1}{2}a + \frac{1}{2}a$ is not weakly equivalent to $a$ when $a$ is strongly ergodic. By Theorem 1.2 in , the Bernoulli shift $\Gamma \curvearrowright ([0,1]^\Gamma,\lambda^\Gamma)$ with $\lambda$ Lebesgue measure on $[0,1]$ is strongly ergodic when $\Gamma$ is nonamenable. Thus when $\Gamma$ is nonamenable, $\mathrm{A}_\sim(\Gamma,X,\mu)$ is not a convex space, only a weak convex space. We now prove Theorem \begin{proof} \textbf{\textbf{(of Theorem )}} Write $A = \mathrm{A}_\sim(\Gamma,X,\mu)$. Let $B = \cch(\ex(A))$ and suppose toward a contradiction that there exists $x \in A \setminus B$. Since $B$ is compact, $d(x,B) > 0$. Let $\alpha = \sup_{y \in A} d(y,B)$ and let $C = \{y \in A: d(y,B) = \alpha\}$. Then $C$ is nonempty, disjoint from $B$ and $C$ is a face of $A$.\\ \\ Let $\mathcal{F}$ be the family of faces of $C$, ordered by reverse inclusion. Suppose $\{F_i\}_{i \in I}$ is a linearly ordered subset of $\mathcal{F}$ and consider $\bigcap_{i \in I} F_i$. If $x,y \in C$ and $0 <t <1$ are such that $tx + (1-t)y \in \bigcap_{i \in I} F_i$, then $x,y \in F_i$ for each $i$ since each $F_i$ is a face. Hence $\bigcap_{i \in I} F_i$ is a face. It is nonempty by compactness. So Zorn's Lemma guarantees there exist minimal elements of $\mathcal{F}$. Let $F$ be such a minimal element.\\ \\ Choose $y \in F$ and suppose toward a contradiction that there exists $y' \in F$ with $y' \notin \cch(\{y\})$. Then $\cch(\{y\})$ is a compact convex set, so letting $G = \left \{z \in F: d(z,\cch(\{y\})) = \sup_{w \in F} d(w,\cch(\{y\})) \right \}$, $G$ is a nonempty face of $F$ disjoint from $\cch(\{y\})$, contradiction the minimality of $F$. So for all $y \in F$ we have $F \subseteq \cch(\{y\})$. Fix such a $y$. Note that $\cch(\{y\}) = \ch(\{y\})$. We claim that $y$ is an extreme point of $C$. Assuming this, since $C$ is a face of $A$ we have that $y$ is an extreme point of $A$ and we have a contradiction to the hypothesis that $C \cap B = \emptyset$.\\ \\ Suppose first that there do not exist $a,b \in C$ and $0 < t <1$ such that $y = ta + (1-t)b$. Then $y$ is an extreme point of $C$ be definition. So let $a,b \in C$ and $0 < t < 1$ be such that $y = ta + (1-t)b$. We must show that $y \sim a \sim b$. Since $F$ is a face of $C$, we have $a,b \in F$. Thus we can write $a = \sum_{i=1}^n s_i y$ and $b = \sum_{i=1}^k r_i y$ for $s_i,r_i \in [0,1]$. By Proposition and associativity we have $y \sim \left( \sum_{i=1}^n t s_i y + \sum_{i=1}^k (1-t)r_i y \right)$. Since $0 < t < 1$, iterating this argument we find that for any $\delta > 0$, there is $m \in \mathbb{N}$ and $(\lambda_i)_{i=1}^m \subseteq [0,1]$ such that $\lambda_i \leq \delta$ for all $i$ and $y \sim \sum_{i=1}^m \lambda_i y$.\\ \\ We claim that this implies $y \sim \kappa y + (1-\kappa)y$ for all $\kappa \in [0,1]$. Note that $\kappa y + (1-\kappa) y$ is isomorphic to $\iota_{\kappa, 1-\kappa} \times y$, where $\iota_{\kappa, 1- \kappa}$ is the trivial action of $\Gamma$ on $(\{0,1\},m_\kappa)$ where $m_\kappa( \{0\}) = \kappa$ and $m_\kappa( \{1\}) = 1-\kappa$. Hence $y$ is a factor of $\kappa y + (1-\kappa)y$ and it thus suffices to show $\kappa y + (1-\kappa)y \prec y$.\\ \\ Let $X_1,X_2$ be two copies of $X$, let $n,k \in \mathbb{N}$, $\epsilon > 0$ and a partition $\mathcal{P} = (P_i)_{i=1}^k$ of $X_1 \sqcup X_2$ be given. As before, we get a partition $\mathcal{P}_1 = (P^1_i)_{i=1}^k$ with $P^1_i = P_i \cap X_1$ of $X_1$ and similarly a partition $\mathcal{P}_2 = (P^2_i)_{i=1}^k$ with $P^2_i = P_i \cap X_2$ of $X_2$. Now, choose $\delta < \frac{\epsilon}{2}$. Then we can find $m$ and $(\lambda_p)_{p=1}^m$ such that $y \sim \sum_{p=1}^m \lambda_p y$ and for some $l \leq m$ we have $\kappa - \frac{\epsilon}{2} \leq \sum_{p=1}^l \lambda_p \leq \kappa$. Let now $X'_p$ be a copy of $X$ for each $p \leq m$, and for $q \in \{0,1\}$ let $P^q_{i,p}$ be the corresponding copy of $P^q_i$ sitting in $X'_p$. Let $\mathcal{Q} = (Q_i)_{i=1}^k$ be the partition of $\bigsqcup_{p=1}^m X'_p$ given by $Q_i = \left( \bigsqcup_{p=1}^l P^1_{i,p} \right) \sqcup \left( \bigsqcup_{p= l+1}^m P^2_{i,p} \right)$. Then for $s \leq n$ and $i,j \leq k$ we have \begin{align*} \Bigg \vert (\kappa \mu & + (1-\kappa) \mu)( \gamma_s^{\kappa y + (1-\kappa) y} P_i \cap P_j) - \left ( \sum_{p=1}^m \lambda_p \mu \right ) \left( \gamma_s^{\sum_{p=1}^m \lambda_p y} Q_i \cap Q_j \right) \Bigg \vert\\ & \leq \left \vert \kappa \mu(\gamma_s^y P_i^1 \cap P_j^1) - \left( \sum_{p=1}^l \lambda_p \mu (\gamma_s^y P^1_{i,p} \cap P^1_{j,p} ) \right) \right \vert + \left \vert (1- \kappa) \mu(\gamma_s^y P_i^2 \cap P_j^2) - \left( \sum_{p=l+1}^m \lambda_p \mu (\gamma_s^y P^2_{i,p} \cap P^2_{j,p} ) \right) \right \vert \end{align*} \begin{align*} & = \left \vert \kappa \mu(\gamma_s^y P_i^1 \cap P_j^1) - \left( \sum_{p=1}^l \lambda_p \right) \mu (\gamma_s^y P^1_i \cap P^1_j ) \right \vert + \left \vert (1-\kappa) \mu(\gamma_s^y P_i^2 \cap P_j^2) - \left( \sum_{p=l+1}^m \lambda \right) \mu (\gamma_s^y P^2_i \cap P^2_j )\right \vert \\ &= \left \vert \left ( \kappa - \sum_{p=1}^l \lambda_p \right) \mu( \gamma^y_s P^1_i \cap P^1_j) \right \vert + \left \vert \left( (1-\kappa) - \sum_{p=l+1}^m \lambda_p \right) \mu(\gamma_s^y P^2_i \cap P^2_j) \right \vert \\ & \leq \left \vert \left ( \kappa - \sum_{p=1}^l \lambda_p \right) \right \vert + \left \vert \left( (1-\kappa) - \sum_{p=l+1}^m \lambda_p \right) \right \vert \leq \frac{\epsilon}{2} + \frac{\epsilon}{2} = \epsilon. \end{align*} Since $y \sim \sum_{p=1}^m \lambda_p y$, $\kappa y + (1-\kappa)y \prec y$ and we are done. \end{proof} We note that a metrizable topological vector space $V$ is locally convex if and only if its topology is induced by a countable family of seminorms $ \left(| \cdot |^V_n \right)_{n=1}^\infty$. Then $p(v,w) = \sum_{n=1}^\infty \frac{1}{2^n} | v - w |^V_n$ is a compatible metric on $V$, which is easily seen to obey Lemma . Thus the technique used to prove Theorem works to prove the metrizable case of the classical Krein-Milman theorem using only the convex and metric structure of $V$, not the vector space structure in the form of linear functionals.\\ \\ Before proving Theorem , we briefly discuss the ergodic decomposition in the context of weak equivalence classes. Suppose $a \in \mathrm{A}(\Gamma,X,\mu)$ and $a = \int_Z a_z d \eta(z)$ is the ergodic decomposition of $a$, that is to say we have a factor map $\pi: (X,\mu) \to (Z,\eta)$ such that if $\mu = \int_Z \mu_z d \eta(z)$ is the disintegration of $\mu$ over $(Z,\eta)$ via $\pi$ then $\mu_z( \pi^{-1}(z) ) = 1$ and $\Gamma \curvearrowright^a (\pi^{-1}(z),\mu_z)$ is isomorphic to $a_z$. Furthermore, the assignment $z \mapsto \mu_z$ from $(Z,\eta) \to M_a(X)$ is Borel, where $M_a(X)$ is the space of $a$-invariant probablity measures on $X$ (we may assume here that $X$ is a Polish space). Recall that $A^*_\sim(\Gamma)$ is the space of weak equivalence classes of all measure-preserving actions of $\Gamma$, including those actions on finite space. $A^*_\sim(\Gamma)$ is topologized using the exact same metric as we use to topologize $\mathrm{A}_\sim(\Gamma,X,\mu)$. We would like to conclude that the assignment $z \mapsto [a_z]$ is measurable from $(Z,\eta)$ to $A^*_\sim(\Gamma)$, where $[a_z]$ is the weak equivalence class of $a_z$. This is a consequence of the following lemma. \begin{lemma} Let $\Gamma \curvearrowright^a Y$ be a Borel action of $\Gamma$ on a Polish space $Y$. Then the map $\Theta$ from $M_a(Y)$ to $A^*_\sim(\Gamma)$ given by $\nu \mapsto [a_\nu]$ is Borel, where $[a_\nu]$ is the weak equivalence class of the measure preserving action $a_\nu = \Gamma \curvearrowright^a (Y,\nu)$. \end{lemma} \begin{proof} Fix a measure $\nu \in M_a(Y)$ and consider $\Theta^{-1}(U)$ where \[U = \{ [a] \in A_{\sim}^*(\Gamma): d_H( C_{n,k}(a_\nu),C_{n,k}(a)) < \epsilon \mbox{ for all }n,k \leq N \} \] for some $N \in \mathbb{N}$ and $\epsilon > 0$, so $U$ is a basic open neighborhood of $\Theta(\nu) = a_\nu$. Since \[ U = \bigcup_{m=1}^\infty \bigcap_{n,k =1}^N \left \{ [b] \in A_{\sim}^*(\Gamma): d_H(C_{n,k}(a_\nu), C_{n,k}(b) ) \leq \epsilon - \frac{1}{m} \right\}, \] it suffices to show $\Theta^{-1}(V)$ is Borel for a set $V$ of the form \[V = \{ [b] \in A_{\sim}^*(\Gamma): d_H( C_{n,k}(a_\nu), C_{n,k}(b) ) \leq r \}. \] Fixing $n$ and $k$ we write $C(b)$ for $C_{n,k}(b)$. Now, let $K$ and $L$ be compact subsets of a compact Polish space $W$ with metric $p$, let $D_K$ be dense in $K$ and $D_L$ be dense in $L$. We have \begin{align*} d_H(K,L) \leq r & \iff \max_{x \in K} \inf_{y \in L} p(x,y) \leq r \mbox{ and } \max_{y \in L} \inf_{x \in K} p(y,x) \leq r \\ & \iff ( \forall x \in K )(\forall \delta > 0 )( \exists y \in L)( p(x,y) < r + \delta )\\ &\hspace{2 in} \wedge ( \forall y \in L )( \forall \delta > 0)( \exists x \in K)( p(y,x) < r + \delta ) \\ & \iff (\forall x \in D_K)( \forall \delta > 0)( \exists y \in D_L)( p(x,y) < r + \delta) \\ & \hspace{2 in}\wedge (\forall y \in D_L)(\forall \delta > 0) (\exists y \in D_L)( p(y,x) < r + \delta ) \end{align*} If $\mathbf{L}$ is a countable algebra generating the Borel $\sigma$-algebra $\mathbf{B}(Y)$ of $Y$, then $\mathbf{L}$ is dense in $\mathrm{MALG}(Y,\rho)$ for any Borel probability measure $\rho$ on $Y$. Regarding a partition of $Y$ into $k$ pieces as an element of $\mathbf{B}(Y)^N$ and considering $\mathbf{L}^k$, we see that there exists a fixed countable family $(\mathcal{A}_m)_{m=1}^\infty$ of partitions of $Y$ such that for any Borel probability measure $\rho$ on $Y$, $(\mathcal{A}_m)_{m=1}^\infty$ is dense in the set of $k$-partitions of $X$ with topology inherited from $\mathrm{MALG}(Y,\rho)$. We may further assume that each element of each $\mathcal{A}_m$ is clopen. This implies that the set $(M^{\mathcal{A}_m}(a_\rho))_{m=1}^\infty$ is dense in $C(a_\rho)$ for any Borel probability measure $\rho$. Therefore we have \begin{align*} V & = \left( \bigcap_{m=1}^\infty \bigcap_{l = 1}^\infty \bigcup_{i=1}^\infty \left \{ b \in A^*_\sim(\Gamma): \left \vert M^{\mathcal{A}_m}(a_\nu) - M^{\mathcal{A}_i}(b) \right \vert < r + \frac{1}{l} \right \} \right) \\ &\hspace{1 in} \cap \left( \bigcap_{m=1}^\infty \bigcap_{l = 1}^\infty \bigcup_{i=1}^\infty \left \{ b \in A^*_\sim(\Gamma): \left \vert M^{\mathcal{A}_i}(a_\nu) - M^{\mathcal{A}_m}(b) \right \vert < r + \frac{1}{l} \right \} \right). \end{align*} Now, $|M^{\mathcal{A}_i}(a_\nu) - M^{\mathcal{A}_m}(a_{\rho})| < s$ if and only if $|\nu(\gamma^a A_i^j \cap A_i^t) - \rho( \gamma^a A_m^i \cap A_m^t)| <s$ for all $A_i^j,A_i^t \in \mathcal{A}_i$ and $A_m^i,A_m^t \in \mathcal{A}_m$. Since for any pair $J_1,J_2 \subseteq Y$ the set $\{\rho: |\nu(J_1) - \nu(J_2)| < s\}$ is Borel, we see \[\Theta^{-1} \left( \left \{ b \in A^*_\sim(\Gamma): \left \vert M^{\mathcal{A}_i}(a_\nu) - M^{\mathcal{A}_m}(b) \right \vert < r + \frac{1}{l} \right \} \right) \] is Borel and consequently $\Theta^{-1}(V)$ is Borel. \end{proof} We now prove Theorem \begin{proof}\textbf{\textbf{(of Theorem )}} Let $\Theta: Z \to A^*_\sim(\Gamma)$ be the map sending each point in $z$ to the weak equivalence class $[a_z]$, so $\Theta$ is measurable by Lemma . Suppose towards a contradiction that the theorem fails. Then for every set $Z' \subseteq Z$ with $\eta(Z') = 1$, there is more than one weak equivalence class in the set $\{[a_z]: z \in Z' \}$. Equivalently, the measure $\Theta_* \eta$ on $A^*_\sim(\Gamma)$ is not supported on a single point. We can thus split $A^*_\sim(\Gamma)$ into two disjoint sets $Y_1,Y_2$ such that $0 < \Theta_*\eta(Y_1), \Theta_*\eta(Y_2) <1$. Letting $A_i = \Theta^{-1}(Y_i)$, we get disjoint measurable sets $A_1,A_2 \subseteq Z$ such that $0 < \eta(A_1), \eta(A_2) < 1$ and for all $z \in A_1$ and all $w \in A_2$ we have that $z \nsim w$.\\ \\ Recall that for a measure-preserving action $b$ of $\Gamma$ and $n,k \in \mathbb{N}$ the set $C_{n,k}(a) \subseteq [0,1]^{n \times k \times k}$ was defined in Section . \begin{lemma} For any action $b$ of $\Gamma$ on a probability space $(Y,\nu)$, we have $\cch(C_{n,k}(b)) \subseteq C_{n,k}(\iota \times b)$. \end{lemma} \begin{proof} Write $C_{n,k}(b) = C(b)$. Suppose $x \in \cch(C(b))$. Then we can find points $(x_i)_{i=1}^\infty$ such that $\lim_{i \to \infty} x_i = x$ and each $x_i$ has the form $x_i = \sum_{j=1}^{j(i)} \alpha_i^j x_i^j$ for $(x_i^j)_{j=1}^{j(i)} \subseteq C(b)$ and $(\alpha_i^j)_{j=1}^{j(i)} \subseteq [0,1]$ with $\sum_{j=1}^{j(i)} \alpha^i_j = 1$ for each $i$. Without loss of generality we may assume that each $x_i^j$ has the form $M^{\mathcal{A}_i^j}(b)$ for a partition $\mathcal{A}_i^j = (A_{i,l}^j)_{l=1}^k$ of $Y$ into $k$ pieces. Fixing $i$ consider the action $\sum_{j=1}^{j(i)} \alpha_i^j b$ on the space $\left( \bigsqcup_{j=1}^{j(i)} Y_j, \sum_{j=1}^{j(i)} \alpha_i^j \nu_j \right)$, where each $(Y_j,\nu_j)$ is a copy of $(Y,\nu)$. Let $\mathcal{B} = (B_l)_{l=1}^k$ be the partition of $\bigsqcup_{j=1}^{j(i)} Y_j$ given by letting $B_l = \bigsqcup_{j=1}^{j(i)} A_{i,l}^j$, where $A_{i,l}^j$ sits inside the $j$ copy of $Y$. For any $p \leq n$ and $l,m \leq k$ and $x \in [0,1]^{n \times k \times k}$ let $(x)_{p,l,m}$ be the $p,l,m$ coordinate of $x$. We then have \begin{align*} \left( M^{\mathcal{B}} \left ( \sum_{j=1}^{j(i)} \alpha_i^j b \right) \right)_{p,l,m} &= \left( \sum_{j=1}^{j(i)} \alpha_i^j \nu_j \right) \left ( \gamma_p^{\sum_{j=1}^{j(i)} \alpha_i^j b} B_l \cap B_m \right) \\ & = \sum_{j=1}^{j(i)} \left( \alpha_i^j \nu_j (\gamma_p^b A_{i,l}^j \cap A_{i,m}^j)\right) \\ & = \sum_{j=1}^{j(i)} \alpha_i^j \left( M^{\mathcal{A}_i^j}(b) \right)_{p,l,m} \end{align*} Therefore \[ M^{\mathcal{B}} \left ( \sum_{j=1}^{j(i)} \alpha_i^j b \right) = \sum_{j=1}^{j(i)} \alpha_i^j \left( M^{\mathcal{A}_i^j}(b) \right) = x_i. \] We have shown that $x_i \in C \left( \sum_{j=1}^{j(i)} \alpha_i^j b \right)$. Since $\sum_{j=1}^{j(i)} \alpha_i^j b$ is a factor of $b \times \iota$, we have $x_i \in C(b \times \iota)$. Since $\lim_{i \to \infty} x_i = x$ and $C(b \times \iota)$ is closed, the lemma follows. \end{proof} It is clear that for any two measure-preserving actions $b,c$ we have $b \prec c$ if and only if $C_{n,k}(b) \subseteq C_{n,k}(c)$ for all $n,k$. We claim that there are disjoint subsets $A_3, A_4 \subseteq Z$ of positive measure such that for some pair $n_0,k_0$, every $z \in A_3$ and every $w \in A_4$ we have $C_{n_0,k_0}(a_z) \nsubseteq \cch(C_{n_0,k_0}(a_w))$. For $z \in A_3$ let $R_z = \{ w \in A_2: a_z \nprec a_w \}$. Since $a_z$ is ergodic, $a_z \prec a_w \times \iota$ implies $a_z \prec a_w$. Therefore $R_z = \{w \in A_2: a_z \nprec a_w \times \iota \}$. \\ \\ Assume first that there is a set $D_3 \subseteq A_1$ with $\eta(D_3) > 0$ such that for each $z \in D_3$ we have $\eta(R_z) > 0$. Write $\hat{K}$ for $\cch(K)$. By Lemma we can write $R_z = \bigcup_{n,k = 1}^\infty R_z^{n,k}$ where $R_z^{n,k} = \left \{ w \in A_2: C_{n,k}(a_z) \nsubseteq \widehat{C_{n,k}(a_w)} \right \}$. Thus for each $z$ there is a lexicographically least pair $(n_z,k_z)$ such that $\eta(R_z^{n_z,k_z}) > 0$. Therefore there is a pair $n_0,k_0$ and a set $D_4 \subseteq D_3$ such that $\eta(D_4) > 0$ and for all $z \in D_4$ we have $\eta(R_z^{n_0,k_0}) > 0$. Fixing $n_0$ and $k_0$ we write $C(b)$ for $C_{n_0,k_0}(b)$. Let $(w_j)_{j=1}^\infty \subseteq A_2$ be a sequence of points such that the family $\left(\widehat{C(a_{w_j})} \right)_{j=1}^\infty$ is dense in the space $\left \{ \widehat{C(a_w)}: w \in A_2 \right\}$ with respect to the Hausdorff metric $d_H$ on the space on compact subsets of $[0,1]^{n_0 \times k_0 \times k_0}$. Let then $F_{j,l} = \left\{ w \in A_2: d_H\left( \widehat{C(a_w)},\widehat{C(a_{w_j})} \right) < \frac{1}{l} \right\}$.\\ \\ Fix $z \in D_4$ and choose $w \in R_z^{n_0,k_0}$. By hypothesis there is $\epsilon > 0$ such that $C(a_z) \nsubseteq B_{\epsilon} \left(\widehat{C(a_w)} \right)$, where $B_{\epsilon}(K)$ denotes the ball of radius $\epsilon$ around $K$. Then if we choose $j$ so that $d_H \left(\widehat{C(a_{w_j})},\widehat{C(w)} \right) < \frac{\epsilon}{2}$ and $l$ so that $\frac{1}{l} < \frac{\epsilon}{2}$ we have $w \in F_{j,l} \subseteq R_z^{n_0,k_0}$. Hence there is a subset $\mathcal{J} \subseteq \mathbb{N}^2$ such that $R_z^{n_0,k_0} = \bigcup_{(j,l) \in \mathcal{J}} F_{j,l}$. So for each $z$ we can choose a lexicographically least pair $(j_z,l_z)$ such that $\eta(F_{j_z,l_z}) > 0$ and $F_{j_z,l_z} \subseteq R_z^{n_0,k_0}$. There is then a pair $(j_0,l_0)$ and a set $E_3 \subseteq D_3$ with $\eta(E_3) > 0$ such that $\eta(F_{j_0,l_0}) > 0$ and for all $z \in E_3$ and all $w \in F_{j_0,l_0}$ we have $C(a_z) \nsubseteq \widehat{C(a_w)}$. So take $A_3 = E_3$ and $A_4 = F_{j_0,l_0}$. Thus we are left with the case $\eta(R_z) = 0$ for almost all $z \in A_1$. Then for almost all $w \in A_2$ and almost all $z \in A_1$ we must have $a_w \nprec a_z$, so a symmetric argument gives the claim.\\ \\ Given a (real) topological vector space $V$, we say a hyperplane in $V$ is a set of the form $H_{\ell,\alpha} = \{v \in V: \ell(v) = \alpha \}$ for some continuous linear functional $\ell$ and $\alpha \in \mathbb{R}$. Given disjoint compact subsets $W_1,W_2 \subseteq V$ we say that $H_{\ell,\alpha}$ separates $W_1$ from $W_2$ if $W_1 \subseteq \{v \in V: \ell(v) < \alpha \}$ and $W_2 \subseteq \{v \in V: \ell(v) > \alpha \}$. \begin{lemma} Let $S \subseteq \mathbb{R}^n$ be compact. Then there is a countable family $(H_i)_{i=1}^\infty$ of hyperplanes such that for any $x \in S$ and any compact convex $W \subseteq S$ there is $i$ so $H_i$ separates $\{x\}$ from $W$. \end{lemma} \begin{proof} Let $(\ell_j)_{j=1}^\infty$ be a countable set of linear functionals which is dense in the $\sup$ norm on $S$. Enumerate $\mathbb{Q}$ as $(q_m)_{m=1}^\infty$ and let $H_{j,m} = \{s \in S: \ell_j(s) = q_m\}$. Given $x$ and $W$, by Hahn-Banach find a linear functional $\ell$ and $\alpha \in \mathbb{R}$ so that $H = H_{\ell,\alpha}$ separates $x$ from $W$. Let $r = \min \left( \inf_{h \in H} ||x -h ||, \inf_{ \substack{ h \in H, \\ w \in W}} || h - w|| \right)$ so $r > 0$. Then choose $m$ so $|q_m - \alpha| < \frac{r}{2}$ and $j$ so $\sup_{s \in S} |\ell(s) - \ell_j(s)| < \frac{r}{2}$. Then $H_{j,m}$ separates $x$ from $W$. \end{proof} Now take $S = [0,1]^{n_0 \times k_0 \times k_0}$ and fix a family $(H_i)_{i=1}^\infty$ of hyperplanes as in the lemma. Since $\widehat{C(a_w)}$ is compact convex for each $w \in A_4$ and for all $z \in A_3$ we have $C(a_z) \nsubseteq \widehat{C(a_w)}$, for each pair $(z,w) \in A_3 \times A_4$ there is an index $i(z,w)$ and a point $x_{z,w} \in C(a_z)$ such that $H_{i(z,w)}$ separates $x_{z,w}$ from $\widehat{C(a_w)}$. Fix $z \in A_3$. Taking $(w_j)_{j=1}^\infty$ as before, for $(j,l) \in \mathbb{N}^2$ let $G_{j,l} = \left \{w \in A_4: d_H\left( \widehat{C(a_w)}, \widehat{C(a_{w_j})}\right) < \frac{1}{l} \right \}$. Choosing $w \in A_4$, let $\epsilon = d_H \left( \widehat{C(w)}, H_{i(z,w)} \right)$ so $\epsilon > 0$. Finding $j_{z,w}$ so $d_H \left( \widehat{C(w_j)}, \widehat{C(w)} \right) < \frac{\epsilon}{2}$ and $l_{z,w}$ so $\frac{1}{l} < \frac{\epsilon}{2}$ we have $w \in G_{j_{z,w},l_{z,w}}$ and $H_{i(z,w)}$ separates $x_{z,w}$ from $\widehat{C(u)}$ for all $u \in G_{j_{z,w},l_{z,w}}$. Then we have $A_4 = \bigcup_{\substack{(j_{z,w},l_{z,w}): \\w \in A_4 }} G_{j_{z,w},l_{z,w}}$ so we can find $w_0$ so that $\eta\left(G_{j_{z,w_0},l_{z,w_0}} \right) > 0$. Let then $G_z = G_{j_{z,w_0},l_{z,w_0}}$, $x_z = x_{z,w_0}$ and $i(z) = i(z,w_0)$ so that $H_{i(z)}$ separates $x_z$ from $\widehat{C(u)}$ for all $u \in G_z$. Since the $G_z$ were chosen from a countable family, we can find a set $A_5 \subseteq A_3$ of positive measure such that $G_z = G$ is the same for all $z \in A_5$. We can then find an index $i$ and a set $A_6 \subseteq A_5$ of positive measure such that for all $z \in K$, $H_i = H$ separates $x_z$ from $\widehat{C(u)}$ for all $u \in G$. $H$ splits $[0,1]^{n \times k \times k}$ into two closed convex sets $H_+$ and $H_-$, where $H_+$ contains the $x_z$ and $H_-$ contains the $C(u)$.\\ \\ For $S \subseteq Z$ with $\eta(S) > 0$ let $\eta_S = \frac{ \eta \upharpoonright S}{\eta(S)}$ be normalized measure on $S$. By Lemma we have $C \left ( \int_G a_u d \eta_G(u) \right ) \subseteq \cch \left( \bigcup_{u \in G} C(u) \right) \subseteq H_-$. Write $A_6 = \bigcup_{p = 1}^\infty A_6^p$, where $A_6^p = \left \{ z \in A_6: d_H( x_z, H) \geq \frac{1}{p} \right \}$ and find $p$ so $\eta(A_6^p) > 0$. Letting $K = A_6^p$, for all $z \in K$, $x_z$ is an element of the closed convex set $H_+^p = \{ y \in H_+: d_H(y,H) \geq \frac{1}{p} \}$ and $H_+^p$ is disjoint from $H_-$. We have $\int_K x_z d \eta_K(z) \in C \left ( \int_K a_z d \eta_K(z) \right )$ and $\int_K x_z d \eta_K(z) \in H_+^p$. Since $C \left ( \int_G a_u d \eta_G(u) \right ) \subseteq H_-$ we see that $C \left ( \int_K a_z d \eta_K(z) \right ) \nsubseteq C \left ( \int_G a_u d \eta_G(u) \right )$ and it follows that $\int_K a_z d \eta_K(z) \nsim \int_G a_u d \eta_G(u)$. Let $L_1 = K, L_2 = G$ then there is $i \in \{1,2\}$ with $\int_{L_i} a_z d \eta_{L_i}(z) \nsim a$. Since $0 < \eta(L_i) < 1$, we can write \[a = \eta(L_i) \left ( \int_{L_i} a_z d \eta_{L_i}(z) \right) + \eta( Z \setminus L_i) \left ( \int_{Z \setminus L_i} a_z d \eta_{Z \setminus L_i}(z) \right) \] which contradicts our assumption that $a$ is an extreme point. \end{proof} We now prove Theorem . Recall that the uniform topology on $\mathrm{Aut}(X,\mu)$ is given by the metric $d_u(T,S) = \mu(\{x: Tx \neq Sx \})$. If $\mathcal{P} = \{P_1,\ldots,P_p \}$ is a partition of a space on which $\mathbb{F}_N$ acts by an action $a$, $J \subseteq \mathbb{F}_N$ is finite and $\tau: J \to p$ let $P^a_{\tau} = \bigcap_{\gamma \in J} \gamma^a P_{\tau(\gamma)}$. \begin{proof} \textbf{\textbf{(of Theorem )}} Let $a$ be a free action of $\mathbb{F}_N$. By replacing $a$ with $a \times \iota$ if necessary, we may assume that for each $n,k$ the set $C_{n,k}(a)$ is closed and convex. Fix integers $n_0$ and $k_0$ and $\epsilon > 0$. It is enough to find a free ergodic action $b$ of $\mathbb{F}_N$ such that for all $n \leq n_0$ and $k \leq k_0$ we have $d_H (C_{n,k}(a),C_{n,k}(b)) < \epsilon$. Let $\{ \gamma_1,\ldots,\gamma_{n_0} \} = F_0$ be the finite subset of $\mathbb{F}_N$ under consideration. Let $s = s_{\mathbb{F}_N}$ be the Bernoulli shift of $\mathbb{F}_N$ acting on $\left( 2^{\mathbb{F}_N}, \nu \right)$ where $\nu$ is the product measure. For any action $c$ of $\mathbb{F}_N$ on $(X,\mu)$ and $\gamma \in \mathbb{F}_N$ we have \[ \{ (x,y) \in X \times 2^{\mathbb{F}_N}: \gamma^{c \times s} (x,y) \neq \gamma^{a \times x}(x,y) \} = \{ x \in X: \gamma^c x \neq \gamma^a x \} \times Y \] and hence \[ (\mu \times \nu)(\{ (x,y) \in X \times 2^{\mathbb{F}_N}: \gamma^{c \times s} (x,y) \neq \gamma^{a \times x}(x,y) \}) = \mu(\{ x \in X: \gamma^c x \neq \gamma^a x \}). \] Assume $d_u(\gamma^a,\gamma^c) < \frac{\epsilon}{16}$ for all $\gamma \in F_0$. Then for any measurable partition $\mathcal{A} = A_1,\ldots,A_k$ of $X \times 2^{\mathbb{F}_N}$, all $\gamma \in F_0$ and all $i,j \leq k$ we have \[| (\mu \times \nu)(\gamma^{a \times s} A_i \cap A_j) - (\mu \times \nu)(\gamma^{c \times s} A_i \cap A_j)| < \frac{\epsilon}{16} \] for all $\gamma \in F_0$. In the notation of Section , $\rho \left(M_{n,k}^\mathcal{A}(a \times s),M_{n,k}^\mathcal{A}(c \times s) \right ) < \frac{\epsilon}{16}$ where $\rho$ is the supremum metric on $[0,1]^{n \times k \times k}$. Choose a finite collection $\mathscr{L}$ of measurable subsets of $X \times 2^{\mathbb{F}_N}$ such that for every measurable partition $\mathcal{A}$ of $X \times 2^{\mathbb{F}_N}$ there is a partition $\mathcal{B} \subseteq \mathscr{L}$ such that $\rho\left( M^\mathcal{A}_{n,k}(a \times s), M^\mathcal{B}_{n,k}(a \times s) \right) < \frac{\epsilon}{16}$. Then for every such $\mathcal{A}$ there exists $\mathcal{B} \subseteq \mathscr{L}$ such that $\rho \left( M^\mathcal{A}_{n,k}(c \times s), M^\mathcal{B}_{n,k}( c \times s) \right) < \frac{3 \epsilon}{16}$. \\ \\ For $\gamma \in \mathbb{F}_N$ let $\pi_{\gamma}: 2^{\mathbb{F}_N} \to 2 $ be projection onto the $\gamma$ coordinate. For $i \in \{0,1\}$ let $S_i = \pi_e^{-1}(\{i\})$ and put $\mathcal{S} = \{S_1,S_2\}$. Choose now a finite partition $\mathcal{R} = \{R_1,\ldots,R_r \}$ of $X$ and a finite subset $F \subseteq \mathbb{F}_N$ containing $F_0$ such that for every $A \in \mathcal{L}$ there are sets $R_j$ with $1 \leq j \leq r$ and a family of functions $(\tau_j)_{j=1}^t$ with $\tau_j: F \to 2$ such that \[ \mu \left ( \left( \bigsqcup_{j=1}^t R_j \times S^s_{\tau_j} \right) \triangle A \right) < \frac{\epsilon}{16}. \] Write $\mathcal{P} =\mathcal{R} \times \mathcal{S}$. We can identify a function $\theta: F \to r \times 2$ with a pair $(\sigma,\tau)$ where $\sigma:F \to r$ and $\tau:F \to 2$ so \[ P^{c \times s}_{\theta} = \bigcap_{\gamma \in F} \gamma^b P^{c \times s}_{\theta(\gamma)} = \left( \bigcap_{\gamma \in F}\gamma^c R_{\sigma(\gamma)} \right) \times \left( \bigcap_{\gamma \in F} \gamma^s S_{\tau(\gamma)} \right) = R^c_{\sigma} \times S^s_{\tau}.\] Note that for any $j \leq r$, $R_j \times S^s_{\tau}$ is a finite disjoint union of sets of the form $R^c_{\sigma} \times S^s_{\tau}$, hence any $A \in \mathscr{L}$ is within $\frac{\epsilon}{16}$ of finite disjoint union of sets of the form $P^{c \times s}_{\theta}$ for $\theta: F \to r \times 2$.\\ \\ Let $\delta = \frac{\epsilon}{4 (2r)^{2|F|}}$. Fix an ergodic action $c$ of $\mathbb{F}_N$ such that $d_u\left(\gamma^a,\gamma^c \right) < \frac{\delta^2}{32 |F|^2 (2r)^{|F|^2}}$ for all $\gamma \in F$. (For example use the fact that the ergodic automorphisms are uniformly dense in $\mathrm{Aut}(X,\mu)$ to move one of the generators $\gamma$ of $\mathbb{F}_N$ so it acts ergodically but is still sufficiently close to $\gamma^a$). Then clearly $d_H(C_{n,k}(a),C_{n,k}(c)) < \frac{\epsilon}{2}$ for all $n \leq n_0$ and $k \leq k_0$. Let $b = c \times s$. Since $c$ is ergodic and $s$ is free and mixing, $b$ is free and ergodic. Thus it is sufficient to show $d_H(C_{n,k}(c),C_{n,k}(b)) < \frac{\epsilon}{2}$ for all $n \leq n_0$, $k \leq k_0$. Since $c \prec b$, it is sufficient to show that for every partition $\mathcal{A}$ of $X \times 2^{\mathbb{F}_N}$ there is a partition $\mathcal{C}$ of $X$ such that $\rho \left(M^\mathcal{A}_{n,k}(b),M^\mathcal{C}_{n,k}(c) \right) < \frac{\epsilon}{2}$. By our previous reasoning, for each partition $\mathcal{A} = (A_1,\ldots,A_k)$ of $X \times 2^{\mathbb{F}_N}$ there is a partition $\mathcal{B}$ whose pieces are disjoint unions of sets of the form $P^b_\theta$ for $\theta: F \to r \times 2$ such that $\rho \left( M^{\mathcal{A}}_{n,k}(b), M^{\mathcal{B}}_{n,k}(b) \right) < \frac{\epsilon}{4}$. \begin{claim} There is a partition $\mathcal{Q}$ of $X$ indexed by $r \times 2$ such that for every $\theta: J \to r \times 2$ with $J \subseteq F_0 F$ we have $|(\mu \times \nu)(P^b_{\theta}) - \mu(Q^c_{\theta})| < \delta$. \end{claim} Suppose the claim holds. Regard $\mathbb{F}_N$ as acting on $\bigcup_{J \subseteq \mathbb{F}_N} \{\theta: J \to 2 \times r \}$ by shift, $\gamma \cdot \theta(\gamma') = \theta(\gamma^{-1}\gamma')$. Thus the domain $\mathrm{dom}(\gamma \cdot \theta) = \gamma \mathrm{dom}(\theta)$. Then for any $\theta,\kappa: F \to 2 \times r$ and $\gamma \in F_0$ we have \[ \gamma^b P^b_{\theta} \cap P^b_{\kappa} = \begin{cases} P^b_{\gamma \cdot \theta \cup \kappa} &\mbox{ if }\gamma \cdot \theta \mbox{ and } \kappa \mbox{ are compatible},\\ \emptyset &\mbox{ if not.} \end{cases} \] and similarly \[ \gamma^c Q^c_{\theta} \cap Q^c_{\kappa} = \begin{cases} Q^c_{\gamma \cdot \theta \cup \kappa} &\mbox{ if }\gamma \cdot \theta \mbox{ and } \kappa \mbox{ are compatible},\\ \emptyset &\mbox{ if not.} \end{cases} \] Therefore the claim gives $|(\mu \times \nu)(\gamma^b P^b_{\theta} \cap P^b_{\kappa}) - \mu(\gamma^c Q^c_{\theta} \cap Q^c_{\kappa})| < \delta$ for all $\theta,\kappa: F \to r \times 2$. So if $\mathcal{B} = \{B_1,\ldots,B_k\}$ is a partition such that $B_i = \bigsqcup_{s=1}^t P^b_{\theta_i(s)}$ for functions $\theta_i(s):F \to r \times 2$ and we let $C_i = \bigsqcup_{s=1}^t Q^c_{\theta_i(s)}$ then we have \begin{align*} |(\mu \times \nu)(\gamma^b B_i \cap B_j) - \mu(\gamma^c C_i \cap C_j)| &= \left \vert (\mu \times \nu) \left( \bigsqcup_{s,s' = 1}^t \gamma^b P^b_{\theta_i(s)} \cap P^b_{\theta_j(s')} \right) - \mu\left( \bigsqcup_{s,s' =1}^t \gamma^c Q^c_{\theta_i(s)} \cap Q^c_{\theta_j(s')} \right) \right \vert \\ & \leq t^2 \delta \leq (2r)^{2|F|} \delta < \frac{\epsilon}{4}, \end{align*} since $t \leq (2r)^{|F|}$. Taking $\mathcal{C} = (C_i)_{i=1}^k$ we get $\rho \left( M^\mathcal{B}_{n,k}(b),M^\mathcal{C}_{n,k}(c) \right) < \frac{\epsilon}{4}$, which implies the theorem.\\ \\ It remains to show Claim . This part of the argument follows the proof of Theorem 1 in and the extensions of these ideas developed in . Let $G = F_0 F$. Assume without loss of generality that $G$ is closed under taking inverses. Note that it suffices to prove the claim for $\theta$ defined on all of $G$. In order to find $\mathcal{Q}$ we will find a partition $\mathcal{T} = \{ T_1,T_2\}$ and set $Q_{i,j} = R_i \cap T_j$ for $1 \leq i \leq r$, $1 \leq j \leq 2$. Thus we are looking for $\mathcal{T} = \{T_1,T_2\}$ such that for all $(\tau,\sigma)$ with $\sigma:G \to r$ and $\tau: G \to 2$ we have \[ |(\mu \times \nu)(R^c_{\sigma} \times S^s_{\tau}) - \mu(R^c_{\sigma} \cap T^c_\tau)| < \delta. \] Note that $\nu(S^s_\tau) = 2^{-|G|}$ for any such $\tau$ so we are looking for $\mathcal{T}$ such that $\left \vert 2^{-|G|}\mu(R^c_{\sigma}) - \mu(R^c_{\sigma} \cap T^c_{\tau}) \right \vert < \delta$. The idea is that a random $\mathcal{T}$ should have this property.\\ \\ Without loss of generality we may assume $X$ is a compact metric space with a compatible metric $p$. For $\eta > 0$ let \[ D_\eta = \{x \in X: \mbox{ for all } \gamma, \gamma' \in G, \gamma_1 \neq \gamma_2 \mbox{ implies } p(\gamma_1^c x, \gamma_2^c x) > \eta \}\] and \[ E_\eta = \{(x,x') \in D_\eta^2: \mbox{ for all } \gamma_1,\gamma_2 \in G, p(\gamma_1^c x, \gamma_2^c x') > \eta\}. \] \begin{lemma} There is $\eta >0$ such that $\mu(D_\eta) > 1 - \frac{\delta^2}{16 (2r)^{|F|^2}}$ and $\mu^2(X^2 \setminus E_\eta) < \frac{\delta^2}{16 (2r)^{2|F|}}$. \end{lemma} \begin{proof} Clearly if $\eta_1 < \eta_2$ then $D_{\eta_2} \subseteq D_{\eta_1}$. We have $X \setminus \bigcup_{\eta > 0} D_\eta = \{ x \in X:$ for some $\gamma_1 \neq \gamma_2 \in G$, $\gamma_1^c x = \gamma_2^c x \}$. Now since $a$ is free, if $\gamma_1^c x = \gamma_2^c x$ then we must have $\gamma_i^c x \neq \gamma_i^a x$ for some $i \in \{1,2\}$. Each $\gamma \in G$ is a product $f_1 f_2$ for $f_1 \in F_0$ and $f_2 \in F$, thus for any $\gamma \in G$ we have \[ d_u \left( \gamma^c,\gamma^a \right) < d_u(f_1^a,f_1^c) + d_u(f_2^a,f_2^c) < \frac{\delta^2}{16 |F|^2 (2r)^{|F|^2}} \] since $f_i \in F$. Therefore \[ \mu ( \{ x: \mbox{ for some }\gamma \in G, \gamma^c x \neq \gamma^a x \}) < |G| \frac{\delta^2}{16 |F|^2 (2r)^{|F|^2}} < \frac{\delta^2}{16 (2r)^{2|F|}}.\] and hence $\mu \left( X \setminus \bigcup_{\eta > 0} D_\eta \right) < \frac{\delta^2}{16 (2r)^{|F|^2}}$. So we can find $\eta = \eta_0$ such that $D_{\eta_0}$ satisfies the lemma. Now for any $\eta > 0$, \begin{align*} D_{\eta_0}^2 \setminus \bigcup_{\eta > 0} E_\eta &= \{(x,x') \in D_{\eta_0}^2: \mbox{ for all } \eta > 0 \mbox{ there exist }\gamma_1,\gamma_2 \in G \mbox{ such that } p(\gamma_1 x,\gamma_2 x') < \eta \}\\ &= \{(x,x') \in D_{\eta_0}^2: \mbox{ there exist }\gamma_1,\gamma_2 \in G \mbox{ such that } \gamma_1 x = \gamma_2 x' \}. \end{align*} For a fixed $x$, $\{(x,x') \in D_{\eta_0}^2: \mbox{ there exist }\gamma_1,\gamma_2 \in G \mbox{ such that } \gamma_1 x = \gamma_2 x' \}$ is finite so $\mu \left( D_{\eta_0}^2 \setminus \bigcup_{\eta > 0} E_\eta \right)$ has measure $0$ by Fubini and hence we have the lemma for $E_\eta$. \end{proof} Let $\mathcal{Y} = \{Y_1,\ldots,Y_m\}$ be a partition of $X$ into pieces with diameter $< \frac{\eta}{4}$. For $x \in X$ let $Y(x)$ be the unique $l \leq m$ such that $x \in Y_i$. Let $\kappa$ be the uniform (= product) probability measure on $2^m$ and for each $\omega \in 2^m$ define a partition $Z(\omega) = \{Z^\omega_1,Z^\omega_2\}$ by letting $x \in Z^\omega_i$ if and only if $\omega(Y(x)) = i$. Thus we have a random variable $Z: (2^m,\kappa) \to \mathrm{MALG}(X,\mu)^2 $ given by $\omega \mapsto Z(\omega)$. Fix now $\tau:G \to 2$ and an arbitrary subset $A \subseteq X$. We compute the expected value of $\mu(Z(\omega)_\tau \cap A)$. Let $\chi_B$ be the characteristic function of $B$. \begin{align} \mathbb{E}[\mu(Z_\tau \cap A)] &= \int_{2^m} \mu(Z(\omega)_\tau \cap A) d \kappa(\omega) \nonumber \\ & = \int_{2^m} \int_X \chi_{Z(\omega)_\tau \cap A}(x) d \mu(x) d \kappa^m(\omega) \nonumber \\ &= \int_A \int_{2^m} \chi_{Z(\omega)_\tau}(x) d \kappa(\omega) d \mu(x) \nonumber \\ & = \int_{D_\eta \cap A} \int_{2^m} \chi_{Z(\omega)_\tau}(x) d \kappa(\omega) d \mu(x) + \int_{A \setminus D_\eta} \int_{2^m} \chi_{Z(\omega)_\tau}(x) d \kappa(\omega) d \mu(x) \end{align} Now if $x \in D_\eta$ then for all $\gamma_1 \neq \gamma_2 \in G$ we have $p(\gamma_1^c x, \gamma_2^c x) \geq \eta$ so that $Y(\gamma_1^c x) \neq Y(\gamma_2^c x)$ and hence the events $\omega(Y(\gamma_1^c x)) = i$ and $\omega(Y(\gamma_2^c x)) = j$ are independent. We have $x \in \gamma^c Z(\omega)_{\tau(\gamma)}$ if and only if $\omega(Y((\gamma^{-1})^c x)) = \tau(\gamma)$, so if $x \in D_\eta$ and $\gamma_1 \neq \gamma_2 \in G$ the events $x \in \gamma^c Z(\omega)_{\tau(\gamma_1)}$ and $x \in \gamma^c Z(\omega)_{\tau(\gamma_2)}$ are independent. So for $x \in D_\eta$, \begin{align} \int_{2^m} \chi_{Z(\omega)_\tau}(x) d \kappa(\omega) &= \kappa (\{\omega: x \in \gamma^c Z(\omega)_{\tau(\gamma)} \mbox{ for all } \gamma \in G \}) \nonumber \\ & = \prod_{\gamma \in G}\kappa \left ( \left\{ \omega: \omega(Y((\gamma^{-1})^cx)) = \tau(\gamma) \right\} \right) = 2^{-|G|} \end{align} Since $\mu(X \setminus D_\eta) < \frac{\delta^2}{16 (2r)^{|F|^2}}$, we have $2^{-|G|}\left(\mu(A) - \frac{\delta^2}{16 (2r)^{|F|^2}} \right) \leq (6) \leq 2^{-|G|}\mu(A)+ \frac{\delta^2}{16 (2r)^{|F|^2}}$ and thus $\left \vert \mathbb{E}[\mu(Z_\tau \cap A)] - \mu(A)2^{-|G|} \right \vert < \frac{\delta^2}{16 (2r)^{|F|^2}}.$ We now compute the second moment of $\mu(Z_\tau \cap A)$, in order to estimate its variance. \begin{align} \mathbb{E}\left[\mu(Z_\tau \cap A)^2 \right] & = \int_{2^m} \mu(Z_\tau(\omega) \cap A)^2 d \kappa (\omega) \nonumber \\ & = \int_{2^m} \left( \int_A \chi_{Z_\tau(\omega)}(x) d \mu(x) \right)^2 d \kappa(\omega) \nonumber \\ & = \int_{2^m} \int_{A^2} \chi_{Z_\tau(\omega)}(x_1) \chi_{Z_\tau(\omega)}(x_2) d \mu^2(x_1,x_2) d \kappa(\omega)\nonumber \\ & = \int_{A^2} \int_{2^m} \chi_{Z_\tau(\omega)}(x_1) \chi_{Z_{\tau}(\omega)}(x_2) d \kappa (\omega) d \mu^2(x_1,x_2) \nonumber \\ & = \int_{A^2 \cap E_\eta} \int_{2^m} \chi_{Z_\tau(\omega)}(x_1) \chi_{Z_{\tau}(\omega)}(x_2) d \kappa (\omega) d \mu^2(x_1,x_2) \nonumber \\ & \hspace{1 in}+ \int_{A^2 \setminus E_\eta} \int_{2^m} \chi_{Z_\tau(\omega)}(x_1) \chi_{Z_{\tau}(\omega)}(x_2) d \kappa (\omega) d \mu^2(x_1,x_2) \end{align} Now if $(x_1,x_2) \in E_\eta$ then for any pair $\gamma_1,\gamma_2 \in G$ we have $p(\gamma^c_1 x_1,\gamma^c_2 x_2) > \eta$ so that $Y(\gamma^c_1 x_1) \neq Y(\gamma^c_2 x_2)$ and thus for a fixed pair $(x_1,x_2)$ the events $\omega(Y(\gamma^{-1})^c x_1) = \tau(\gamma)$ for all $\gamma \in G$ and $\omega(Y(\gamma^{-1})^c x_2) = \tau(\gamma)$ for all $\gamma \in G$ are independent. Hence for a fixed $(x_1,x_2) \in E_\eta$ we have \begin{align*} \int_{2^m} \chi_{Z_\tau(\omega)}(x_1) \chi_{Z_\tau(\omega)}(x_2) d \kappa(\omega) & = \kappa(\{ \omega: x_1 \in \gamma^c Z(\omega)_{\tau(\gamma)} \mbox{ and } x_2 \in \gamma^c Z(\omega)_{\tau(\gamma)} \mbox{ for all } \gamma \in G \}) \\ & = \kappa(\{ \omega: \omega(Y((\gamma^{-1})^cx_1) = \tau(\gamma) \mbox{ and } \omega(Y((\gamma^{-1})^c) x_2) = \tau(\gamma) \mbox{ for all } \gamma \in G \}) \\ & =\kappa \left ( \left\{ \omega: \omega(Y((\gamma^{-1})^cx_1)) = \tau(\gamma) \mbox{ for all }\gamma \in G \right\} \right) \\ & \hspace{1 in} \cdot \kappa \left ( \left\{ \omega: \omega(Y((\gamma^{-1})^cx_2)) = \tau(\gamma) \mbox{ for all } \gamma \in G \right\} \right) \\ & = 2^{-2|G|} \end{align*} by $(7)$ and the fact that $E_\eta \subseteq D_\eta^2$. Since $\mu^2(A \setminus E_\eta) < \frac{\delta^2}{16 (2r)^{|F|^2}}$ we see $\left(\mu(A)^2 - \frac{\delta^2}{16 (2r)^{|F|^2}} \right) 2^{-2|G|} \leq (8) \leq 2^{-2|G|} \mu(A)^2 + \frac{\delta^2}{16 (2r)^{|F|^2}}$ and hence $\left \vert \mathbb{E}[\mu(Z_\tau \cap A)^2] - \mu(A)^2 2^{-2|G|} \right \vert < \frac{\delta^2}{16 (2r)^{|F|^2}}$. Therefore \begin{align*} \mathrm{Var}(\mu(Z_\tau \cap A)) &= \mathbb{E}[\mu(Z_\tau \cap A)^2] - \mathbb{E}[\mu(Z_\tau \cap A)]^2 \\ & \leq \left \vert \mathbb{E}[\mu(Z_\tau \cap A)^2] - \mu(A)^2 2^{-2|G|} \right \vert + \mu(A)^2 2^{-2|G|} - \left ( - \left \vert \mathbb{E}[\mu(Z_\tau \cap A)] - \mu(A)2^{-|G|} \right \vert + \mu(A)2^{-|G|} \right)^2 \\ & \leq \frac{\delta^2}{16(2r)^{|F|^2}} + \mu(A)^2 2^{-2|G|} - \left( - \frac{\delta^2}{16(2r)^{|F|^2}} + \mu(A)2^{-|G|} \right) ^2 \\ & = \frac{\delta^2}{16(2r)^{|F|^2}} - \frac{\delta^4}{(16(2r)^{|F|^2})^2} + 2 \mu(A)2^{-|G|} \frac{\delta^2}{16 ( 2r)^{|F|^2}} \leq \frac{\delta^2}{8 (2r)^{|F|^2}} . \end{align*} Therefore Chebyshev's inequality for $\mu(Z_\tau \cap A)$ gives \begin{align*}\kappa \left( \left\{ \omega: |\mu(Z_\tau(\omega) \cap A) - \mathbb{E}[\mu(Z_\tau \cap A)] | \geq \frac{\delta}{2} \right \} \right) &\leq \frac{\mathrm{Var}(\mu(Z_\tau \cap A))}{\left(\frac{\delta}{2}\right)^2} \\ & \leq \frac{1}{2 (2r)^{|F|^2} } \end{align*} Now since $\left \vert \mathbb{E}[\mu(Z_\tau \cap A)] - \mu(A)2^{-|G|} \right \vert < \frac{\delta}{2}$ we have \[\kappa \left( \left\{ \omega: \left \vert \mu(Z_\tau(\omega) \cap A) - \mu(A)2^{-|G|} \right \vert \geq \delta \right \} \right) \leq \frac{1}{2 (2r)^{|F|^2} }.\] Since this is true for each $\tau \in 2^G$ we have \[\kappa \left( \left\{ \omega:\left \vert \mu(Z_\tau(\omega) \cap A) - \mu(A)2^{-|G|}\right \vert| \geq \delta \mbox{ for some } \tau: G \to 2 \right \} \right ) \leq \frac{1}{2 r^{|F|^2} }.\] Finally, letting $A$ range over the sets $R_\sigma$ for $\sigma \in r^G$ we get \[\kappa \left ( \left\{ \omega: \left \vert \mu(Z_\tau(\omega) \cap R^c_\sigma) - \mu(R^c_\sigma)2^{-|G|} \right \vert \geq \delta \mbox{ for some } \tau: G \to 2 \mbox{ and } \sigma: G \to r \right \} \right) \leq \frac{1}{2} .\] Then any member of the nonempty complement of \[ \left\{ \omega: \left \vert \mu(Z_\tau(\omega) \cap R^c_\sigma) - \mu(R^c_\sigma)2^{-|G|} \right \vert \geq \delta \mbox{ for some } \tau: G \to 2 \mbox{ and } \sigma: G \to r \right \} \] works as $\mathcal{T}$. This completes the proof of Theorem . \end{proof} We note that the proof of Theorem goes through for any group $\Gamma$ such that an arbitrary free action can be approximated in the uniform topology by ergodic actions - for example any group of the form $\mathbb{Z} * H$. Such an approximation is impossible if $\Gamma$ has property $\mathrm{(T)}$, and in this case the extreme points of $\mathrm{FR}_{\sim_s}(\Gamma,X,\mu)$ are closed. Therefore the following question is natural. \begin{question} Let $\Gamma$ be a group without property $\mathrm{(T)}$. Can every free action of $\Gamma$ be approximated in the uniform topology of $A(\Gamma,X,\mu)$ by ergodic actions? \end{question} \section{The space of stable weak equivalence classes.} $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ can be given the structure of a weak convex space in exactly the same way as $\mathrm{A}_\sim(\Gamma,X,\mu)$. Moreover, it is clear that for any $a \in \mathrm{A}(\Gamma,X,\mu)$ and $t \in [0,1]$ we have $a \sim_s ta + (1-t)a$, so $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ is in fact a convex space. Recall that the metric $d_s$ on $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ is defined by $d_s(a,b) = d(a \times \iota,b \times \iota)$ where $d$ is the metric on $\mathrm{A}_\sim(\Gamma,X,\mu)$. \begin{proposition} For any $a,b,c \in \mathrm{A}(\Gamma,X,\mu)$ and $t \in [0,1]$, we have $d_s(ta + (1-t)c,tb + (1-t)c) \leq td_s(a,b)$. \end{proposition} It is clear that $(ta + (1-t)c) \times \iota \sim t(a \times \iota) + (1-t)(c \times \iota)$, so it suffices to show the following. \begin{proposition} For any $a,b,c \in \mathrm{A}(\Gamma,X,\mu)$ and $t \in [0,1]$ we have $d(ta + (1-t)c,tb + (1-t)c) \leq t d(a,b)$. \end{proposition} \begin{proof} Fix $n,k$ and write $C(a) = C_{n,k}(a)$ in order to show that $d_H(C(ta + (1-t)c), C(tb + (1-t)c)) \leq t d_H(C(a),C(b))$. Fix $\epsilon > 0$. Let $\mathcal{P} = (P_i)_{i=1}^n$ be a partition of $ X_1 \sqcup X_2$ where $X_1$ and $X_2$ are disjoint copies of $X$. Let $P_i^l = P_i \cap X_l$ for $l \in \{1,2\}$. Find a partition $\mathcal{Q} = (Q_i)_{i=1}^n$ such that for $i,j \leq n$ and $p \leq k$ we have \[ | \mu(\gamma_p^a P^1_i \cap P^1_j) - \mu(\gamma_p^b Q_i \cap Q_j)| < d_H(C(a),C(b)) + \epsilon.\] Then if we take $Q_i' = Q_i \sqcup P^2_i$ for all $i,j \leq n$, \begin{align*} | (t\mu + (1-t) \mu)&( \gamma_p^{ta + (1-t)c} P_i \cap P_j) - (t \mu + (1-t) \mu)(\gamma_p^{tb + (1-t)c} Q_i' \cap Q_j')| \\ & = | t \mu( \gamma_p^a P^1_i \cap P^1_j) + (1-t) \mu( \gamma_p^c P^2_i \cap P^2_j) - t\mu( \gamma_p^b Q_i \cap Q_j) - (1-t) \mu(\gamma_p^c P^2_i \cap P^2_j) | \\ & = | t \mu( \gamma_p^a P^1_i \cap P^1_j) +t\mu( \gamma_p^b Q_i \cap Q_j)| \leq t (d_H(C(a),C(b)) + \epsilon). \end{align*} \end{proof} Theorem now follows from Proposition and Corollary 12 in . Tucker-Drob and Bowen have obtained the next result independently of the author. \begin{proposition} The extreme points of $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ are precisely those stable weak equivalence classes which contain an ergodic action. \end{proposition} \begin{proof} Suppose that $a$ is ergodic and we have $a \sim_s tb + (1-t)c$ for $t \in (0,1)$. Therefore $a \prec \iota \times (tb + (1-t)c) \sim t( b \times \iota) + (1-t)(c \times \iota)$. Since $a$ is ergodic, Theorem 3.11 in implies that $a \prec b$ and $a \prec c$. Suppose toward a contradiction that $b \nprec_s c$, so that for some $n,k$ we have $C_{n,k}(b) \nsubseteq \cch( C_{n,k}(c) )$. Fixing $n,k$ write $C(d)$ for $C_{n,k}(d)$. Let $\alpha = \sup_{x \in C(b)} p(x, \cch(C(c)) )$ where $p$ is the metric on $[0,1]^{n \times k \times k}$. Choose $x_0 \in C(b)$ so that $p(x_0, \cch(C(c))) = \alpha$. Choose $y_0 \in \cch(C(c))$ so that $p(x_0,y_0) = \alpha$. Consider the point $tx_0 + (1-t) y_0 \in \cch(C(tb + (1-t)c))$. It is easy to see that \[ p(tx + (1-t) z, ty + (1-t) z) \leq tp(y,z) \] for any $x,y,z$ so we have \begin{align*} p(t x_0 + (1-t) y_0, x_0) & = p(t x_0 + (1-t)y_0, tx_0 + (1-t) x_0) \\ & \leq (1-t) p(x_0,y_0) < \alpha \end{align*} since $0 < t$. Since $\alpha = \inf_{y \in \cch(C(c))} p(x_0,y)$ we see that $tx_0 + (1-t) y_0 \notin \cch(C(c))$ and hence $\cch(C(tb+(1-t)c) \nsubseteq \cch(C(c))$. Since for any two actions $d,e$ we have $d \prec_s e$ if and only if $\cch(C_{n,k}(d)) \subseteq \cch(C_{n,k}(e))$ for all $n,k$ this implies that $tb + (1-t)c \nprec_s c$. But $tb + (1-t)c \prec_s a \prec c$ by hypothesis, so we have a contradiction and we conclude $b \prec_s c$. A symmetric argument shows $c \prec_s b$, so $b \sim_s c$. Since $\mathrm{A}_{\sim_s}(\Gamma,X,\mu)$ obeys $(2)$ of Definition , we get that $a \sim_s b \sim_s c$. Therefore if a stable weak equivalence class contains an ergodic action, it is an extreme point of $A_{\sim_s}(\Gamma,X,\mu)$. On the other hand, an argument identical to the proof of Theorem shows that if the stable weak equivalence class of an action $a$ is an extreme point of $A_{\sim_s}(\Gamma,X,\mu)$ then if we write $a = \int_Z a_z d \eta(z)$ then there is an ergodic action $b$ such that $a_z \sim_s b$ for all $z \in Z$. Thus $a \sim_s b \times \iota \sim_s b$ and we see that $a$ is stably weakly equivalent to an ergodic action. \end{proof} \bibliographystyle{plain} \bibliography{bibliography} Department of Mathematics\\ California Institute of Technology\\ Pasadena CA, 91125\\ \texttt{pjburton@caltech.edu} |
1501.04080 | Title: Differentially Private Bayesian Optimization
Abstract: Bayesian optimization is a powerful tool for fine-tuning the hyper-parameters
of a wide variety of machine learning models. The success of machine learning
has led practitioners in diverse real-world settings to learn classifiers for
practical problems. As machine learning becomes commonplace, Bayesian
optimization becomes an attractive method for practitioners to automate the
process of classifier hyper-parameter tuning. A key observation is that the
data used for tuning models in these settings is often sensitive. Certain data
such as genetic predisposition, personal email statistics, and car accident
history, if not properly private, may be at risk of being inferred from
Bayesian optimization outputs. To address this, we introduce methods for
releasing the best hyper-parameters and classifier accuracy privately.
Leveraging the strong theoretical guarantees of differential privacy and known
Bayesian optimization convergence bounds, we prove that under a GP assumption
these private quantities are also near-optimal. Finally, even if this
assumption is not satisfied, we can use different smoothness guarantees to
protect privacy.
Body: In general, our aim will be to protect the privacy of a validation dataset of sensitive records $\Vc \subseteq \Xc$ (where $\Xc$ is the collection of all possible records) when the results of Bayesian optimization depends on $\Vc$. \paragraph{Bayesian optimization.} Our goal is to maximize an unknown function $f_{\Vc}\colon \Lambda \rightarrow \mathbb{R}$ that depends on some validation dataset ${\Vc} \subseteq \Xc$: \begin{equation} \max_{\lambda \in \Lambda} f_{\Vc}(\lambda). \end{equation} It is important to point out that all of our results hold for the general setting of eq.~(), but throughout the paper, we use the vocabulary of a common application: that of machine learning hyper-parameter tuning. In this case $f_{\Vc}(\lambda)$ is the gain of a learning algorithm evaluated on validation dataset $\Vc$ that was trained with hyper-parameters $\lambda \in \Lambda \subseteq \Rbb^d$. As evaluating $f_{\Vc}$ is expensive (e.g., each evaluation requires training a learning algorithm), Bayesian optimization gives a procedure for selecting a small number of locations to sample $f_{\Vc}$: $[\lambda_1, \ldots, \lambda_T] \!=\! \blam_{T} \!\in\! \Rbb^{d\times T}$. Specifically, given a current sample $\lambda_t$, we observe a validation gain $v_t$ such that $v_t = f_{\Vc}(\lambda_t) + \alpha_t$, where $\alpha_t \sim {\cal N}(0,\sigma^2)$ is Gaussian noise with possibly non-zero variance $\sigma^2$. Then, given $v_t$ and previously observed values $v_1, \ldots, v_{t-1}$, Bayesian optimization updates its belief of $f_{\Vc}$ and samples a new hyper-parameter $\lambda_{t+1}$. Each step of the optimization proceeds in this way. To decide which hyper-parameter to sample next, Bayesian optimization places a prior distribution over $f_{\Vc}$ and updates it after every (possibly noisy) function observation. One popular prior distribution over functions is the Gaussian process ${\cal G}{\cal P}\bigl(\mu(\cdot),k(\cdot,\cdot)\bigr)$ \citep{rasmussen2006gaussian}, parameterized by a mean function $\mu(\cdot)$ (we set $\mu = 0$, w.l.o.g.) and a kernel covariance function $k(\cdot,\cdot)$. Functions drawn from a Gaussian process have the property that any finite set of values of the function are normally distributed. Additionally, given samples $\blam_{T} = [\lambda_1, \ldots, \lambda_T]$ and observations $\vb_T = [v_1, \ldots, v_T]$, the GP posterior mean and variance has a closed form: \begin{align} \mu_T(\lambda) =& \; k(\lambda,\blam_T)( \K_T + \sigma^2 \I )^{-1} \vb_T \nonumber \\ k_T(\lambda,\lambda') =&\; k(\lambda,\lambda') - k(\lambda,\blam_T)( \K_T + \sigma^2 \I )^{-1} k(\blam_T,\lambda') \nonumber \\ \sigma^2_T(\lambda) =& \; k_T(\lambda,\lambda), \end{align} where $k(\lambda,\blam_T) \!\in\! \Rbb^{1 \times T}$ is evaluated element-wise on each of the $T$ columns of $\blam_T$. As well, $\K_T = k(\X_T,\X_T) \!\in\! \Rbb^{T \times T}$ and $\lambda \in \Lambda$ is any hyper-parameter. As more samples are observed, the posterior mean function $\mu_T(\lambda)$ approaches $f_{\Vc}(\lambda)$. One well-known method to select hyper-parameters $\lambda$ maximizes the \emph{upper-confidence bound} (UCB) of the posterior GP model of $f_{\Vc}$ \citep{auer2002finite,srinivas2009gaussian}: \begin{align} \lambda_{t+1} \triangleq \argmax_{\lambda \in \Lambda} \mu_t(\lambda) + \sqrt{\beta_{t+1}}\sigma_t(\lambda), \end{align} where $\beta_{T+1}$ is a parameter that trades off the \emph{exploitation} of maximizing $\mu_t(\lambda)$ and the \emph{exploration} of maximizing $\sigma_t(\lambda)$. \citet{srinivas2009gaussian} proved that given certain assumptions on $f_{\Vc}$ and fixed, non-zero observation noise $(\sigma^2 \!>\! 0)$, selecting hyper-parameters $\lambda$ to maximize eq.~() is a no-regret Bayesian optimization procedure: $\lim_{T \to \infty} \frac{1}{T} \sum_{t = 1}^T f_{\Vc}(\lambda^*) - f_{\Vc}(\lambda_t) = 0$, where $f_{\Vc}(\lambda^*)$ is the maximizer of eq.~(). For the no-noise setting, \citet{de2012exponential} give a UCB-based no-regret algorithm. \paragraph{Contributions.} Alongside maximizing $f_{\Vc}$, we would like to guarantee that if $f_{\Vc}$ depends on (sensitive) validation data, we can release information about $f_{\Vc}$ so that the data ${\Vc}$ remains private. Specifically, we may wish to release (a) our best guess $\hat{\lambda} \triangleq \argmax_{t \leq T} f_{\Vc}(\lambda_t)$ of the true (unknown) maximizer $\lambda^*$ and (b) our best guess $f_{\Vc}(\hat{\lambda})$ of the true (also unknown) maximum objective $f_{\Vc}(\lambda^*)$. The primary question this work aims to answer is: how can we release private versions of $\hat{\lambda}$ and $f_{\Vc}(\hat{\lambda})$ that are close to their true values, or better, the values $\lambda^*$ and $f_{\Vc}(\lambda^*)$? We give two answers to these questions. The first will make a Gaussian process assumption on $f_{\Vc}$, which we describe immediately below. The second, described in Section , will utilize Lipschitz and convexity assumptions to guarantee privacy in the event the GP assumption does not hold. \paragraph{Setting.} For our first answer to this question, let us define a Gaussian process over hyper-parameters $\lambda,\lambda' \in \Lambda$ \emph{and} datasets $\Vc, \Vc' \!\subseteq\! \Xc$ as follows: $\GP\bigl(0, k_1(\Vc,\Vc') \otimes k_2(\lambda,\lambda')\bigr)$. A prior of this form is known as a multi-task Gaussian process \citep{bonilla2008multi}. Many choices for $k_1$ and $k_2$ are possible. The function $k_1(\Vc,\Vc')$ defines a set kernel (e.g., a function of the number of records that differ between $\Vc$ and $\Vc'$). For $k_2$, we focus on either the squared exponential: $k_2(\lambda,\lambda') = \exp\bigl(-\| \lambda - \lambda' \|_2^2 / (2 \ell^2)\bigr)$ or Mat\'{e}rn kernels: (e.g., for $\nu = 5/2$, $k_2(\lambda,\lambda') = (1 + \sqrt{5} r / \ell + (5 r^2) / (3\ell^2) ) \exp(- \sqrt{5} r /\ell )$, for $r = \| \lambda - \lambda' \|_2$), for a fixed $\ell$, as they have known bounds on the maximum information gain \citep{srinivas2009gaussian}. Note that as defined, the kernel $k_2$ is normalized (i.e., $k_2(\lambda,\lambda) = 1$). \begin{assumption} We have a problem of type (), where all possible dataset functions $[f_1, \ldots, f_{2^{|\Xc|}}]$ are GP distributed $\GP\bigl(0, k_1(\Vc,\Vc') \otimes k_2(\lambda,\lambda')\bigr)$ for known kernels $k_1, k_2$, for all $\Vc,\Vc' \!\subseteq\! \Xc$ and $\lambda,\lambda' \!\in\! \Lambda$, where $|\Lambda| \!\leq\! \infty$. \end{assumption} Similar Gaussian process assumptions have been made in previous work \citep{srinivas2009gaussian}. For a result in the no-noise observation setting, we will make use of the assumptions of \citet{de2012exponential} for our privacy guarantees, as described in Section . \subsection{Differential Privacy} One of the most widely accepted frameworks for private data release is \emph{differential privacy} \citep{dwork2006calibrating}, which has been shown to be robust to a variety of privacy attacks \citep{ganta2008composition,sweeney1997weaving,narayanan2008robust}. Given an algorithm $\Ac$ that outputs a value $\lambda$ when run on dataset $\Vc$, the goal of differential privacy is to `hide' the effect of a small change in $\Vc$ on the output of $\Ac$. Equivalently, an attacker should not be able to tell if a private record was swapped in $\Vc$ just by looking at the output of $\Ac$. If two datasets $\Vc,\Vc'$ differ by swapping a single element, we will refer to them as \emph{neighboring} datasets. Note that any non-trivial algorithm (i.e., an algorithm $\Ac$ that outputs different values on $\Vc$ and $\Vc'$ for some pair $\Vc,\Vc' \subseteq \Xc$) must include some amount of randomness to guarantee such a change in $\Vc$ is unobservable in the output $\lambda$ of $\Ac$ \citep{dwork2013algorithmic}. The level of privacy we wish to guarantee decides the amount of randomness we need to add to $\lambda$ (better privacy requires increased randomness). Formally, the definition of differential privacy is stated below. \begin{define} A randomized algorithm $\Ac$ is $(\epsilon,\delta)$-\textbf{differentially private} for $\epsilon,\delta \geq 0$ if for all $\lambda \in \emph{\mbox{Range}}(\Ac)$ and for all neighboring datasets $\Vc,\Vc'$ (i.e., such that $\Vc$ and $\Vc'$ differ by swapping one record) we have that \begin{align} \Prob\bigl[\Ac(\Vc) = \lambda\bigr] \leq e^{\epsilon} \Prob\bigl[ \Ac(\Vc') = \lambda\bigr] + \delta. \end{align} \end{define} The parameters $\epsilon, \delta$ guarantee how private $\Ac$ is; the smaller, the more private. The maximum privacy is $\epsilon = \delta = 0$ in which case eq.~() holds with equality. This can be seen by the fact that $\Vc$ and $\Vc'$ can be swapped in the definition, and thus the inequality holds in both directions. If $\delta=0$, we say the algorithm is simply $\epsilon$-differentially private. For a survey on differential privacy we refer the interested reader to \citet{dwork2013algorithmic}. There are two popular methods for making an algorithm $\epsilon$-differentially private: (a) the Laplace mechanism \citep{dwork2006calibrating}, in which we add random noise to $\lambda$ and (b) the exponential mechanism \citep{mcsherry2007mechanism}, which draws a random output $\tilde{\lambda}$ such that $\tilde{\lambda} \approx \lambda$. For each mechanism we must define an intermediate quantity called the \emph{global sensitivity} describing how much $\Ac$ changes when $\Vc$ changes. \begin{define} (Laplace mechanism) The \textbf{global sensitivity} of an algorithm $\Ac$ over all neighboring datasets $\Vc,\Vc'$ (i.e., $\Vc,\Vc'$ differ by swapping one record) is \begin{align} \Delta_{\Ac} \triangleq \max_{\Vc,\Vc' \subseteq \Xc } \| \Ac(\Vc) - \Ac(\Vc') \|_1. \nonumber \end{align} (Exponential mechanism) The \textbf{global sensitivity} of a function $q\colon \!\Xc \!\times\! \Lambda \!\rightarrow\! \mathbb{R}$ over all neighboring datasets $\Vc,\Vc'$ is \begin{equation} \Delta_q \triangleq \max_{\substack{\Vc,\Vc' \subseteq \Xc \\ \lambda \in \Lambda}} \| q(\Vc,\lambda) - q(\Vc',\lambda) \|_1. \nonumber \end{equation} \end{define} The Laplace mechanism hides the output of $\Ac$ by perturbing its output with some amount of random noise. \begin{define} Given a dataset $\Vc$ and an algorithm $\Ac$, the \textbf{Laplace mechanism} returns $\Ac(\Vc) + \omega$, where $\omega$ is a noise variable drawn from $\Lap(\Delta_{\Ac} / \epsilon)$, the Laplace distribution with scale parameter $\Delta_{\Ac} / \epsilon$ (and location parameter $0$). \end{define} The exponential mechanism draws a slightly different $\tilde{\lambda}$ that is `close' to $\lambda$, the output of $\Ac$. \begin{define} Given a dataset $\Vc$ and an algorithm $\Ac(\Vc) \!=\! \argmax_{\lambda \in \Lambda} q(\Vc,\lambda)$, the \textbf{exponential mechanism} returns $\tilde{\lambda}$, where $\tilde{\lambda}$ is drawn from the distribution $\frac{1}{Z} \exp\bigl( \epsilon q(\Vc,\lambda) / (2 \Delta_q )\bigr)$, and $Z$ is a normalizing constant. \end{define} Given $\Lambda$, a possible set of hyper-parameters, we derive methods for privately releasing the best hyper-parameters and the best function values $f_{\Vc}$, approximately solving eq.~(). We first address the setting with observation noise $(\sigma^2 \!>\! 0)$ in eq.~() and then describe small modifications for the no-noise setting. For each setting we use the UCB sampling technique in eq.~() to derive our private results. We have introduced methods for privately releasing the best hyper-parameters and validation accuracies in the case of exact and noisy observations. Our work makes use of the differential privacy framework, which has become commonplace in private machine learning \citep{dwork2013algorithmic}. We believe we are the first to demonstrate differentially private quantities in the setting of global optimization of expensive (possibly nonconvex) functions, through the lens of Bayesian optimization. One key future direction is to design techniques to release each sampled hyper-parameter and validation accuracy privately (during the run of Bayesian optimization). This requires analyzing how the maximum upper-confidence bound changes as the validation dataset changes. Another interesting direction is extending our guarantees in Sections and to other acquisition functions. For the case of machine learning hyper-parameter tuning our results are designed to guarantee privacy of the validation set only (it is equivalent to guarantee that the training set is never allowed to change). To simultaneously protect the privacy of the training set it may be possible to use techniques similar to the training stability results of \citet{chaudhuri2013stability}. Training stability could be guaranteed, for example, by assuming an additional training set kernel that bounds the effect of altering the training set on $f$. We leave developing these guarantees for future work. As practitioners begin to use Bayesian optimization in practical settings involving sensitive data, it suddenly becomes crucial to consider how to preserve data privacy while reporting accurate Bayesian optimization results. This work presents methods to achieve such privacy, which we hope will be useful to practitioners and theorists alike. In hyper-parameter tuning it may be reasonable to assume that we can observe function evaluations exactly: $v_{\Vc,t} = f_{\Vc}(\lambda_t)$. First note that we can use the same algorithm to report the maximum $\lambda$ in the no-noise setting. Theorems and still hold (note that $q = 0$ in Theorem ). However, we cannot readily report a private maximum $f$ as the information gain $\gamma_T$ in Theorems and approaches infinity as $\sigma^2 \rightarrow 0$. Therefore, we extend results from the previous section to the exact observation case via the regret bounds of \citet{de2012exponential}. Algorithm demonstrates how to privatize the maximum $f$ in the exact observation case. \begin{algorithm} \caption{ Private Bayesian Opt. (noise free obs.)} \begin{algorithmic} \STATE \textbf{Input:} $\Vc$; $\Lambda \subseteq \mathbb{R}^d$; $T$; $(\epsilon,\delta)$; $A, \tau$; assumptions on $f_\Vc$ in \citet{de2012exponential} \STATE Run method of \citet{de2012exponential}, resulting in noise free observations: $f_{\Vc}(\lambda_1), \ldots, f_{\Vc}(\lambda_T)$ \STATE $c = 2\sqrt{\bigl(1 - k(\Vc,\Vc')\bigr)\log(2|\Lambda|/\delta)}$ \STATE Draw $\theta \sim \Lap\!\Big[ \frac{A}{\epsilon} e\!^{- \frac{ 2 \tau }{ (\log 2)^{d/4}}} + \frac{c}{\epsilon} \Big]$ \STATE \textbf{Return:} $\tilde{f} = \max_{2 \leq t \leq T} f_{\Vc}(\lambda_t) + \theta$ \end{algorithmic} \end{algorithm} \subsection{Private near-maximum validation gain} We demonstrate that releasing $\tilde{f}$ in Algorithm is private (Theorem ) and that a small amount of noise is added to make $\tilde{f}$ private (Theorem ). To do so, we derive the global sensitivity of $\max_{2 \leq t \leq T} f(\lambda_t)$ in Algorithm independent of the maximum information gain $\gamma_T$ via \citet{de2012exponential}. Then we prove releasing $\tilde{f}$ is $(\epsilon,\delta)$-differentially private and that $\tilde{f}$ is almost $\max_{2 \leq t \leq T} f(\lambda_t)$. \paragraph{Global sensitivity.} The following Theorem gives a bound on the global sensitivity of the maximum $f$. \begin{thm} Given Assumption and the assumptions in Theorem 2 of \citet{de2012exponential}, for neighboring datasets $\Vc,\Vc'$ we have the following global sensitivity bound (in the Laplace mechanism sense), \begin{align} | \max_{2 \leq t \leq T} f'(\lambda_t) - \max_{2 \leq t \leq T} f(\lambda_t) | \leq A e^{- \frac{ 2 \tau }{ (\log 2)^{d/4}}}+ c \nonumber \end{align} w.p.\ at least $1-\delta$ for $c \!=\! 2\sqrt{\bigl(1 \!-\! k(\Vc,\Vc')\bigr)\log(2 |\Lambda|/\delta)}$. \end{thm} We leave the proof to the supplementary material. Given this sensitivity, we may apply the Laplace mechanism to release $\tilde{f}$. \begin{coro} Let $\Ac(\Vc)$ denote Algorithm run on dataset $\Vc$. Given assumption and that $f$ satisfies the assumptions of \citet{de2012exponential}, $\tilde{f}$ is $(\epsilon,\delta)$-differentially private, with respect to any neighboring dataset $\Vc'$, \emph{i.e.,} \begin{align} \Prob\bigl[ \Ac(\Vc) = \tilde{f}\bigr] \leq e^\epsilon \Prob\bigl[ \Ac(\Vc') = \tilde{f} \bigr] + \delta. \nonumber \end{align} \end{coro} Even though we must add noise to the maximum $f$ we show that $\tilde{f}$ is still close to the optimal $f(\lambda^*)$. \begin{thm} Given the assumptions of Theorem , we have the utility guarantee for Algorithm : \begin{align} | \tilde{f} - f(\lambda^*) | \leq \Omega + a \Big( \textstyle{\frac{\Omega}{\epsilon} + \frac{c}{\epsilon}} \Big) \nonumber \end{align} w.p. at least $1 \!-\! (\delta + e^{-a})$ for $\Omega \!=\! Ae^{- \frac{ 2 \tau }{ (\log 2)^{d/4}}}$. \end{thm} Machine learning is increasingly used in application areas with sensitive data. For example, hospitals use machine learning to predict if a patient is likely to be readmitted soon~\citep{yu2013predicting}, webmail providers classify spam emails from non-spam~\citep{weinberger2009feature}, and insurance providers forecast the extent of bodily injury in car crashes~\citep{chong2005traffic}. In these scenarios data cannot be shared legally, but companies and hospitals may want to share hyper-parameters and validation accuracies through publications or other means. However, data-holders must be careful, as even a small amount of information can compromise privacy. Which hyper-parameter setting yields the highest accuracy can reveal sensitive information about individuals in the validation or training data set, reminiscent of reconstruction attacks described by \citet{dwork2013algorithmic} and \citet{dinur2003revealing}. For example, imagine updated hyper-parameters are released right after a prominent public figure is admitted to a hospital. If a hyper-parameter is known to correlate strongly with a particular disease the patient is suspected to have, an attacker could make a direct correlation between the hyper-parameter value and the individual. To prevent this sort of attack, we develop a set of algorithms that automatically fine-tune the hyper-parameters of a machine learning algorithm while provably preserving differential privacy~\citep{dwork2006calibrating}. Our approach leverages recent results on Bayesian optimization~\citep{snoek2012practical,hutter2011sequential,bergstra2012random,gardner2014bayesian}, training a Gaussian process (GP)~\citep{rasmussen2006gaussian} to accurately predict and maximize the validation gain of hyper-parameter settings. We show that the GP model in Bayesian optimization allows us to release noisy final hyper-parameter settings to protect against aforementioned privacy attacks, while only sacrificing a tiny, bounded amount of validation gain. Our privacy guarantees hold for releasing the best hyper-parameters and best validation gain. Specifically our contributions are as follows: \begin{itemize} \item We derive, to the best of our knowledge, the first framework for Bayesian optimization with provable differential privacy guarantees, \item We develop variations both with and without observation noise, and \item We show that even if our validation gain is not drawn from a Gaussian process, we can guarantee differential privacy under different smoothness assumptions. \end{itemize} We begin with background on Bayesian optimization and differential privacy we will use to prove our guarantees. Even if our our true validation score $f$ is not drawn from a Gaussian process (Assumption ), we can still guarantee differential privacy for releasing its value after Bayesian optimization $f^{\mbox{\scriptsize{BO}}} \!=\! \max_{t \leq T} f(\lambda_t)$. In this section we describe a different functional assumption on $f$ that also yields differentially private Bayesian optimization for the case of machine learning hyper-parameter tuning. Assume we have a (nonsensitive) training set $\Tc \!=\! \{(\x_i,y_i)\}_{i=1}^n$, which, given a hyperparameter $\lambda$ produces a model $\w(\lambda)$ from the following optimization, \begin{align} \w_\lambda = \argmin_{\w} \overbrace{\frac{\lambda}{2} \| \w \|^2_2 + \frac{1}{n} \sum_{i=1}^n \ell(\w, \x_i, y_i)}^{O_{\lambda}(\w)}, \end{align} The function $\ell$ is a training loss function (e.g., logistic loss, hinge loss). Given a (sensitive) validation set $\Vc = \{(\xbar_i,\ybar_i)\}_{i=1}^m \subseteq \Xc$ we would like to use Bayesian optimization to maximize a validation score $f_{\Vc}$. \begin{assumption} Our true validation score $f_{\Vc}$ is \begin{align} f_{\Vc}(\w(\lambda)) = -\frac{1}{m} \sum_{i=1}^m g(\w_\lambda, \xbar_i, \ybar_i), \nonumber \end{align} where $g(\cdot)$ is a validation loss function that is $L$-Lipschitz in $\w$ (e.g., ramp loss, normalized sigmoid \citep{huang2014ramp}). Additionally, the training model $\w_\lambda$ is the minimizer of eq.~() for a training loss $\ell(\cdot)$ that is $1$-Lipschitz in $\w$ and convex (e.g., logistic loss, hinge loss). \end{assumption} Algorithm describes a procedure for privately releasing the best validation accuracy $f^{\BO}$ given assumption . Different from previous algorithms, we may run Bayesian optimization in Algorithm with any acquisition function (e.g., expected improvement \citep{mockus1978application}, UCB) and privacy is still guaranteed. \begin{algorithm} \caption{ Private Bayesian Opt. (Lipschitz and convex) } \begin{algorithmic} \STATE \textbf{Input:} $\Tc$ size $n$; $\Vc$ size $m$; $\Lambda$; $\lambda_{\min}$; $\lambda_{\max}$; $\epsilon$; $T$; $L$; $d$ \STATE Run Bayesian optimization for $T$ timesteps, observing: $f_{\Vc}(\w_{\lambda_1}), \ldots, f_\Vc(\w_{\lambda_T})$ for $\{ \lambda_1, \ldots, \lambda_T \} = \Lambda_{T,\Vc} \subseteq \Lambda$ \STATE $f^\BO = \max_{t \leq T} f_\Vc(\w_{\lambda_t})$ \STATE $g^* = \max_{(\x,y) \in \Xc, \w \in \Rc^d } g(\w,\x,y)$ \STATE Draw $\theta \sim \Lap\Big[ \frac{1}{\epsilon} \min\{ \frac{g^*}{m}, \frac{L}{m\lambda_{\min} }\} + \frac{(\lambda_{\max} - \lambda_{\min})L }{\epsilon \lambda_{\max} \lambda_{\min}} \Big]$ \STATE \textbf{Return:} $\tilde{f}_L = f^\BO + \theta$ \end{algorithmic} \end{algorithm} Similar to Algorithms 1 and 2 we use the Laplace mechanism to mask the possible change in validation accuracy when $\Vc$ is swapped with a neighboring validation set $\Vc'$. Different from the work of \citet{chaudhuri2013stability} changing $\Vc$ to $\Vc'$ may also lead to Bayesian optimization searching different hyper-parameters, $\Lambda_{T,\Vc}$ vs. $\Lambda_{T,\Vc'}$. Therefore, we must bound the \emph{total global sensitivity} of $f$ with respect to $\Vc$ \emph{and $\lambda$}, \begin{define} The \textbf{total global sensitivity} of $f$ over all neighboring datasets $\Vc,\Vc'$ is \begin{align} \Delta_f \triangleq \max_{ \substack{\Vc,\Vc' \subseteq \Xc \\ \lambda, \lambda' \in \Lambda}} | f_{\Vc}(\w_\lambda) - f_{\Vc'}(\w_{\lambda'}) |. \nonumber \end{align} \end{define} In the following theorem we demonstrate that we can bound the change in $f$ for arbitrary $\lambda < \lambda'$. \begin{thm} Given assumption , for neighboring $\Vc,\Vc'$ and arbitrary $\lambda < \lambda'$ we have that, \begin{align} \textstyle | f_{\Vc}(\w_\lambda) \!-\! f_{\Vc'}(\w_{\lambda'}) | \leq \frac{(\lambda' \!-\! \lambda)L}{\lambda' \lambda} + \min \! \big\{ \frac{g^*}{m}, \frac{L}{m\lambda_{\min}} \big\} \nonumber \end{align} where $L$ is the Lipschitz constant of $f$, $g^* = \max_{(\x,y) \in \Xc, \w \in \Rc^d } g(\w,\x,y)$, and $m$ is the size of $\Vc$. \end{thm} \emph{Proof.} Applying the triangle inequality yields \begin{align} | f_{\Vc}(\w_\lambda) - f_{\Vc'}(\w_{\lambda'}) | \leq& \; | f_{\Vc}(\w_{\lambda}) - f_{\Vc}(\w_{\lambda'}) | \nonumber \\ +& \; | f_{\Vc}(\w_{\lambda'}) - f_{\Vc'}(\w_{\lambda'}) |. \nonumber \end{align} This second term is bounded by \citet{chaudhuri2013stability} in the proof of Theorem 4. The only difference is, as we are not adding random noise to $\w_{\lambda'}$ we have that $| f_{\Vc}(\w_{\lambda'}) - f_{\Vc'}(\w_{\lambda'}) | \leq \min \{ g^*/m, L/(m\lambda_{\min})$ \}. To bound the first term, let $O_\lambda(\w)$ be the value of the objective in eq.~() for a particular $\lambda$. Note that $O_\lambda(\w)$ and $O_{\lambda'}(\w)$ are $\lambda$ and $\lambda'$-strongly convex. Define \begin{align} \textstyle h(\w) = O_{\lambda'}(\w) - O_{\lambda}(\w) = \frac{\lambda' - \lambda}{2} \| \w \|^2_2. \end{align} Further, define the minimizers $\w_\lambda \!=\! \argmin_\w O_{\lambda}(\w)$ and ${\w}_{\lambda'} \!=\! \argmin_\w [O_{\lambda}(\w) + h(\w)]$. This implies that \begin{align} \nabla O_{\lambda}(\w_\lambda) = \nabla O_{\lambda}({\w}_{\lambda'}) + \nabla h({\w}_{\lambda}) = 0. \end{align} Given that $O_\lambda$ is $\lambda$-strongly convex \citep{shalev2007online}, and by the Cauchy-Schwartz inequality, \begin{align} \lambda & \; \|\w_{\lambda} \!-\! \w_{\lambda'}\|^2_2 \leq \Big[\nabla O_\lambda(\w_\lambda) - \nabla O_{\lambda}({\w}_{\lambda'})\Big]^\top\Big[\w_\lambda \!-\! {\w}_{\lambda'}\Big] \nonumber \\ \leq& \; \nabla h(\w_{\lambda'})^\top\Big[\w_{\lambda} \!-\! \w_{\lambda'}\Big] \leq \|\nabla h(\w_{\lambda'}) \|_2 \|\w_{\lambda} \!-\! \w_{\lambda'}\|_2. \nonumber \end{align} Rearranging, \begin{align} \frac{1}{\lambda} \|\nabla h(\w_{\lambda'}) \|_2 = \Big\| \frac{\lambda' \!-\! \lambda}{2} \nabla \|\w_{\lambda'}\|_2^2 \Big\|_2 \geq \| \w_{\lambda} \!-\! \w_{\lambda'} \|_2 \end{align} Now as $\w_{\lambda'}$ is the minimizer of $O_{\lambda'}$ we have, \begin{align} \textstyle \nabla \| \w_{\lambda'} \|_2^2 = \frac{2}{\lambda'} \big[ - \frac{1}{n} \sum_{i=1}^n \nabla \ell(\w_{\lambda'}, \x_i, y_i) \big]. \nonumber \end{align} Substituting this value of $\w_{\lambda'}$ into eq.~() and noting that we can pull the positive constant term $(\lambda' - \lambda)/2$ out of the norm and drop the negative sign in the norm gives us \begin{align} \textstyle \frac{1}{\lambda} \|\nabla h(\w_{\lambda'}) \|_2 \!=\! \frac{\lambda' \!-\! \lambda}{\lambda \lambda'} \Big\| \frac{1}{n} \sum_{i=1}^n \nabla \ell(\w_{\lambda'}, \x_i, y_i) \Big\|_2 \!=\! \frac{\lambda' \!-\! \lambda}{\lambda \lambda'}. \nonumber \end{align} The last equality follows from the fact that the loss $\ell$ is $1$-Lipschitz by Assumption and the triangle inequality. Thus, along with eq.~(), we have \begin{align} \textstyle \| \w_{\lambda} \!-\! \w_{\lambda'} \|_2 \leq \frac{1}{\lambda} \|\nabla h(\w_{\lambda'}) \|_2 \leq \frac{\lambda' - \lambda}{\lambda \lambda'}. \nonumber \end{align} Finally, as $f$ is $L$-Lipschitz in $\w$, \begin{align} \textstyle |f_{\Vc}(\w_\lambda) - f_{\Vc}(\w_{\lambda'})| \leq L \| \w_{\lambda} - \w_{\lambda'} \|_2 \leq L \frac{\lambda' - \lambda}{\lambda \lambda'} \nonumber \end{align} Combining the result of \citet{chaudhuri2013stability} with the above expression completes the proof. \hfill$\blacksquare$ Given a finite set of possible hyperparameters $\Lambda$, we would like to bound $| f_\Vc^* - f_{\Vc'}^* |$; the best validation score found when running Bayesian optimization on $\Vc$ vs. $\Vc'$. Note that, by Theorem 7, \begin{align} \textstyle | f_\Vc^* - f_{\Vc'}^* | \leq & \; \max_{\lambda,\lambda'} | f_{\Vc}(\w_\lambda) \!-\! f_{\Vc'}(\w_{\lambda'}) | \nonumber \\ \leq & \; \textstyle{\frac{(\lambda_{\max} \!-\! \lambda_{\min})L}{\lambda_{\max} \lambda_{\min}} + \min \! \Big\{ \frac{g^*}{m}, \frac{L}{m\lambda_{\min}}} \Big\}, \nonumber \end{align} as $(\lambda' \!-\! \lambda)/(\lambda' \lambda)$ is strictly increasing in $\lambda'$ strictly decreasing in $\lambda$. Given this sensitivity of $f^*$ we can use the Laplace mechanism to hide changes in the validation set as follows. \begin{coro} Let $\Ac(\Vc)$ denote Algorithm applied on dataset $\Vc$. Given assumption , $\tilde{f}_L$ is $\epsilon$-differentially private, i.e., $\Prob\bigl[ \Ac(\Vc) = \tilde{f}_L\bigr] \leq e^\epsilon \Prob\bigl[ \Ac(\Vc') = \tilde{f}_L\bigr]$ \end{coro} We leave the proof to the supplementary material. Further, by the exponential tails of the Laplace mechanism we have the following utility guarantee, \begin{thm} Given the assumptions of Theorem , we have the following utility guarantee for $\tilde{f}_L$ w.r.t. $f^\BO$, \begin{align} \textstyle | \tilde{f}_L - f^\BO | \leq a\Big[ \frac{1}{\epsilon m} \min\{ g^*, \frac{L}{\lambda_{\min}} \} + \frac{(\lambda_{\max} - \lambda_{\min}) L }{ \epsilon \lambda_{\max} \lambda_{\min} } \Big] \nonumber \end{align} with probability at least $1 - e^{-a}$. \end{thm} \emph{Proof.} This follows exactly from the tail bound on Laplace random variables, given in the beginning of the proof of Theorem 6. \hfill$\blacksquare$ \section{With observation noise} In general cases of Bayesian optimization, observation noise occurs in a variety of real-world modeling settings such as sensor measurement prediction \citep{krause2008near}. In hyper-parameter tuning, noise in the validation gain may be as a result of noisy validation or training features. \begin{algorithm} \caption{ Private Bayesian Opt. (noisy observations)} \begin{algorithmic} \STATE \textbf{Input:} $\Vc$; $\Lambda \subseteq \mathbb{R}^d$; $T$; $(\epsilon,\delta)$; $\sigma^2_{\Vc,0}$; $\gamma_T$ \STATE $\mu_{\Vc,0} = 0$ \FOR{ $t = 1 \ldots T$ } \STATE $\beta_{t} \!=\! 2\log( |\Lambda| t^2 \pi^2 / (3 \delta) )$ \STATE $\lambda_{t} \triangleq \argmax_{\lambda \in \Lambda} \mu_{\Vc, t-1}(\lambda) + \sqrt{\beta_{t}}\sigma_{\Vc, t-1}(\lambda)$ \STATE Observe validation gain $v_{\Vc,t}$, given $\lambda_t$ \STATE Update $\mu_{\Vc, t}$ and $\sigma^2_{\Vc, t}$ according to \eqref{eq:post_GP} \ENDFOR \STATE $c \!=\! 2 \sqrt{\bigl(1 \!-\! k(\Vc,\Vc')\bigr) \log \big(3 |\Lambda|/\delta\big) }$ \STATE $q \!=\! \sigma \sqrt{4 \log(3/\delta)}$ \STATE $C_1 = 8/\log(1 + \sigma^{-2})$ \STATE Draw $\tilde{\lambda} \in \Lambda$ w.p. $\Prob[\lambda] \propto \mbox{exp}\Big(\frac{\epsilon \mu_{\Vc,T}(\lambda)}{2(2 \sqrt{\beta_{T+1}} + c)}\Big)$ \STATE $v^* \!=\! \max_{t \leq T} v_{\Vc,t}$ \STATE Draw $\theta\sim\! \Lap\!\Big[ \frac{\sqrt{C_1 \beta_T \gamma_T}}{\epsilon\sqrt{T}} + \frac{c}{\epsilon} + \frac{q}{\epsilon} \Big]$ \STATE $\tilde{v}=v^*+\theta$ \STATE \textbf{Return:} $\tilde{\lambda},\tilde{v}$ \end{algorithmic} \end{algorithm} In the sections that follow, although the quantities $f, \mu, \sigma, v$ all depend on the validation dataset $\Vc$, for notational simplicity we will occasionally omit the subscript $\Vc$. Similarly, for $\Vc'$ we will often write: $f', \mu', {\sigma'}^{2}, v'$. \subsection{Private near-maximum hyper-parameters} In this section we guarantee that releasing $\tilde{\lambda}$ in Algorithm is private (Theorem ) and that it is near-optimal (Theorem ). Our proof strategy is as follows: we will first demonstrate the global sensitivity of $\mu_T(\lambda)$ with probability at least $1\!-\!\delta$. Then we will show show that releasing $\tilde{\lambda}$ via the exponential mechanism is $(\epsilon,\delta)$-differentially private. Finally, we prove that $\mu_T(\tilde{\lambda})$ is close to $f(\lambda^*)$, the true maximizer of eq.~(). \paragraph{Global sensitivity.} As a first step we bound the global sensitivity of $\mu_T(\lambda)$ as follows: \begin{thm} Given Assumption , for any two neighboring datasets $\Vc,\Vc'$ and for all $\lambda \!\in\! \Lambda$ with probability at least $1- \delta$ there is an upper bound on the global sensitivity (in the exponential mechanism sense) of $\mu_T$: \begin{equation} | \mu'_T(\lambda) - \mu_T(\lambda) | \leq 2\sqrt{\beta_{T+1}} + \sigma_1\sqrt{2\log\left(3 |\Lambda| / \delta\right)}, \nonumber \end{equation} for $\sigma_1\!=\!\sqrt{2\bigl(1\!-\! k_1(\Vc,\Vc')\bigr)}$, $\beta_{t} \!=\! 2\log\Bigl( \! |\Lambda| t^2 \pi^2 /(3 \delta) \! \Bigr)$. \end{thm} \emph{Proof.} Note that, by applying the triangle inequality twice, for all $\lambda \in \Lambda$, \begin{align} &| \mu'_T(\lambda) - \mu_T(\lambda) | \leq \; | \mu'_T(\lambda) - f'(\lambda) | + | f'(\lambda) - \mu_T(\lambda) | \nonumber\\ &\leq | \mu'_T(\lambda) - f'(\lambda) | + | f'(\lambda) - f(\lambda) | + |f(\lambda)-\mu_T(\lambda)|.\nonumber \end{align} We can now bound each one of the terms in the summation on the right hand side (RHS) with probability at least $\frac{\delta}{3}$. According to \citet{srinivas2009gaussian}, Lemma 5.1, we obtain $| \mu'_T(\lambda) - f'(\lambda) |\!\leq\! \sqrt{\beta_{T+1}} \sigma'_T(\lambda)$. The same can be applied to $| f(\lambda)-\mu_T(\lambda) |$. As $\sigma'_T(\lambda)\!\leq\! 1$, because $k(\lambda,\lambda)\!=\!1$, we can upper bound both terms by $2\sqrt{\beta_{T+1}}$. In order to bound the remaining (middle) term on the RHS recall that for a random variable $Z \!\sim\! {\cal N}(0,1)$ we have: $\Prob\bigl[|Z| \!>\! \gamma\bigr]\! \leq\! e^{-\gamma^2/2}$. For variables $Z_1,\dots Z_n\!\sim\! {\cal N}(0,1)$, we have, by the union bound, that $\Prob\bigl[\forall i,\ |Z_i|\!\leq \!\gamma\bigr]\geq 1-ne^{-\gamma^2/2} \triangleq 1-\frac{\delta}{3}$. If we set $Z=\frac{|f(\lambda) - f'(\lambda)|}{\sigma_1}$ and $n\!=\!|\Lambda|$, we obtain $\gamma \!=\! \sqrt{2\log\bigl( 3 | \Lambda | / \delta\bigr)}$, which completes the proof. \hfill$\blacksquare$ We remark that all of the quantities in Theorem are either given or selected by the modeler (e.g, $\delta, T$). Given this upper bound we can apply the exponential mechanism to release $\tilde{\lambda}$ privately, as per Definition~: \begin{coro} Let $\Ac(\Vc)$ denote Algorithm applied on dataset $\Vc$. Given Assumption , $\tilde{\lambda}$ is $(\epsilon,\delta)$-differentially private, i.e., $\Prob\bigl[ \Ac(\Vc) \!=\! \tilde{\lambda}\bigr] \!\leq\! e^\epsilon \Prob\bigl[ \Ac(\Vc') \!=\! \tilde{\lambda} \bigr] \!+\! \delta$, for any pair of neighboring datasets $\Vc, \Vc'$. \end{coro} We leave the proof of Corollary to the supplementary material. Even though we must release a noisy hyper-parameter setting $\tilde{\lambda}$, it is in fact near-optimal. \begin{thm} Given Assumption the following near-optimal approximation guarantee for releasing $\tilde{\lambda}$ holds: \begin{align} \textstyle \mu_T(\tilde{\lambda}) \geq f(\lambda^*) - 2\sqrt{\beta_{T}} - q - \frac{2 \Delta}{\epsilon}( \log |\Lambda| + a ) \nonumber \end{align} w.p.\ $\geq 1 - (\delta + e^{-a})$, where $\Delta = 2\sqrt{\beta_{T+1}} + c$ (for $\beta_{T+1}$, $c$, and $q$ defined as in Algorithm ). \end{thm} \emph{Proof.} In general, the exponential mechanism selects $\tilde{\lambda}$ that is close to the maximum $\lambda$ \citep{mcsherry2007mechanism}: \begin{equation} \textstyle \mu_T(\tilde{\lambda}) \geq \max_{\lambda \in \Lambda} \mu_T(\lambda) - \frac{2 \Delta}{\epsilon}( \log |\Lambda| + a ), \end{equation} with probability at least $1 - e^{-a}$. Recall we assume that at each optimization step we observe noisy gain $v_t = f(\lambda_t) + \alpha_t$, where $\alpha_t \sim {\cal N}(0, \sigma^2)$ (with fixed noise variance $\sigma^2 \!>\! 0$). As such, we can lower bound the term $\max_{\lambda \in \Lambda} \mu_T(\lambda)$: \begin{align} \max_{\lambda \in \Lambda} \mu_T(\lambda) \geq& \; \underbrace{f(\lambda_T) + \alpha_T }_{v_T} \nonumber \\ f(\lambda^*) - \max_{\lambda \in \Lambda} \mu_T(\lambda) \leq& \; f(\lambda^*) - f(\lambda_T) + \alpha_T \nonumber \\ f(\lambda^*) - \max_{\lambda \in \Lambda} \mu_T(\lambda) \leq& \; 2\sqrt{\beta_T}\sigma_{T-1}(\lambda_T) + \alpha_T \nonumber \\ \max_{\lambda \in \Lambda} \mu_T(\lambda) \geq& \; f(\lambda^*) - 2\sqrt{\beta_T} + \alpha_T, \end{align} where the third line follows from \citet{srinivas2009gaussian}: Lemma 5.2 and the fourth line from the fact that $\sigma_{T-1}(\lambda_T) \leq 1$. As in the proof of Theorem , given a normal random variable $Z \!\sim\! {\cal N}(0,1)$ we have: $\Prob[ |Z| \!\leq\! \gamma] \!\geq\! 1 \!-\! e^{-\gamma^2/2} := 1 \!-\! \frac{\delta}{2}$. Therefore if we set $Z \!=\! \frac{\alpha_T}{\sigma}$ we have $\gamma \!=\! \sqrt{2 \log(2/\delta)}$. This implies that $|\alpha_T| \!\leq\! \sigma \sqrt{2 \log(2/\delta)} \!\leq\! \sqrt{4 \log(3/\delta)} \!=\! q$ (as defined in Algorithm ) with probability at least $1 \!-\! \frac{\delta}{2}$. Thus, we can lower bound $\alpha_T$ by $-q$. We can then lower bound $\max_{\lambda \in \Lambda} \mu_T(\lambda)$ in eq.~() with the right hand side of eq.~(). Therefore, given the $\beta_T$ in Algorithm , \citet{srinivas2009gaussian}, Lemma 5.2 holds with probability at least $1 - \frac{\delta}{2}$ and the theorem statement follows. \hfill$\blacksquare$ \subsection{Private near-maximum validation gain} In this section we demonstrate releasing the validation gain $\tilde{v}$ in Algorithm is private (Theorem ) and that the noise we add to ensure privacy is bounded with high probability (Theorem ). As in the previous section our approach will be to first derive the global sensitivity of the maximum $v$ found by Algorithm . Then we show releasing $\tilde{v}$ is $(\epsilon,\delta)$-differentially private via the Laplace mechanism. Perhaps surprisingly, we also show that $\tilde{v}$ is close to $f(\lambda^*)$. \paragraph{Global sensitivity.} We bound the global sensitivity of the maximum $v$ found with Bayesian optimization and UCB: \begin{thm} Given Assumption , and neighboring $\Vc,\Vc'$, we have the following global sensitivity bound (in the Laplace mechanism sense) for the maximum $v$, w.p. $\geq\! 1\!-\!\delta$ \begin{align} | \max_{t \leq T} v'_t - \max_{t \leq T} v_t | \leq \frac{\sqrt{C_1 \beta_T \gamma_T}}{\sqrt{T}} + c + q. \nonumber \end{align} where the maximum Gaussian process information gain $\gamma_T$ is bounded above for the squared exponential and Mat\'{e}rn kernels \citep{srinivas2009gaussian}. \end{thm} \emph{Proof.} For notational simplicity let us denote the regret term as $\Omega \triangleq \sqrt{C_1 T \beta_T \gamma_T}$. Then from Theorem 1 in \citet{srinivas2009gaussian} we have that \begin{equation} \frac{\Omega}{T} \geq f(\lambda^*) - \frac{1}{T} \sum_{t=1}^T f(\lambda_t) \geq f(\lambda^*) - \max_{t \leq T} f(\lambda_t). \end{equation} This implies $f(\lambda^*) \leq \max_{t \leq T} f(\lambda_t) + \frac{\Omega}{T}$ with probability at least $1 - \frac{\delta}{3}$ (with appropriate choice of $\beta_T$). Recall that in the proof of Theorem we showed that $|f(\lambda) - f'(\lambda)| \leq c$ with probability at least $1 - \frac{\delta}{3}$ (for $c$ given in Algorithm 1). This along with the above expression imply the following two sets of inequalities with probability greater than $1- \frac{2\delta}{3}$: \begin{align} \textstyle f'(\lambda^*) - c \leq f(\lambda^*) <& \; \max_{t \leq T} f(\lambda_t) + \textstyle{\frac{\Omega}{T}}; \nonumber \\ f(\lambda^*) - c \leq f'(\lambda^*) <& \; \max_{t \leq T} f'(\lambda_t) + \textstyle{\frac{\Omega}{T}}. \nonumber \end{align} These, in turn, imply the two sets of inequalities: \begin{align} \max_{t \leq T} f'(\lambda_t) \leq f'(\lambda^*) <& \; \max_{t \leq T} f(\lambda_t) + \textstyle{\frac{\Omega}{T}} + c; \nonumber \\ \max_{t \leq T} f(\lambda_t) \leq f(\lambda^*) <& \; \max_{t \leq T} f'(\lambda_t) + \textstyle{\frac{\Omega}{T}} +c. \nonumber \end{align} This implies $| \max_{t \leq T} f'(\lambda_t) - \max_{t \leq T} f(\lambda_t) | \leq \frac{\Omega}{T} + c$. That is, the global sensitivity of $\max_{t \leq T} f(\lambda_t)$ is bounded. Given the sensitivity of the maximum $f$, we can readily derive the sensitivity of maximum $v$. First note that we can use the triangle inequality to derive \begin{align} | \max_{t \leq T} v'(\lambda_t) - \max_{t \leq T} v(\lambda_t) | \leq& \; | \max_{t \leq T} v(\lambda_t) - \max_{t \leq T} f(\lambda_t) | \nonumber \\ +& \; | \max_{t \leq T} v'(\lambda_t) - \max_{t \leq T} f'(\lambda_t) | \nonumber \\ +& \; | \max_{t \leq T} f'(\lambda_t) - \max_{t \leq T} f(\lambda_t) |. \nonumber \end{align} We can immediately bound the final term on the right hand side. Note that as $v_t = f(\lambda_t) + \alpha_t$, the first two terms are bounded above by $|\alpha|$ and $|\alpha'|$, where $\alpha = \{\alpha_{\lceil t \rceil} \mid \lceil t \rceil \triangleq \argmax_{t \leq T} |\alpha_t|\}$ (similarly for $\alpha'$). This is because, in the worst case, the observation noise shifts the observed maximum $\max_{t \leq T} v_t$ up or down by $\alpha$. Therefore, let $\hat{\alpha} = \alpha$ if $|\alpha| > |\alpha'|$ and $\hat{\alpha} = \alpha'$ otherwise, so that we have: \begin{equation} | \max_{t \leq T} v'(\lambda_t) - \max_{t \leq T} v(\lambda_t) | \leq \textstyle{\frac{\Omega}{T}} + c + |2\hat{\alpha}|. \nonumber \end{equation} Although $|\hat{\alpha}|$ can be arbitrarily large, recall that for $Z \sim {\cal N}(0,1)$ we have: $\Prob[ |Z| \leq \gamma] \geq 1 - e^{-\gamma^2/2} \triangleq 1 - \frac{\delta}{3}$. Therefore if we set $Z = \frac{2\hat{\alpha}}{\sigma \sqrt{2}}$ we have $\gamma = \sqrt{2 \log(3/\delta)}$. This implies that $|2 \hat{\alpha}| \leq \sigma \sqrt{4 \log(3/\delta)} = q$ with probability at least $1 - \frac{\delta}{3}$. Therefore, if Theorem 1 from \citet{srinivas2009gaussian} and the bound on $| f(\lambda) - f'(\lambda)|$ hold together with probability at least $1- \frac{2\delta}{3}$ as described above, the theorem follows directly. \hfill$\blacksquare$ As in Theorem each quantity in the above bound is given in Algorithm ($\beta$, $c$, $q$), given in previous results \citep{srinivas2009gaussian} ($\gamma_T$, $C_1$) or specified by the modeler ($T$, $\delta$). Now that we have a bound on the sensitivity of the maximum $v$ we will use the Laplace mechanism to prove our privacy guarantee (proof in supplementary material): \begin{coro} Let $\Ac(\Vc)$ denote Algorithm run on dataset $\Vc$. Given Assumption , releasing $\tilde{v}$ is $(\epsilon,\delta)$-differentially private, i.e., $\Prob[ \Ac(\Vc) \!=\! \tilde{v}] \!\leq\! e^\epsilon \Prob[ \Ac(\Vc') \!=\! \tilde{v} ] \!+\! \delta.$ \end{coro} Further, as the Laplace distribution has exponential tails, the noise we add to obtain $\tilde{v}$ is not too large: \begin{thm} Given the assumptions of Theorem , we have the following bound, \begin{equation} \textstyle | \tilde{v} - f(\lambda^*)| \leq \sqrt{2 \log (2T / \delta)} + \frac{\Omega}{T} + a \Big( \frac{\Omega}{\epsilon T} + \frac{c}{\epsilon} + \frac{q}{\epsilon} \Big), \nonumber \end{equation} with probability at least $1- (\delta + e^{-a})$ for $\Omega = \sqrt{C_1 T \beta_T \gamma_T}$. \end{thm} \emph{Proof (Theorem ).} Let $Z$ be a Laplace random variable with scale parameter $b$ and location parameter $0$; $Z \sim \mathop{Lap}(b)$. Then $\Prob\bigl[|Z| \leq ab\bigr] = 1 - e^{-a}$. Thus, in Algorithm , $| \tilde{v} - \max_{t \leq T} v_t | \leq ab$ for $b = \frac{\Omega}{\epsilon T} + \frac{c}{\epsilon} + \frac{q}{\epsilon}$ with probability at least $1 - e^{-a}$. Further observe, \begin{align} \textstyle ab \geq & \; \max_{t \leq T} v_t - \tilde{v} \geq ( \max_{t \leq T} f(\lambda_t) - \max_{t \leq T} |\alpha_t| ) - \tilde{v} \nonumber \\ \geq & \; f(\lambda^*) - \textstyle{\frac{\Omega}{T}} - \max_{t \leq T} |\alpha_t| - \tilde{v} \end{align} where the second and third inequality follow from the proof of Theorem 3 (using the regret bound of \citet{srinivas2009gaussian}: Theorem 1). Note that the third inequality holds with probability greater than $1 - \frac{\delta}{2}$ (given $\beta_t$ in Algorithm ). The final inequality implies $f(\lambda^*) - \tilde{v} \leq \max_{t \leq T} |\alpha_t| + \frac{\Omega}{T} + ab$. Also note that, \begin{align} ab \geq & \; \tilde{v} - \max_{t \leq T} v_t \geq \tilde{v} - (\max_{t \leq T} f(\lambda_t) + \max_{t \leq T} |\alpha_t| ) \nonumber \\ \geq & \; \tilde{v} - f(\lambda^*) - \textstyle{\frac{\Omega}{T}} - \max_{t \leq T} |\alpha_t| \end{align} This implies that $f(\lambda^*) - \tilde{v} \geq - \max_{t \leq T} | \alpha_t | - \frac{\Omega}{T} - ab$. Thus we have that $| \tilde{v} - f(\lambda^*)| \leq \max_{t \leq T} |\alpha_t| + \frac{\Omega}{T} + ab$. Finally, because $|\alpha_t|$ could be arbitrarily large we give a high probability upper bound on $|\alpha_t|$ for all $t$. Recall that for $Z_1, \ldots Z_n \sim {\cal N}(0,1)$ we have by the tail probability bound and union bound that $\Prob\bigl[\forall t,\ |Z_t|\!\leq \!\gamma\bigr]\!\geq\! 1-ne^{-\gamma^2/2} \triangleq 1-\frac{\delta}{2}$. Therefore, if we set $Z_t = \alpha_t$ and $n = T$, we obtain $\gamma \!=\! \sqrt{2 \log (2T / \delta)}$. As defined $\gamma \geq \max_{t \leq T} |\alpha_t|$. \hfill$\blacksquare$ We note that, because releasing either $\tilde{\lambda}$ or $\tilde{v}$ is $(\epsilon,\delta)$-differentially private, by Corollaries and , releasing both private quantities in Algorithm 1 guarantees $(2\epsilon,2\delta)$-differential privacy for validation dataset $\Vc$. This is due to the composition properties of $(\epsilon,\delta)$-differential privacy \citep{dwork2006our} (in fact stronger composition results can be demonstrated, \citep{dwork2013algorithmic}). \twocolumn[ \icmltitle{Differentially Private Bayesian Optimization} \icmlauthor{Matt J. Kusner}{mkusner@wustl.edu} \icmlauthor{Jacob R. Gardner}{gardner.jake@wustl.edu} \icmlauthor{Roman Garnett}{garnett@wustl.edu} \icmlauthor{Kilian Q. Weinberger}{kilian@wustl.edu} \icmladdress{Computer Science \& Engineering, Washington University in St. Louis} \icmlkeywords{boring formatting information, machine learning, ICML} \vskip 0.3in ] \begin{abstract} Bayesian optimization is a powerful tool for fine-tuning the hyper-parameters of a wide variety of machine learning models. The success of machine learning has led practitioners in diverse real-world settings to learn classifiers for practical problems. As machine learning becomes commonplace, Bayesian optimization becomes an attractive method for practitioners to automate the process of classifier hyper-parameter tuning. A key observation is that the data used for tuning models in these settings is often sensitive. Certain data such as genetic predisposition, personal email statistics, and car accident history, if not properly private, may be at risk of being inferred from Bayesian optimization outputs. To address this, we introduce methods for releasing the best hyper-parameters and classifier accuracy privately. Leveraging the strong theoretical guarantees of differential privacy and known Bayesian optimization convergence bounds, we prove that under a GP assumption these private quantities are also near-optimal. Finally, even if this assumption is not satisfied, we can use different smoothness guarantees to protect privacy. \end{abstract} \section{Introduction} \input{intro.tex} \section{Background} \input{background.tex} \input{noise.tex} \section{Without observation noise} \input{exact.tex} We prove Corollary and Theorem in the supplementary material. We have demonstrated that in the noisy and noise-free settings we can release private near-optimal hyper-parameter settings $\tilde{\lambda}$ and function evaluations $\tilde{v}, \tilde{f}$. However, the analysis thus far assumes the hyper-parameter set is finite: $| \Lambda | \!<\! \infty$. It is possible to relax this assumption, using an analysis similar to \citep{srinivas2009gaussian}. We leave this analysis to the supplementary material. \section{Without the GP assumption} \input{no_gp.tex} \section{Results} \input{results.tex} \section{Related work} \input{related.tex} \section{Conclusion} \input{conclusion.tex} \bibliography{bibliography} \bibliographystyle{icml2015} There has been much work towards differentially private convex optimization \citep{chaudhuri2011differentially, kifer2012private, duchi2013local, song2013stochastic, jain2014near, bassily2014private}. The work of \citet{bassily2014private} established upper and lower bounds for the excess empirical risk of $\epsilon$ and $(\epsilon,\delta)$-differentially private algorithms for many settings including convex and strongly convex risk functions that may or may not be smooth. There is also related work towards private high-dimensional regression, where the dimensions outnumber the number of instances \citep{kifer2012private,thakurta2013differentially}. In such cases the Hessian becomes singular and so the loss is nonconvex. However, it is possible to use the restricted strong convexity of the loss in the regression case to guarantee privacy. Differential privacy has been shown to be achievable in online and interactive kernel learning settings \citep{jain2011differentially,thakurta2013nearly,jain2013differentially,mishraprivate}. In general, non-private online algorithms are closest in spirit to the methods of Bayesian optimization. However, all of the previous work in differentially private online learning represents a dataset as a sequence of bandit arm pulls (the equivalent notion in Bayesian optimization is function evaluations $f(\x_t)$). Instead, we consider functions in which changing a single dataset entry possibly affects \emph{all future function evaluations}. Closest to our work is that of \citet{chaudhuri2013stability}, who show that given a fixed set of hyper-parameters which are always evaluated for any validation set, they can return a private version of the index of the best hyper-parameter, as well as a private model trained with that hyper-parameter. Our setting is strictly more general in that, if the validation set changes, Bayesian optimization could search completely different hyper-parameters. Bayesian optimization, largely due to its principled handling of the exploration/exploitation trade-off of global, black-box function optimization, is quickly becoming the global optimization paradigm of choice. Alongside promising empirical results there is a wealth of recent work on convergence guarantees for Bayesian optimization, similar to those used in this work \citep{srinivas2009gaussian,de2012exponential}. \citet{vazquez2010convergence} and \citet{bull2011convergence} give regret bounds for optimizing the expected improvement acquisition function each optimization step. BayesGap \citep{hoffman2014correlation} gives a convergence guarantee for Bayesian optimization with budget constraints. Bayesian optimization has also been extended to multi-task optimization \citep{BaBrKeSe13,swersky2013multi}, the setting where multiple experiments can be run at once \citep{azimi2012hybrid,snoek2012practical}, and to constrained optimization \citep{gardner2014bayesian}. In this section we examine the validity of our multi-task Gaussian process assumption on $[f_1, \ldots, f_{2^{|\Xc|}}]$. Specifically, we search for the most likely value of the multi-task Gaussian process covariance element $k_1(\Vc,\Vc')$, for classifier hyper-parameter tuning. Larger values of $k_1(\Vc,\Vc')$ correspond to a smaller global sensitivity bounds in Theorems , , and leading to improved privacy guarantees. For our setting of hyper-parameter tuning, each $\lambda = [C, \gamma^2]$ are hyper-parameters for training a kernelized support vector machine (SVM) \citep{cortes1995support,scholkopf2001learning} with cost parameter $C$ and radial basis kernel width $\gamma^2$. The value $f_{\Vc}(\lambda)$ is the accuracy of the SVM model trained with hyper-parameters $\lambda$ on $\Vc$. To search for the most likely $k_1(\Vc,\Vc')$ we start by sampling $100$ different SVM hyper-parameter settings $\lambda_1,\ldots,\lambda_{100}$ from a Sobol sequence and train an SVM model for each on the Forest UCI dataset ($36,603$ training inputs). We then randomly sample $100$ i.i.d.\ validation sets $\Vc$. Here we describe the evaluation procedure for a fixed validation set size, which corresponds to a single curve in Figure (as such, to generate all results we repeat this procedure for each validation set size in the set $\{1000,2000,3000,5000,15000\}$). For each of the $100$ validation sets we randomly add or remove an input to form a neighboring dataset $\Vc'$. We then evaluate each of the trained SVM models on all $100$ datasets $\Vc$ and their pairs $\Vc'$. This results in two $100 \times 100$ (number of datasets, number of trained SVM models) function evaluation matrices $\mathbf{F}_{\Vc}$ and $\mathbf{F}_{\Vc'}$. Thus, $[\mathbf{F}_{\Vc}]_{ij}$ is the validation accuracy on the $i^{\scriptsize \mbox{th}}$ validation set $\Vc_i$ using the $j^{\scriptsize \mbox{th}}$ SVM model. The likelihood of function evaluations for a dataset pair $(\Vc_i,\Vc_i')$, for a value of $k_1(\Vc_i,\Vc_i')$, is given by the marginal likelihood of the multi-task Gaussian process: \begin{align} \Prob \begin{pmatrix} [\mathbf{F}_{\Vc}]_i \\ [\mathbf{F}_{\Vc'}]_i \\ \end{pmatrix} \sim {\cal N}( \mu_{100}, \sigma_{100}^2 ), \end{align} where $[\mathbf{F}_{\Vc}]_i = [f_{\Vc}(\lambda_1), \ldots, f_{\Vc}(\lambda_{100})]$ (similarly for $\Vc'$) and $\mu_{100}$ and $\sigma_{100}^2$ are the posterior mean and variance of the multi-task Gaussian process using kernel $k_1(\Vc,\Vc') \otimes k_2(\lambda,\lambda')$ after observing $[\mathbf{F}_{\Vc}]_i$ and $[\mathbf{F}_{\Vc'}]_i$ (for more details see \citet{bonilla2008multi}). As $\mu_{100}$ and $\sigma_{100}^2$ depend on $k_1(\Vc,\Vc')$, we treat it as a free-parameter and vary its value from $0.05$ to $0.95$ in increments of $0.05$. For each value, we compute the marginal likelihood () for all validation datasets ($\Vc_i$ for $i = 1,\ldots,100$). As each $\Vc_i$ is sampled i.i.d.\, the joint marginal likelihood is simply the product of all $\Vc_i$ likelihoods. Computing this joint marginal likelihood for each $k_1(\Vc,\Vc')$ value yields a single curve of Figure . As shown, the largest values of $k_1(\Vc,\Vc') = 0.95$ is most likely, meaning that $c$ in the global sensitivity bounds is quite small, leading to private values that are closer to their true optimums. |
1501.04081 | Title: Demonstration of Cathode Emittance Dominated High Bunch Charge Beams in
a DC gun-based Photoinjector
Abstract: We present the results of transverse emittance and longitudinal current
profile measurements of high bunch charge (greater than or equal to 100 pC)
beams produced in the DC gun-based Cornell Energy Recovery Linac Photoinjector.
In particular, we show that the cathode thermal and core beam emittances
dominate the final 95% and core emittance measured at 9-9.5 MeV. Additionally,
we demonstrate excellent agreement between optimized 3D space charge
simulations and measurement, and show that the quality of the transverse laser
distribution limits the optimal simulated and measured emittances. These
results, previously thought achievable only with RF guns, demonstrate that DC
gun based photoinjectors are capable of delivering beams with sufficient single
bunch charge and beam quality suitable for many current and next generation
accelerator projects such as Energy Recovery Linacs (ERLs) and Free Electron
Lasers (FELs).
Body: \title{Demonstration of Cathode Emittance Dominated High Bunch Charge Beams in a DC gun-based Photoinjector} \author{Colwyn Gulliford}\homepage{co-first author}\email{cg248@cornell.edu} \author{Adam Bartnik}\homepage{co-first author}\email{acb20@cornell.edu} \author{Ivan Bazarov} \author{Bruce Dunham} \author{Luca Cultrera} \affiliation{CLASSE, Cornell University} \date{\today} \begin{abstract} We present the results of transverse emittance and longitudinal current profile measurements of high bunch charge ($\geq$100 pC) beams produced in the DC gun-based Cornell Energy Recovery Linac Photoinjector. In particular, we show that the cathode thermal and core beam emittances dominate the final 95\ \end{abstract} \pacs{PACS numbers?} \maketitle Linear electron accelerators boast a wide range of current and planned applications in the physical sciences. Examples include: x-ray sources , electron-ion coolers , Ultra-fast Electron Diffraction (UED) experiments , and fixed-target nuclear physics experiments . A key feature of many of these applications is the potential to produce beams where the initial beam quality, set by the source, dominates the final beam quality at the usage point. This has lead to the design of a next generation of machines, such as high energy Energy Recovery Linacs (ERLs) , and Free Electron Lasers (FELs) which could provide diffraction limited hard x-rays with orders of magnitude brighter beams than modern storage rings. The successful design and implementation of such machines has the potential to impact an impressively broad range of research in physics, chemistry, biology, and engineering. For next generation high energy x-ray sources like the proposed Linac Coherent Light Source-II (LCLS-II) , the creation (at MHz repetition rates) and effective transport of multi-MeV beams with high bunch charges ($\geq$100 pC), picosecond bunch lengths, and sub-micron normalized transverse emittances represents a beam dynamics regime previously thought attainable only with RF gun based photoinjectors . In this letter, we challenge this assumption, and show that the DC gun-based Cornell ERL injector can produce cathode emittance dominated beams which meet the bunch charge, emittance, and peak current specifications of a next generation light source. In doing so, we also demonstrate excellent agreement between measurement and simulation of the injector, and show that ultimate optimization of the emittance in high-brightness photoinjectors may require advanced transverse laser shaping along with the use of low intrinsic emittance photocathodes. Before discussing our experimental results, we review the definitions of the key figures of merit for beam quality in high-brightness accelerators relevant for this work: emittance and brightness. For the beam densities encountered in this work ($10^{17}$-$10^{18}$ $\mathrm{e}/\mathrm{m}^3)$, classical relativistic Hamiltonian mechanics, with a self-interaction described by a space charge potential, sufficiently approximates the single bunch beam dynamics . In this model, the Hamiltonian for each bunch separates into a sum over $N=q/e$ Hamiltonians of the same form, reducing the 6N-D ensemble phase space volume conserved in Liouville's theorem to the 6D phase space of a single bunch. In the absence of coupling between each of the 2D canonical phase spaces ($x_i$, $P_{x_i}$), the conserved 6D phase space volume separates into three conserved 2D volumes. From a physics stand point, these three conserved quantities represent the most fundamental definition of 2D emittance. However, as a figure of merit, this definition fails to capture the effect of distortions of the phase space due to non-linear fields. This typically motivates the definition of the (normalized) rms emittance: $\epsilon_{n,x_i} = \frac{1}{mc}\sqrt{\langle x_i^2\rangle\langle p_{x_i}^2\rangle-\langle x_ip_{x_i}\rangle^2}$. Note the use of the mechanical momenta $p_{x_i}$. Under the above assumptions, conservation of this emittance follows directly from Liouville's theorem, provided the forces on the bunch are linear. In practice, bunches in high-brightness photoinjectors experience both non-linear and longitudinally correlated fields arising from space charge and time-dependent RF fields. Left unchecked, these fields lead to emittance growth along the beamline. The mitigation of these effects, known as emittance compensation , determines the degree to which the cathode emittance, given by: \begin{eqnarray} \epsilon_{n,x}=\frac{1}{mc} \sigma_{x,0}\sigma_{p_x,0} = \sigma_x\sqrt{\frac{\mathrm{MTE}}{mc^2}}, \end{eqnarray} dominates the beam quality downstream. Here $\sigma_{p_x,0}$ is the momentum variance intrinsic to the cathode material, which can be expressed in terms of the mean (kinetic) energry of the photoemitted electrons (MTE), and $\sigma_{x,0}$ is the spatial variance of the laser distribution. The rms emittance motivates a simple definition for the average transverse (normalized) brightness, defined generally as the particle flux per unit 4D transverse phase space volume : \begin{eqnarray} \overline{\mathcal{B}}_{n} = \frac{\bar{I}}{\epsilon_{n,x}\epsilon_{n,y}}, \end{eqnarray} where $\bar{I}$ is the average beam current. To characterize the contribution of the central core of the phase space to the emittance, as well as provide a pratical means to compare non-Gaussian beams, we define the emittance vs. fraction curve (see for details): for every area in phase space $a$, we find a bounding contour $D(a)$ which maximizes the enclosed fraction $f$ of beam particles. The rms emittance computed for the particles inside $D(a)$ defines the corresponding fractional emittance $\epsilon_{n,x}(f)$. From this we define the core emittance as \begin{eqnarray} \epsilon_{n,x}^{\mathrm{core}}=\left.d\epsilon/df\right|_{f=0} = \frac{1}{4\pi\rho_0}, \end{eqnarray} where $\rho_0$ is the peak in the phase space distribution function (typically the centroid). The corresponding brightness as a function of fraction and peak brightness follow directly from Eqns.~(-): \begin{eqnarray} \overline{\mathcal{B}}_{n}(f) = \frac{\bar{I}f^2}{\epsilon_{n,x}(f)\epsilon_{n,y}(f)},\hspace{0.5cm} \overline{\mathcal{B}}^{\mathrm{peak}}_{n}=\left.\overline{\mathcal{B}}_{n}\right|_{\mathrm{core}}. \end{eqnarray} In addition to defining the principle figures of merit for high-brightness accelerators, Eqns.~(-) make clear the importance of preserving the thermal and cathode core emittances: the degree of conservation sets the extent to which the intrinsic cathode MTE determines the downstream beam quality. For cathode emittance dominated beams, any improvement in the MTE translates into immediate improvement in the final beam quality. With this property in mind, we turn to the main purpose of this work: demonstrating that the Cornell ERL injector, a 5-15 MeV machine featuring a DC gun followed by a short SRF linac, can produce beams with a high degree of emittance preservation in the beam dynamics regime set by next generation light sources. Originally designed to create low emittance, moderate bunch charge ($\leq$77 pC) beams at high (1.3 GHz) repetition rate for a full hard x-ray ERL, the Cornell injector currently holds the world record for high average current from a photoinjector with cathode lifetimes suitable for an operating user facility , as well as the record for lowest demonstrated emittance from a DC gun-based photoinjector at bunch charges of 19 and 77 pC . As of this work, the Cornell injector remains largely the same as described in , with the most notable difference being the current operation of the DC gun at 395 kV. For the measurements in this work, we used the LCLS-II injector 95\ \begin{table}[htb] \caption{LCLS-II Injector Specifications} \begin{ruledtabular} \begin{tabular}{ c | ccc} Bunch Charge & 20 pC & 100 pC & 300 pC \\ \hline 95\ Peak Current & 5 A & 10 A & 30 A \end{tabular} \end{ruledtabular} \end{table} For all direct phase space and longitudinal profile data taken with our Emittance Measurement System (EMS), we exclusively used a 50 MHz laser to limit the beam power deposited into the interceptive EMS diagnostics. This laser system produces 520 nm, 1 ps rms pulses with comparable pulse energy to the primary 1.3 GHz laser used for full repetition rate experiments . Four rotatable birefringent crystals temporally shape the primary pulses by splitting each into 16 copies with tunable relative intensities set by the crystal rotation angles. For transverse shaping, we used a beam expander and pinhole to clip the Gaussian laser distribution at roughly the half maximum intensity (truncation fraction of 50\ All measurements were performed using a single NaKSb cathode with a $140\pm10$ meV MTE. In order to determine the injector settings that produce optimal emittances and peak currents, we ran Multi-Objective Genetic Algorithm (MOGA) optimizations using the 3D space charge code General Particle Tracer (GPT) . For each of the LCLS-II nominal charges, the optimizer simultaneously minimized both the emittance and rms bunch length at the location of the EMS in the simulated injector, subject to realistic constraints on all relevant injector and beam parameters. For a detailed description of our 3D injector model, refer to . Additionally, we provided the optimizer with a realistic simulation of the laser distribution, and allowed the optimizer to vary the transverse pinhole size and truncation fraction, as well as the rotation of the longitudinal shaping crystal angles. The resulting Pareto fronts (shown later) provided injector settings that simultaneously satisfied both the 95\ For each bunch charge we loaded the corresponding optimal settings into the injector and tuned the machine to produce the lowest emittances possible while still meeting the peak current targets. All critical machine parameters matched those chosen by the optimizer to within 5\ The thermal emittances were computed directly from the measured transverse laser profiles according to Eqn.~(). In order to characterize the initial temporal laser shape, we measured the longitudinal electron beam current profile at near zero charge (0.02 $\pm$ 0.01 pC) with all RF cavities off. Finally, we loaded the corresponding machine settings and measured laser distributions for each bunch charge into our virtual accelerator GUI , and ran 250k macro-particle GPT simulations for comparison with measurement. Fig.~ shows the measured laser distributions on a CCD camera located at the same distance from the clipping pinhole as the cathode. To match the optimization results as best as possible, we tuned the laser spot size on the laser CCD so that the edge truncation fraction was $50\ Fig.~ displays the measured horizontal and vertical projected phases spaces corresponding to the best measured emittances. Note the use of the normalized mechanical momenta $\gamma\beta_{x_i} = p_{x_i}/mc$. One striking feature seen in Fig.~ is the overall symmetry between the horizontal and vertical phase spaces. Fig.~ shows the comparison of the measured (red) and simulated longitudinal current profiles (blue). In addition to the excellent agreement seen between measurement and simulation, we note that all peak current targets were met. Table~ displays the thermal and core emittance at the cathode and the resulting measured 95\ We estimate a $\pm6$\ The random error between subsequent measurements using the EMS was typically $\leq$1\ \begin{table} \caption{ (a) Measured horizontal (vertical) thermal 95\ \subtable[\hspace{0.2cm}Horizontal (vertical) projected emittance data.]{ \begin{ruledtabular} \begin{tabular}{c | ccc} Charge & Thermal $\epsilon_{n}$ ($\mathrm{\mu m}$) & 95\ \hline 20 pC & 0.12 (0.11) & 0.18 (0.19) & 67 (58) \\ 100 pC & 0.24 (0.23) & 0.30 (0.32) & 80 (72) \\ 300 pC & 0.42 (0.41) & 0.62 (0.60) & 67 (68) \end{tabular} \end{ruledtabular} }\\ \subtable[\hspace{0.2cm}Horizontal (vertical) projected core emittance data.]{ \begin{ruledtabular} \begin{tabular}{c | ccc} Charge & Cathode $\epsilon_{n,\mathrm{core}}$ ($\mathrm{\mu m}$) & EMS $\epsilon_{n,\mathrm{core}}$ ($\mathrm{\mu m}$) & Ratio ($\ \hline 20 pC & 0.06 (0.06) & 0.09 (0.08) & 67 (75) \\ 100 pC & 0.14 (0.13) & 0.16 (0.16) & 85 (79) \\ 300 pC & 0.26 (0.24) & 0.30 (0.28) & 87 (87) \end{tabular} \end{ruledtabular} } \end{table} We note that this data quantitatively reflects the qualitative symmetry seen in the phase space measurements (Fig.~), and as well as satisfies all of the LCLS-II injector emittance targets. The table also shows the ratio of the thermal emittance and the final 95\ In order to determine the effect of the laser shape on the emittances, we ran a second round of optimizations using the measured transverse laser distributions in Fig.~(a) and the crystal angles used to create the flattop Fig.~(b). All other relevant injector parameters varied as before. Fig.~ shows the \emph{average 100\ For the 100 and 300 pC measurements, this produces roughly a 23\ In this work, we have shown that optimal injector settings found using MOGA optimizations of 3D space charge simulations of the Cornell ERL injector produce machine states that preserve both the measured 95\ \begin{acknowledgments} We acknowledge Jared Maxson for his useful discussions and interest in this work. This work was supported, in part, by the LCLS-II Project and the US Department of Energy, Contract DE-AC02-76SF00515. \end{acknowledgments} \begin{thebibliography}{22} \makeatletter \providecommand \@ifxundefined [1]{ \@ifx{#1\undefined} } \providecommand \@ifnum [1]{ \ifnum #1\expandafter \@firstoftwo \else \expandafter \@secondoftwo \fi } \providecommand \@ifx [1]{ \ifx #1\expandafter \@firstoftwo \else \expandafter \@secondoftwo \fi } \providecommand \natexlab [1]{#1} \providecommand \enquote [1]{``#1''} \providecommand \bibnamefont [1]{#1} \providecommand \bibfnamefont [1]{#1} \providecommand \citenamefont [1]{#1} \providecommand \href@noop [0]{\@secondoftwo} \providecommand \href [0]{\begingroup \@sanitize@url \@href} \providecommand \@href[1]{\@@startlink{#1}\@@href} \providecommand \@@href[1]{\endgroup#1\@@endlink} \providecommand \@sanitize@url [0]{\catcode `\\12\catcode `\$12\catcode `\&12\catcode `\#12\catcode `\^12\catcode `\_12\catcode `\ \providecommand \@@startlink[1]{} \providecommand \@@endlink[0]{} \providecommand \url [0]{\begingroup\@sanitize@url \@url } \providecommand \@url [1]{\endgroup\@href {#1}{\urlprefix }} \providecommand \urlprefix [0]{URL } \providecommand \Eprint [0]{\href } \providecommand \doibase [0]{http://dx.doi.org/} \providecommand \selectlanguage [0]{\@gobble} \providecommand \bibinfo [0]{\@secondoftwo} \providecommand \bibfield [0]{\@secondoftwo} \providecommand \translation [1]{[#1]} \providecommand \BibitemOpen [0]{} \providecommand \bibitemStop [0]{} \providecommand \bibitemNoStop [0]{.\EOS\space} \providecommand \EOS [0]{\spacefactor3000\relax} \providecommand \BibitemShut [1]{\csname bibitem#1\endcsname} \let\auto@bib@innerbib\@empty \bibitem [{\citenamefont {Neil}\ \emph {et~al.}(2000)\citenamefont {Neil}, \citenamefont {Bohn}, \citenamefont {Benson}, \citenamefont {Biallas}, \citenamefont {Douglas}, \citenamefont {Dylla}, \citenamefont {Evans}, \citenamefont {Fugitt}, \citenamefont {Grippo}, \citenamefont {Gubeli}, \citenamefont {Hill}, \citenamefont {Jordan}, \citenamefont {Li}, \citenamefont {Merminga}, \citenamefont {Piot}, \citenamefont {Preble}, \citenamefont {Shinn}, \citenamefont {Siggins}, \citenamefont {Walker},\ and\ \citenamefont {Yunn}}]{ref:TJNAF} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {G.~R.}\ \bibnamefont {Neil}}, \bibinfo {author} {\bibfnamefont {C.~L.}\ \bibnamefont {Bohn}}, \bibinfo {author} {\bibfnamefont {S.~V.}\ \bibnamefont {Benson}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Biallas}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Douglas}}, \bibinfo {author} {\bibfnamefont {H.~F.}\ \bibnamefont {Dylla}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Evans}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Fugitt}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Grippo}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Gubeli}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Hill}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Jordan}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Li}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Merminga}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Piot}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Preble}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Shinn}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Siggins}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Walker}}, \ and\ \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Yunn}},\ }\href {\doibase 10.1103/PhysRevLett.84.662} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {84}},\ \bibinfo {pages} {662} (\bibinfo {year} {2000})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Bazarov}\ \emph {et~al.}(2012)\citenamefont {Bazarov}, \citenamefont {Belomestnykh}, \citenamefont {Bilderback}, \citenamefont {Billing}, \citenamefont {Brock}, \citenamefont {Buckley}, \citenamefont {Chapman}, \citenamefont {Chojnacki}, \citenamefont {Conway}, \citenamefont {Crittenden}, \citenamefont {Dale}, \citenamefont {Dobbins}, \citenamefont {Dunham}, \citenamefont {Ehrlich}, \citenamefont {Ehrlichman}, \citenamefont {Finkelstein}, \citenamefont {Fontes}, \citenamefont {Forster}, \citenamefont {Gray}, \citenamefont {Greenwald}, \citenamefont {Gruner}, \citenamefont {Gulliford}, \citenamefont {Hartill}, \citenamefont {Helmke}, \citenamefont {Hoffstaetter}, \citenamefont {Kazimirov}, \citenamefont {Kaplan}, \citenamefont {Karkare}, \citenamefont {Kostroun}, \citenamefont {Laham}, \citenamefont {Lau}, \citenamefont {Li}, \citenamefont {Liu}, \citenamefont {Liepe}, \citenamefont {Loehl}, \citenamefont {Cultrera}, \citenamefont {Miyajima}, \citenamefont {Mayes}, \citenamefont {Maxson}, \citenamefont {Meseck}, \citenamefont {Mikhailichenko}, \citenamefont {Ouzounov}, \citenamefont {Padamsee}, \citenamefont {Peck}, \citenamefont {Pfeifer}, \citenamefont {Posen}, \citenamefont {Quigley}, \citenamefont {Revesz}, \citenamefont {Rice}, \citenamefont {Sae-Ueng}, \citenamefont {Sagan}, \citenamefont {Sears}, \citenamefont {Shemelin}, \citenamefont {Sinclair}, \citenamefont {Smilgies}, \citenamefont {Smith}, \citenamefont {Smolenski}, \citenamefont {Spethmann}, \citenamefont {Song}, \citenamefont {Tanabe}, \citenamefont {Temnykh}, \citenamefont {Tigner}, \citenamefont {Valles}, \citenamefont {Veshcherevich}, \citenamefont {Wang}, \citenamefont {Woll}, \citenamefont {Xie},\ and\ \citenamefont {Zhao}}]{ref:pddr} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {I.~V.}\ \bibnamefont {Bazarov}}, \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont {Belomestnykh}}, \bibinfo {author} {\bibfnamefont {D.~H.}\ \bibnamefont {Bilderback}}, \bibinfo {author} {\bibfnamefont {M.~G.}\ \bibnamefont {Billing}}, \bibinfo {author} {\bibfnamefont {J.~D.}\ \bibnamefont {Brock}}, \bibinfo {author} {\bibfnamefont {B.~W.}\ \bibnamefont {Buckley}}, \bibinfo {author} {\bibfnamefont {S.~S.}\ \bibnamefont {Chapman}}, \bibinfo {author} {\bibfnamefont {E.~P.}\ \bibnamefont {Chojnacki}}, \bibinfo {author} {\bibfnamefont {Z.~A.}\ \bibnamefont {Conway}}, \bibinfo {author} {\bibfnamefont {J.~A.}\ \bibnamefont {Crittenden}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Dale}}, \bibinfo {author} {\bibfnamefont {J.~A.}\ \bibnamefont {Dobbins}}, \bibinfo {author} {\bibfnamefont {B.~M.}\ \bibnamefont {Dunham}}, \bibinfo {author} {\bibfnamefont {R.~D.}\ \bibnamefont {Ehrlich}}, \bibinfo {author} {\bibfnamefont {M.~P.}\ \bibnamefont {Ehrlichman}}, \bibinfo {author} {\bibfnamefont {K.~D.}\ \bibnamefont {Finkelstein}}, \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Fontes}}, \bibinfo {author} {\bibfnamefont {M.~J.}\ \bibnamefont {Forster}}, \bibinfo {author} {\bibfnamefont {S.~W.}\ \bibnamefont {Gray}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Greenwald}}, \bibinfo {author} {\bibfnamefont {S.~M.}\ \bibnamefont {Gruner}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Gulliford}}, \bibinfo {author} {\bibfnamefont {D.~L.}\ \bibnamefont {Hartill}}, \bibinfo {author} {\bibfnamefont {R.~G.}\ \bibnamefont {Helmke}}, \bibinfo {author} {\bibfnamefont {G.~H.}\ \bibnamefont {Hoffstaetter}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Kazimirov}}, \bibinfo {author} {\bibfnamefont {R.~P.}\ \bibnamefont {Kaplan}}, \bibinfo {author} {\bibfnamefont {S.~S.}\ \bibnamefont {Karkare}}, \bibinfo {author} {\bibfnamefont {V.~O.}\ \bibnamefont {Kostroun}}, \bibinfo {author} {\bibfnamefont {F.~A.}\ \bibnamefont {Laham}}, \bibinfo {author} {\bibfnamefont {Y.~H.}\ \bibnamefont {Lau}}, \bibinfo {author} {\bibfnamefont {Y.}~\bibnamefont {Li}}, \bibinfo {author} {\bibfnamefont {X.}~\bibnamefont {Liu}}, \bibinfo {author} {\bibfnamefont {M.~U.}\ \bibnamefont {Liepe}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Loehl}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Cultrera}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Miyajima}}, \bibinfo {author} {\bibfnamefont {C.~E.}\ \bibnamefont {Mayes}}, \bibinfo {author} {\bibfnamefont {J.~M.}\ \bibnamefont {Maxson}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Meseck}}, \bibinfo {author} {\bibfnamefont {A.~A.}\ \bibnamefont {Mikhailichenko}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Ouzounov}}, \bibinfo {author} {\bibfnamefont {H.~S.}\ \bibnamefont {Padamsee}}, \bibinfo {author} {\bibfnamefont {S.~B.}\ \bibnamefont {Peck}}, \bibinfo {author} {\bibfnamefont {M.~A.}\ \bibnamefont {Pfeifer}}, \bibinfo {author} {\bibfnamefont {S.~E.}\ \bibnamefont {Posen}}, \bibinfo {author} {\bibfnamefont {P.~G.}\ \bibnamefont {Quigley}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Revesz}}, \bibinfo {author} {\bibfnamefont {D.~H.}\ \bibnamefont {Rice}}, \bibinfo {author} {\bibfnamefont {U.}~\bibnamefont {Sae-Ueng}}, \bibinfo {author} {\bibfnamefont {D.~C.}\ \bibnamefont {Sagan}}, \bibinfo {author} {\bibfnamefont {J.~O.}\ \bibnamefont {Sears}}, \bibinfo {author} {\bibfnamefont {V.~D.}\ \bibnamefont {Shemelin}}, \bibinfo {author} {\bibfnamefont {C.~K.}\ \bibnamefont {Sinclair}}, \bibinfo {author} {\bibfnamefont {D.~M.}\ \bibnamefont {Smilgies}}, \bibinfo {author} {\bibfnamefont {E.~N.}\ \bibnamefont {Smith}}, \bibinfo {author} {\bibfnamefont {K.~W.}\ \bibnamefont {Smolenski}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Spethmann}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Song}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Tanabe}}, \bibinfo {author} {\bibfnamefont {A.~B.}\ \bibnamefont {Temnykh}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Tigner}}, \bibinfo {author} {\bibfnamefont {N.~R.~A.}\ \bibnamefont {Valles}}, \bibinfo {author} {\bibfnamefont {V.~G.}\ \bibnamefont {Veshcherevich}}, \bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Wang}}, \bibinfo {author} {\bibfnamefont {A.~R.}\ \bibnamefont {Woll}}, \bibinfo {author} {\bibfnamefont {Y.}~\bibnamefont {Xie}}, \ and\ \bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Zhao}},\ }\href@noop {} {\emph {\bibinfo {title} {Cornell Energy Recovery Linac Project Definition Design Report}}},\ \bibinfo {type} {Tech. Rep.}\ (\bibinfo {institution} {Cornell University},\ \bibinfo {year} {2012})\BibitemShut {NoStop} \bibitem [{\citenamefont {Emma}\ \emph {et~al.}(2010)\citenamefont {Emma}, \citenamefont {Akre}, \citenamefont {Arthur}, \citenamefont {Bionta}, \citenamefont {Bostedt}, \citenamefont {Bozek}, \citenamefont {Brachmann}, \citenamefont {Bucksbaum}, \citenamefont {Coffee}, \citenamefont {Decker}, \citenamefont {Ding}, \citenamefont {Dowell}, \citenamefont {Edstrom}, \citenamefont {Fisher}, \citenamefont {Frisch}, \citenamefont {Gilevich}, \citenamefont {Hastings}, \citenamefont {Hays}, \citenamefont {Hering}, \citenamefont {Huang}, \citenamefont {Iverson}, \citenamefont {Loos}, \citenamefont {Messerschmidt}, \citenamefont {Miahnahri}, \citenamefont {Moeller}, \citenamefont {Nuhn}, \citenamefont {Pile}, \citenamefont {Ratner}, \citenamefont {Rzepiela}, \citenamefont {Schultz}, \citenamefont {Smith}, \citenamefont {Stefan}, \citenamefont {Tompkins}, \citenamefont {Turner}, \citenamefont {Welch}, \citenamefont {White}, \citenamefont {Wu}, \citenamefont {Yocky},\ and\ \citenamefont {Galayda}}]{ref:lcls} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Emma}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Akre}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Arthur}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Bionta}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Bostedt}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Bozek}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Brachmann}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Bucksbaum}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Coffee}}, \bibinfo {author} {\bibfnamefont {F.-J.}\ \bibnamefont {Decker}}, \bibinfo {author} {\bibfnamefont {Y.}~\bibnamefont {Ding}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Dowell}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Edstrom}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Fisher}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Frisch}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Gilevich}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Hastings}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Hays}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Hering}}, \bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Huang}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Iverson}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Loos}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Messerschmidt}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Miahnahri}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Moeller}}, \bibinfo {author} {\bibfnamefont {H.-D.}\ \bibnamefont {Nuhn}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Pile}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Ratner}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Rzepiela}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Schultz}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Smith}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Stefan}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Tompkins}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Turner}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Welch}}, \bibinfo {author} {\bibfnamefont {W.}~\bibnamefont {White}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Wu}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Yocky}}, \ and\ \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Galayda}},\ }\href {http://dx.doi.org/10.1038/nphoton.2010.176} {\bibfield {journal} {\bibinfo {journal} {Nat Photon}\ }\textbf {\bibinfo {volume} {4}},\ \bibinfo {pages} {641} (\bibinfo {year} {2010})}\BibitemShut {NoStop} \bibitem [{ref(2004)}]{ref:eRHIC} \BibitemOpen \href@noop {} {\emph {\bibinfo {title} {eRHIC, A Future Electron-Ion Collider at BNL}}}\ (\bibinfo {organization} {EPAC 2004},\ \bibinfo {address} {Lucerne, Switzerland},\ \bibinfo {year} {2004})\BibitemShut {NoStop} \bibitem [{\citenamefont {Hastings}\ \emph {et~al.}(2006)\citenamefont {Hastings}, \citenamefont {Rudakov}, \citenamefont {Dowell}, \citenamefont {Schmerge}, \citenamefont {Cardoza}, \citenamefont {Castra}, \citenamefont {Gierman}, \citenamefont {Loos},\ and\ \citenamefont {Weber}}]{ref:uedslac1} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {J.~B.}\ \bibnamefont {Hastings}}, \bibinfo {author} {\bibfnamefont {F.~M.}\ \bibnamefont {Rudakov}}, \bibinfo {author} {\bibfnamefont {D.~H.}\ \bibnamefont {Dowell}}, \bibinfo {author} {\bibfnamefont {J.~F.}\ \bibnamefont {Schmerge}}, \bibinfo {author} {\bibfnamefont {J.~D.}\ \bibnamefont {Cardoza}}, \bibinfo {author} {\bibfnamefont {J.~M.}\ \bibnamefont {Castra}}, \bibinfo {author} {\bibfnamefont {S.~M.}\ \bibnamefont {Gierman}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Loos}}, \ and\ \bibinfo {author} {\bibfnamefont {P.~M.}\ \bibnamefont {Weber}},\ }\href@noop {} {\emph {\bibinfo {title} {Ultrafast Time-resolved Electron Diffraction with Megavolt Electron Beams}}},\ \bibinfo {type} {Tech. Rep.}\ \bibinfo {number} {12162}\ (\bibinfo {institution} {SLAC},\ \bibinfo {year} {2006})\BibitemShut {NoStop} \bibitem [{\citenamefont {King}\ \emph {et~al.}(2005)\citenamefont {King}, \citenamefont {Campbell}, \citenamefont {Frank}, \citenamefont {Reed}, \citenamefont {Schmerge}, \citenamefont {Siwick}, \citenamefont {Stuart},\ and\ \citenamefont {Weber}}]{ref:uedutor1} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {W.~E.}\ \bibnamefont {King}}, \bibinfo {author} {\bibfnamefont {G.~H.}\ \bibnamefont {Campbell}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Frank}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Reed}}, \bibinfo {author} {\bibfnamefont {J.~F.}\ \bibnamefont {Schmerge}}, \bibinfo {author} {\bibfnamefont {B.~J.}\ \bibnamefont {Siwick}}, \bibinfo {author} {\bibfnamefont {B.~C.}\ \bibnamefont {Stuart}}, \ and\ \bibinfo {author} {\bibfnamefont {P.~M.}\ \bibnamefont {Weber}},\ }\href {\doibase http://dx.doi.org/10.1063/1.1927699} {\bibfield {journal} {\bibinfo {journal} {Journal of Applied Physics}\ }\textbf {\bibinfo {volume} {97}},\ \bibinfo {eid} {111101} (\bibinfo {year} {2005})}\BibitemShut {NoStop} \bibitem [{\citenamefont {van Oudheusden}\ \emph {et~al.}(2007)\citenamefont {van Oudheusden}, \citenamefont {de~Jong}, \citenamefont {van~der Geer}, \citenamefont {'t~Root}, \citenamefont {Luiten},\ and\ \citenamefont {Siwick}}]{ref:uedutor2} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {van Oudheusden}}, \bibinfo {author} {\bibfnamefont {E.~F.}\ \bibnamefont {de~Jong}}, \bibinfo {author} {\bibfnamefont {S.~B.}\ \bibnamefont {van~der Geer}}, \bibinfo {author} {\bibfnamefont {W.~P. E. M.~O.}\ \bibnamefont {'t~Root}}, \bibinfo {author} {\bibfnamefont {O.~J.}\ \bibnamefont {Luiten}}, \ and\ \bibinfo {author} {\bibfnamefont {B.~J.}\ \bibnamefont {Siwick}},\ }\href {\doibase http://dx.doi.org/10.1063/1.2801027} {\bibfield {journal} {\bibinfo {journal} {Journal of Applied Physics}\ }\textbf {\bibinfo {volume} {102}},\ \bibinfo {eid} {093501} (\bibinfo {year} {2007})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Wang}(2012)}]{ref:uedbnl1} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {X.}~\bibnamefont {Wang}},\ }in\ \href {\doibase 10.1364/ICUSD.2012.IT3D.1} {\emph {\bibinfo {booktitle} {Research in Optical Sciences}}}\ (\bibinfo {publisher} {Optical Society of America},\ \bibinfo {year} {2012})\ p.\ \bibinfo {pages} {IT3D.1}\BibitemShut {NoStop} \bibitem [{\citenamefont {Anthony}\ \emph {et~al.}(2004)\citenamefont {Anthony}, \citenamefont {Arnold}, \citenamefont {Arroyo}, \citenamefont {Baird}, \citenamefont {Bega}, \citenamefont {Biesiada}, \citenamefont {Bosted}, \citenamefont {Breuer}, \citenamefont {Carr}, \citenamefont {Cates}, \citenamefont {Chen}, \citenamefont {Chudakov}, \citenamefont {Cooke}, \citenamefont {Decker}, \citenamefont {Decowski}, \citenamefont {Deur}, \citenamefont {Emam}, \citenamefont {Erickson}, \citenamefont {Fieguth}, \citenamefont {Field}, \citenamefont {Gao}, \citenamefont {Gustafsson}, \citenamefont {Hicks}, \citenamefont {Holmes}, \citenamefont {Hughes}, \citenamefont {Humensky}, \citenamefont {Jones}, \citenamefont {Kaufman}, \citenamefont {Kolomensky}, \citenamefont {Kumar}, \citenamefont {Lhuillier}, \citenamefont {Lombard-Nelsen}, \citenamefont {Mastromarino}, \citenamefont {Mayer}, \citenamefont {McKeown}, \citenamefont {Michaels}, \citenamefont {Olson}, \citenamefont {Paschke}, \citenamefont {Peterson}, \citenamefont {Pitthan}, \citenamefont {Pope}, \citenamefont {Relyea}, \citenamefont {Rock}, \citenamefont {Saxton}, \citenamefont {Shapiro}, \citenamefont {Singh}, \citenamefont {Souder}, \citenamefont {Szalata}, \citenamefont {Tobias}, \citenamefont {Tonguc}, \citenamefont {Turner}, \citenamefont {Tweedie}, \citenamefont {Vacheret}, \citenamefont {Walz}, \citenamefont {Weber}, \citenamefont {Weisend}, \citenamefont {Whittum}, \citenamefont {Woods},\ and\ \citenamefont {Younus}}]{ref:mixingangle1} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Anthony}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Arnold}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Arroyo}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Baird}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Bega}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Biesiada}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Bosted}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Breuer}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Carr}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Cates}}, \bibinfo {author} {\bibfnamefont {J.-P.}\ \bibnamefont {Chen}}, \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Chudakov}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Cooke}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Decker}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Decowski}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Deur}}, \bibinfo {author} {\bibfnamefont {W.}~\bibnamefont {Emam}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Erickson}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Fieguth}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Field}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Gao}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Gustafsson}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Hicks}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Holmes}}, \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Hughes}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Humensky}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Jones}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Kaufman}}, \bibinfo {author} {\bibfnamefont {Y.}~\bibnamefont {Kolomensky}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Kumar}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Lhuillier}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Lombard-Nelsen}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Mastromarino}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Mayer}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {McKeown}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Michaels}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Olson}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Paschke}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Peterson}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Pitthan}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Pope}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Relyea}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Rock}}, \bibinfo {author} {\bibfnamefont {O.}~\bibnamefont {Saxton}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Shapiro}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Singh}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Souder}}, \bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Szalata}}, \bibinfo {author} {\bibfnamefont {W.}~\bibnamefont {Tobias}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Tonguc}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Turner}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Tweedie}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Vacheret}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Walz}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Weber}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Weisend}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Whittum}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Woods}}, \ and\ \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Younus}} (\bibinfo {collaboration} {SLAC E158 Collaboration}),\ }\href {\doibase 10.1103/PhysRevLett.92.181602} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {92}},\ \bibinfo {pages} {181602} (\bibinfo {year} {2004})}\BibitemShut {NoStop} \bibitem [{ref(2014)}]{ref:slacspecs1} \BibitemOpen \href@noop {} {\emph {\bibinfo {title} {Linear Accelerator Design for the LCLS-II FEL Facility}}},\ \bibinfo {number} {THP025}\ (\bibinfo {organization} {FEL 2014},\ \bibinfo {address} {Basel, Switzerland},\ \bibinfo {year} {2014})\BibitemShut {NoStop} \bibitem [{\citenamefont {Sannibale}\ \emph {et~al.}(2012)\citenamefont {Sannibale}, \citenamefont {Filippetto}, \citenamefont {Papadopoulos}, \citenamefont {Staples}, \citenamefont {Wells}, \citenamefont {Bailey}, \citenamefont {Baptiste}, \citenamefont {Corlett}, \citenamefont {Cork}, \citenamefont {De~Santis}, \citenamefont {Dimaggio}, \citenamefont {Doolittle}, \citenamefont {Doyle}, \citenamefont {Feng}, \citenamefont {Garcia~Quintas}, \citenamefont {Huang}, \citenamefont {Huang}, \citenamefont {Kramasz}, \citenamefont {Kwiatkowski}, \citenamefont {Lellinger}, \citenamefont {Moroz}, \citenamefont {Norum}, \citenamefont {Padmore}, \citenamefont {Pappas}, \citenamefont {Portmann}, \citenamefont {Vecchione}, \citenamefont {Vinco}, \citenamefont {Zolotorev},\ and\ \citenamefont {Zucca}}]{ref:LBNL} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Sannibale}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Filippetto}}, \bibinfo {author} {\bibfnamefont {C.~F.}\ \bibnamefont {Papadopoulos}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Staples}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Wells}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Bailey}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Baptiste}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Corlett}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Cork}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {De~Santis}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Dimaggio}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Doolittle}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Doyle}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Feng}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Garcia~Quintas}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Huang}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Huang}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Kramasz}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Kwiatkowski}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Lellinger}}, \bibinfo {author} {\bibfnamefont {V.}~\bibnamefont {Moroz}}, \bibinfo {author} {\bibfnamefont {W.~E.}\ \bibnamefont {Norum}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Padmore}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Pappas}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Portmann}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Vecchione}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Vinco}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Zolotorev}}, \ and\ \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Zucca}},\ }\href {\doibase 10.1103/PhysRevSTAB.15.103501} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. ST Accel. Beams}\ }\textbf {\bibinfo {volume} {15}},\ \bibinfo {pages} {103501} (\bibinfo {year} {2012})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Serafini}\ and\ \citenamefont {Rosenzweig}(1997)}]{ref:rsecomp} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Serafini}}\ and\ \bibinfo {author} {\bibfnamefont {J.~B.}\ \bibnamefont {Rosenzweig}},\ }\href {\doibase 10.1103/PhysRevE.55.7565} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. E}\ }\textbf {\bibinfo {volume} {55}},\ \bibinfo {pages} {7565} (\bibinfo {year} {1997})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Maxson}\ \emph {et~al.}(2013)\citenamefont {Maxson}, \citenamefont {Bazarov}, \citenamefont {Wan}, \citenamefont {Padmore},\ and\ \citenamefont {Coleman-Smith}}]{ref:disorderedheating} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {J.~M.}\ \bibnamefont {Maxson}}, \bibinfo {author} {\bibfnamefont {I.~V.}\ \bibnamefont {Bazarov}}, \bibinfo {author} {\bibfnamefont {W.}~\bibnamefont {Wan}}, \bibinfo {author} {\bibfnamefont {H.~A.}\ \bibnamefont {Padmore}}, \ and\ \bibinfo {author} {\bibfnamefont {C.~E.}\ \bibnamefont {Coleman-Smith}},\ }\href@noop {} {\bibfield {journal} {\bibinfo {journal} {New J. Phys.}\ }\textbf {\bibinfo {volume} {15}} (\bibinfo {year} {2013})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Carlsten}(1989)}]{ref:ecomp} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {B.~E.}\ \bibnamefont {Carlsten}},\ }\href@noop {} {\bibfield {journal} {\bibinfo {journal} {Nuclear Instruments and Methods in Physics Research}\ }\textbf {\bibinfo {volume} {A}},\ \bibinfo {pages} {313} (\bibinfo {year} {1989})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Bazarov}(2012)}]{ref:syncradps} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {I.~V.}\ \bibnamefont {Bazarov}},\ }\href {\doibase 10.1103/PhysRevSTAB.15.050703} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. ST Accel. Beams}\ }\textbf {\bibinfo {volume} {15}},\ \bibinfo {pages} {050703} (\bibinfo {year} {2012})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Bazarov}\ \emph {et~al.}(2009)\citenamefont {Bazarov}, \citenamefont {Dunham},\ and\ \citenamefont {Sinclair}}]{ref:maxbb} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {I.~V.}\ \bibnamefont {Bazarov}}, \bibinfo {author} {\bibfnamefont {B.~M.}\ \bibnamefont {Dunham}}, \ and\ \bibinfo {author} {\bibfnamefont {C.~K.}\ \bibnamefont {Sinclair}},\ }\href {\doibase 10.1103/PhysRevLett.102.104801} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {102}},\ \bibinfo {pages} {104801} (\bibinfo {year} {2009})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Gulliford}\ \emph {et~al.}(2013)\citenamefont {Gulliford}, \citenamefont {Bartnik}, \citenamefont {Bazarov}, \citenamefont {Cultrera}, \citenamefont {Dobbins}, \citenamefont {Dunham}, \citenamefont {Gonzalez}, \citenamefont {Karkare}, \citenamefont {Lee}, \citenamefont {Li}, \citenamefont {Li}, \citenamefont {Liu}, \citenamefont {Maxson}, \citenamefont {Nguyen}, \citenamefont {Smolenski},\ and\ \citenamefont {Zhao}}]{ref:lowemitter} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Gulliford}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Bartnik}}, \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Bazarov}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Cultrera}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Dobbins}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Dunham}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Gonzalez}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Karkare}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Lee}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Li}}, \bibinfo {author} {\bibfnamefont {Y.}~\bibnamefont {Li}}, \bibinfo {author} {\bibfnamefont {X.}~\bibnamefont {Liu}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Maxson}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Nguyen}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Smolenski}}, \ and\ \bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Zhao}},\ }\href {\doibase 10.1103/PhysRevSTAB.16.073401} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. ST Accel. Beams}\ }\textbf {\bibinfo {volume} {16}},\ \bibinfo {pages} {073401} (\bibinfo {year} {2013})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Dunham}\ \emph {et~al.}(2013)\citenamefont {Dunham}, \citenamefont {Barley}, \citenamefont {Bartnik}, \citenamefont {Bazarov}, \citenamefont {Cultrera}, \citenamefont {Dobbins}, \citenamefont {Hoffstaetter}, \citenamefont {Johnson}, \citenamefont {Kaplan}, \citenamefont {Karkare}, \citenamefont {Kostroun}, \citenamefont {Li}, \citenamefont {Liepe}, \citenamefont {Liu}, \citenamefont {Loehl}, \citenamefont {Maxson}, \citenamefont {Quigley}, \citenamefont {Reilly}, \citenamefont {Rice}, \citenamefont {Sabol}, \citenamefont {Smith}, \citenamefont {Smolenski}, \citenamefont {Tigner}, \citenamefont {Vesherevich}, \citenamefont {Widger},\ and\ \citenamefont {Zhao}}]{ref:hcrecord} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Dunham}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Barley}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Bartnik}}, \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Bazarov}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Cultrera}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Dobbins}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Hoffstaetter}}, \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Johnson}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Kaplan}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Karkare}}, \bibinfo {author} {\bibfnamefont {V.}~\bibnamefont {Kostroun}}, \bibinfo {author} {\bibfnamefont {Y.}~\bibnamefont {Li}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Liepe}}, \bibinfo {author} {\bibfnamefont {X.}~\bibnamefont {Liu}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Loehl}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Maxson}}, \bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Quigley}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Reilly}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Rice}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Sabol}}, \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Smith}}, \bibinfo {author} {\bibfnamefont {K.}~\bibnamefont {Smolenski}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Tigner}}, \bibinfo {author} {\bibfnamefont {V.}~\bibnamefont {Vesherevich}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Widger}}, \ and\ \bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Zhao}},\ }\href {\doibase 10.1063/1.4789395} {\bibfield {journal} {\bibinfo {journal} {Applied Physics Letters}\ }\textbf {\bibinfo {volume} {102}},\ \bibinfo {eid} {034105} (\bibinfo {year} {2013})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Zhao}\ \emph {et~al.}(2012)\citenamefont {Zhao}, \citenamefont {Dunham}, \citenamefont {Bazarov},\ and\ \citenamefont {Wise}}]{ref:GHzlaser} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont {Zhao}}, \bibinfo {author} {\bibfnamefont {B.~M.}\ \bibnamefont {Dunham}}, \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Bazarov}}, \ and\ \bibinfo {author} {\bibfnamefont {F.~W.}\ \bibnamefont {Wise}},\ }\href {\doibase 10.1364/OE.20.004850} {\bibfield {journal} {\bibinfo {journal} {Opt. Express}\ }\textbf {\bibinfo {volume} {20}},\ \bibinfo {pages} {4850} (\bibinfo {year} {2012})}\BibitemShut {NoStop} \bibitem [{ref(2011)}]{ref:gpt1} \BibitemOpen \href {http://www.pulsar.nl/gpt/} {\enquote {\bibinfo {title} {Pulsar website for gpt},}\ }\bibinfo {howpublished} {} (\bibinfo {year} {2011})\BibitemShut {NoStop} \bibitem [{\citenamefont {Maxson}\ \emph {et~al.}(2014{\natexlab{a}})\citenamefont {Maxson}, \citenamefont {Bartnik},\ and\ \citenamefont {Bazarov}}]{ref:lasershaping1} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {J.~M.}\ \bibnamefont {Maxson}}, \bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont {Bartnik}}, \ and\ \bibinfo {author} {\bibfnamefont {I.~V.}\ \bibnamefont {Bazarov}},\ }\href@noop {} {\bibfield {journal} {\bibinfo {journal} {Appl. Phys. Lett.}\ }\textbf {\bibinfo {volume} {105}} (\bibinfo {year} {2014}{\natexlab{a}})}\BibitemShut {NoStop} \bibitem [{\citenamefont {Maxson}\ \emph {et~al.}(2014{\natexlab{b}})\citenamefont {Maxson}, \citenamefont {Lee}, \citenamefont {Bartnik}, \citenamefont {Kiefer},\ and\ \citenamefont {Bazarov}}]{ref:lasershaping2} \BibitemOpen \bibfield {author} {\bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Maxson}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Lee}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Bartnik}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Kiefer}}, \ and\ \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Bazarov}},\ }\href@noop {} {\bibfield {journal} {\bibinfo {journal} {submitted to Phys. Rev. ST Accel. Beams}\ } (\bibinfo {year} {2014}{\natexlab{b}})}\BibitemShut {NoStop} \end{thebibliography} |
1501.04083 | Title: Coherence and Rydberg blockade of atomic ensemble qubits
Abstract: We demonstrate $|W\rangle$ state encoding of multi-atom ensemble qubits.
Using optically trapped Rb atoms the $T_2$ coherence time is 2.6(3) ms for
$\bar N=7.6$ atoms and scales approximately inversely with the number of atoms.
Strong Rydberg blockade between two ensemble qubits is demonstrated with a
fidelity of $0.89(1)$ and a fidelity of $\sim \hspace{-.05cm}1.0$ when
postselected on control ensemble excitation. These results are a significant
step towards deterministic entanglement of atomic ensembles.
Body: \title{Coherence and Rydberg blockade of atomic ensemble qubits} \author{{M. Ebert}} \email{mebert@wisc.edu} \author{{M. Kwon}} \author{{T. G. Walker}} \author{{M. Saffman}} \affiliation{ Department of Physics, University of Wisconsin, 1150 University Avenue, Madison, Wisconsin 53706, USA } \date{\today} \begin{abstract} We demonstrate $\ket{W}$ state encoding of multi-atom ensemble qubits. Using optically trapped Rb atoms the $T_2$ coherence time is 2.6(3) ms for $\bar N=7.6$ atoms and scales approximately inversely with the number of atoms. Strong Rydberg blockade between two ensemble qubits is demonstrated with a fidelity of $0.89(1)$ and a fidelity of $\sim \hspace{-.05cm}1.0$ when postselected on control ensemble excitation. These results are a significant step towards deterministic entanglement of atomic ensembles. \end{abstract} \pacs{03.67.-a, 42.50.Dv, 32.80.Rm} \maketitle Qubits encoded in hyperfine states of neutral atoms are a promising approach for scalable implementation of quantum information processing. While a qubit can be encoded in a pair of ground states of a single atom, it is also possible to encode a qubit, or even multiple qubits, in an $N$ atom ensemble by using Rydberg blockade to enforce single excitation of one of the qubit states. Ensemble qubits have several interesting features in comparison to single atom qubits. Using an array of traps it is simpler to prepare many ensemble qubits with $N\ge 1$ for each ensemble, than it is to prepare an array with exactly one atom in each trap which remains an outstanding challenge. In addition, a $\ket{W}$ state ensemble qubit encoding is maximally robust against loss of a single atom, which can be remedied with error correction protocols, while atom loss is a critical error for single atom qubits. Furthermore an ensemble encoding facilitates strong coupling between atoms and light, an essential ingredient for quantum networking protocols and atomic control of photonic interactions in Rydberg blockaded ensembles. As the atom-light coupling strength grows with the number of atoms, recent experiments, and theory proposals are based on ensembles with $N>100$. We are focused here on studying the physics of ensembles for computational qubits and therefore work with smaller ensembles with up to $N\sim 10$ atoms. In this letter we demonstrate and study the coherence and interactions of atomic ensemble qubits. We measure the $T_2$ coherence time of ensemble qubits achieving a ratio of coherence time to single qubit $\pi$ rotation time of $\sim 2600$. We furthermore proceed to demonstrate strong Rydberg blockade between two, spatially separated ensemble qubits. Together with the recent demonstration of entanglement between a Rydberg excited ensemble and a propagating photon these results establish a path towards both local and remote entanglement of arrays of ensemble qubits, which will enable enhanced quantum repeater architectures. The computational basis states of the ensemble qubits are \begin{equation} \ket{\bar 0} = |0_{1} ... 0_{N}\rangle,~~ \ket{\bar 1} = \frac{1}{\sqrt N}\sum_{j=1}^N |0_{1}0_{2}... 1_j ... 0_{N}\rangle, \end{equation} where $\ket{0_j}$ and $\ket{1_j}$ are two ground states of the $j^{\rm th}$ atom in an $N$ atom sample. The state $\ket{\bar 1}$, which is a symmetric superposition of one of the $N$ atoms being excited, is commonly referred to as a $\ket{W}$ state in the quantum information literature. Gate protocols for ensemble qubits differ slightly from the single atom qubit case as all operations must use blockade to prohibit multi-atom excitation. Gate operations are performed via the collective, singly excited Rydberg state $$ \ket{\bar r} = \frac{1}{\sqrt N}\sum_{j=1}^N |0_{1}0_{2}... r_j ... 0_{N}\rangle, $$ where $\ket{r_j}$ is the Rydberg state of the $j^{\rm th}$ atom. A single qubit rotation $R(\theta,\phi)$ with area $\theta$ and phase $\phi$ between ensemble states $\ket{\bar 0}, \ket{\bar 1}$ is implemented as the three pulse sequence $\ket{\bar 1} \xrightarrow[\pi]{\Omega} \ket{\bar r},$ $\ket{\bar r} \xleftrightarrow[R(\theta,\phi)]{\Omega_N} \ket{\bar 0},$ $\ket{\bar r} \xrightarrow[\pi]{\Omega} \ket{\bar 1}$. Note that the coupling strength between states $\ket{\bar 1}, \ket{\bar r}$ is the single atom Rabi frequency $\Omega$ while the coupling between $\ket{\bar 0}, \ket{\bar r}$ is at the collective Rabi frequency $\Omega_N=\sqrt N \Omega.$ Since $\Omega_N$ depends on $N$, the one-qubit gate pulse lengths depend on the number of atoms. A $C_Z$ gate between control and target ensembles $\rm c,t$ is implemented as the three pulse sequence $\ket{\bar 1}_{\rm c} \xrightarrow[\pi]{\Omega} \ket{\bar r}_{\rm c},$ $\ket{\bar 1}_{\rm t} \xleftrightarrow[2\pi]{\Omega} \ket{\bar r}_{\rm t},$ $\ket{\bar r}_{\rm c} \xrightarrow[\pi]{\Omega} \ket{\bar 1}_{\rm c}$. The $C_Z$ gate pulses do not depend on the number of atoms. The $N$ dependence of the one-qubit gates can be strongly suppressed using adiabatic pulse sequences so that high fidelity gate operations are possible with small, but unknown values of $N$. The experimental setting is as described in . In brief we prepare a cold sample of $^{87}$Rb atoms in a magneto-optical trap (MOT) and then load a variable number of atoms into optical dipole traps. The dipole traps shown in Fig. are formed by focusing 1064 nm light to waists ($1/e^2$ intensity radii) of $3.0 ~\mu\rm m$. The atoms are cooled to a temperature of $\sim 150~\mu\rm K$ in 1-1.5 mK deep optical potentials. This gives approximately Gaussian shaped density distributions with typical standard deviations $\sigma_\perp= 0.7~\mu \rm m$ perpendicular to the long trap axis and $\sigma_z=7~\mu\rm m$ parallel to the long axis. The estimated density at trap center is $n/N= 5\times 10^{16} ~\rm m^{-3}$. We apply a bias magnetic field along the trap axis of $B_z=0.24~\rm mT$ and optically pump into $\ket{0}\equiv\ket{5s_{1/2},f=2,m_f=0}$ using $\pi$ polarized 795 nm light resonant with $\ket{5s_{1/2},f=2} \rightarrow \ket{5p_{1/2},f=2}$ and 780 nm repump light resonant with $\ket{5s_{1/2},f=1} \rightarrow \ket{5p_{3/2},f=2}$. Rydberg excitation coupling $\ket{\bar 0}, \ket{\bar r}$ is performed by off-resonant two-photon transitions via $5p_{3/2}$ using counter-propagating $780_0$ and 480 nm light. With $\sigma_+$ polarization for both beams we couple to the Rydberg state $\ket{r}=\ket{nd_{5/2},m_j=5/2}$ which is selected with a $B_z=0.37~\rm mT$ bias field. The other qubit ground state is $\ket{1}\equiv\ket{5s_{1/2},f=1,m_f=0}$. Coupling between $\ket{\bar 1}, \ket{\bar r}$ is performed with $780_1$ and 480 nm light where $780_0$ and $780_1$ have the same propagation vector and polarization but a frequency difference of 6.8 GHz corresponding to the $^{87}$Rb $f=1\leftrightarrow f=2$ clock frequency. In the experiments reported below we used Rydberg levels $97d_{5/2}$ and $111d_{5/2}$. In both cases strong blockade was observed in individual ensembles with no evidence for double excitation of the logical $|\bar 1\rangle$ state. While we do not observe double excitation of $\ket{\bar 1}$, experiments with two ensembles do show evidence for double excitation of the Rydberg state $\ket{\bar r}$, which plays a role in limiting the fidelity with which we can prepare the $|\bar1\rangle$ state. We proceed to demonstrate the coherence of the ensemble states of Eq. () using Ramsey interferometry. The amplitude of the Ramsey signal is used to quantify the presence of $N$ atom entanglement in the ensemble, as has been observed in other recent experiments. Details of the analysis showing that $82\pm6\ \begin{equation} \ket{\psi}= R_1(\pi)R_0(\pi/2) R_1(\pi)G(t)R_1(\pi)R_0(\pi/2)\ket{\bar 0}. \end{equation} Here $R_0(\theta)$ is a pulse of area $\theta$ between states $\ket{\bar 0}, \ket{\bar r}$ and $R_1(\theta)$ is a pulse of area $\theta$ between states $\ket{\bar 1}, \ket{\bar r}$. The first $R_0(\pi/2)$ pulse creates an equal superposition $\frac{\ket{\bar 0}+\ket{\bar r}}{\sqrt2}$. This is then mapped to $\frac{\ket{\bar 0}+\ket{\bar 1}}{\sqrt2}$ with a $R_1(\pi) $ pulse, we wait a gap time $t$ described by an operator $G(t)$, map $\ket{\bar 1}\rightarrow \ket{\bar r}$ with a $R_1(\pi)$ pulse, and then perform another $\pi/2$ pulse between $\ket{\bar 0}, \ket{\bar r}$. Finally, any population left in $\ket{\bar r}$ is mapped back to $\ket{\bar 1}$ with another $R_1(\pi)$ pulse. Atoms in state $\ket{0}$ are then pushed out of the trap using unbalanced radiation pressure from a beam resonant with $\ket{5s_{1/2},f=2}\rightarrow \ket{5p_{3/2}, f=3}$ while the dipole trap light is chopped on and off. For the push out step a bias field is applied along $x$ the narrow axis of the dipole traps, and the circularly polarized push out beam propagates along $x$. This is followed by a measurement of the number of atoms remaining in the dipole trap. The resulting data are shown in Fig. . The amplitude of the Ramsey interference at short gap times is limited by the $|W\rangle$ state preparation fidelity of about 50\ atom number used in the figure. The fidelities of the $R_0(\pi)$ and $R_1(\pi)$ pulses used to prepare $|W\rangle$ are estimated to each be at least 90\ The principal sources of decoherence in this experiment are expected to be magnetic noise, motional dephasing, and atomic collisions. For small atom numbers and low collision rates we fit the Ramsey signal to the expression $v_{\rm b}(t,T_{2})=v_0/[1+(e^{2/3}-1)(\frac{t}{T_{2}})^{2}]^{3/2}$ and in the collision dominated regime we use a Gaussian form $v_{\rm a}(t)=v_0 e^{-(t/T_2)^2}$ where $v_0$ is the amplitude at $t=0$. Both functional forms give the same $T_2$ time within our experimental error bars of $T_2 = 2.6 \pm 0.3~\rm ms$. The $\pi$ pulse times were $0.24~\mu\rm s$ for $\ket{\bar 0}\rightarrow \ket{\bar r}$, $0.06~\mu\rm s$ for the gap between pulses, and $0.68~\mu\rm s$ for $\ket{\bar r}\rightarrow \ket{\bar 1}$ giving a coherence to $R(\pi)$ gate time ratio of approximately $2600$. To further clarify the sensitivity to collisional dephasing Fig. shows the measured $T_2$ for different $\bar N$, including the case of $N=1$ Fock states which are selected using an additional fluorescence measurement before the Ramsey sequence. We see that $T_2\sim 1/{\bar N}$, in contrast to the $1/N^2$ scaling observed for GHZ states. The observed $1/\bar N$ scaling for $|W\rangle$ states is expected for decoherence dominated by collisions since the collision rate per atom is proportional to $\bar N$. For comparison, the $T_2$ time was also measured for product states $\ket{\psi}\sim (\ket{0}-i\ket{1})^{\otimes N}$. These states were prepared using a two-frequency Raman laser coupling $\ket{0}$ and $\ket{1}$ via the $5p_{3/2}$ level as shown in Fig. . Comparison of the $\ket{\bar 1}$ ($\ket{W}$ state) and product state coherence data suggests that for $N \gtrsim 5$ the coherence time is limited by collisions. For $\bar N<5$ as well as for the $N=1$ Fock state data the product states show longer coherence time. The coherence of the $\ket{W}$ states is measured by comparison with a phase reference defined by the beatnote of the $780_0$ and $780_1$ Rydberg lasers which have a measured beatnote linewidth of 100 Hz FWHM. This linewidth is consistent with the observed shorter coherence time of the $\ket{W}$ states compared to the product states which are referenced to the Raman laser beatnote which is in turn locked to a stable 6.8 GHz microwave oscillator. We anticipate that compensated optical traps and dynamical decoupling methods together with an optical lattice to reduce collisional effects can be used to greatly extend these coherence times. To demonstrate ensemble-ensemble blockade we load atoms into control (c) and target (t) dipole traps, optically pump into $\ket{\bar 0}_c\ket{\bar 0 }_t$ and apply one of two sequences. Preparation of a superposition of $\ket{\bar 0}$ and $\ket{\bar 1}$ in the target qubit is effected by the sequence $U_a\ket{\bar 0}_{\rm c}\ket{\bar 0}_{\rm t}=R_{1,\rm t}(\pi)R_{0,\rm t}(\theta)\ket{\bar 0}_{\rm c}\ket{\bar 0}_{\rm t}$. This should ideally leave the qubits in the joint state $\ket{\bar 0}_{\rm c}\left[\cos(\theta/2)\ket{\bar 0}_{\rm t}-\sin(\theta/2)\ket{\bar 1}_{\rm t}\right]$ with the probability of preparing $\ket{\bar 1}_t$ proportional to $\sin^2(\theta/2),$ as is shown in Fig. a). We see the expected time dependence with a peak probability of $P_{\ket{\bar 1},\rm t}\sim 0.52$, consistent with our earlier study of Fock state preparation. Rydberg blockade between two ensembles is observed with the sequence $U_b\ket{\bar 0}_{\rm c}\ket{\bar 0}_{\rm t}=R_{1,c}(\pi)R_{1,t}(\pi)R_{0,t}(\theta)R_{0,c}(\pi)\ket{\bar 0}_{\rm c}\ket{\bar 0}_{\rm t}$. Here we have used state $\ket{\bar 0}$ of the control ensemble to block the target transfer with the final $R_{1,c}(\pi)$ pulse ideally leaving the qubits in the joint state $\ket{\bar 1}_{\rm c}\ket{\bar 0}_{\rm t}$. The data in Fig. a) show a ratio of $P_{\ket{\bar 1},t}(U_b)/P_{\ket{\bar 1},t}(U_a)=0.11(1)$, i.e. a blockade fidelity of 0.89. This implies that the success probability of the transition $R_{0,\rm c}(\pi) \ket{\bar 0}_{\rm c}\rightarrow \ket{\bar r}_c$ is bounded below by the $\ket{\bar 1}_{\rm t}$ population ratio for the two sequences. We infer that at least one atom is excited to the Rydberg state $\ket{r}_{\rm c}$ with probability $\ge 0.89(1)$. As a further check on the inter-site blockade fidelity, events where the control site ends in state $\ket{\bar 1}_{\rm c}$ after sequence $U_b$ are post selected. The observed post-selected target population is shown in Figure b), along with the expected blow-away leakage rate of the control and target sites which is measured to be $0.2\ From the data it can be seen that the post-selected results are consistent with perfect inter-site blockade. The observed high blockade fidelity exceeds that originally achieved in experiments with single atom qubits, and is certainly sufficient to create entanglement between ensemble qubits. What has so far limited a demonstration of deterministic entanglement is the relatively low probability of up to 62\ We believe this effect is due to long range interactions, where the amplitude for Rydberg atom excitation in the target site is sufficiently blockaded to prevent it from making the transfer to $\ket{\bar 1}_{\rm t}$ with any significant probability, yet the target ensemble Rydberg excitation still interacts with the control ensemble strongly enough to disrupt the control ensemble state transfer. A similar situation of partial blockade together with decoherence of multi-atom ground-Rydberg Rabi oscillations was reported earlier in . A two-atom Rydberg interaction effect should scale with the Rydberg double excitation probability, i.e. $P_2 \propto \Omega_{\bar{N}}^2/{\sf B}^2$, where $\sf B$ is the ensemble mean blockade shift. To check this, we extract the slopes from linear fits to the $P_{\ket{\bar 1}_{\rm c}}(\theta)$ data for small $\theta$ and compare to the scaling parameter \begin{equation} F = \Omega_{\bar{N}_{\rm t}}^2 \left[\frac{(n/n_0)^{12}}{(R/R_0)^6}\right]^{-2} \propto P_{\rm double}. \end{equation} Here $n$ is the Rydberg principal quantum number and $R$ is the site - site separation. The larger $F$ is for a given set of parameters, the stronger the Rydberg-Rydberg interaction, and thus the larger the slope of $d P_{\ket{\bar 1}_{\rm c}}(\theta)/d\theta.$ Indeed, this is the behavior we observe, as shown in Fig. b), for a range of $\bar{N}$, $R$, and $n$. This interaction effect hints at the possible mechanism responsible for the observed reduction in the probability $P_{\ket{\bar 1}}$ of preparing the collective qubit state in a single ensemble. The spatial extent of one ensemble is $\sim 2\sigma_z=14~\mu\rm m$ giving a length scale in between the lower two data sets in Fig. a). The intra-ensemble Rydberg interactions are significantly stronger than between atoms located in different ensembles at the same separation because the dipole-dipole interaction angular factors favor atom pairs separated along $z$. These considerations imply that lack of perfect blockade leading to long range Rydberg-Rydberg interactions in a single ensemble only partially explains the observed maximum of $P_{\ket{\bar 1}}=0.62$ . Another candidate explanation is very strong interactions at short range in a single ensemble which mix levels together and open anti-blockade resonance channels. The doubly excited molecular energy structure becomes difficult to calculate with confidence at short range, with many molecular potentials near resonant. For our typical Rydberg state $97d_{5/2}$ this characteristic separation is $\sim 5$ $\mu$m, and for a 6 atom sample with our ensemble spatial distributions an average of 7 atom pairs out of 15 have $R< 5~\mu\rm m$. We conjecture that the strong, short range interactions give an amplitude for double excitation, resulting in Rydberg-Rydberg interactions which dephase the ground-Rydberg rotations needed for state preparation, thereby limiting the probability of preparing the ensemble $\ket{\bar 1}$ state. A related reduction of the fidelity of Rydberg mediated atom-photon coupling in dense ensembles due to Rydberg-ground state interactions has also been observed. In conclusion, we have demonstrated the coherence of ensemble qubit basis states. The coherence time scales approximately inversely with the number of atoms, but is still several ms and $2600$ times longer than our characteristic gate time for $N\sim 10$. Additionally we have demonstrated inter-ensemble blockade with a fidelity of 0.89 and $\sim 1.0$ when post-selecting on control ensemble excitation. We identified Rydberg-Rydberg interactions from weak double excitations, either at long or short range, as a possible mechanism limiting the fidelity of ensemble state preparation. Future work towards ensemble entanglement and quantum computation will explore the use of a background optical lattice to better localize the ensembles while limiting uncontrolled short range interactions. This work was funded by NSF grant PHY-1104531 and the AFOSR Quantum Memories MURI. \bibliographystyle{apsrev4-1} \input{EnsembleQubits_resubmit_arxiv_v3.bbl} \newpage {\bf Coherence and Rydberg blockade of atomic ensemble qubits\\Supplementary Material} \section{Multipartite W-State Entanglement Verification} In order to demonstrate multipartite entanglement it is necessary to show that the results obtained in a measurement cannot be reproduced with a separable state. Thus we require that the $N$-particle state in question $|\psi^{N}\rangle$ satisfies \begin{equation} |\psi^{N}\rangle \neq |\psi_A^{K}\rangle \otimes |\psi_B^{N-K}\rangle, \end{equation} for any K in the range $N/2\leq K<N$. In this supplemental material we evaluate the observed signatures of $W$-state entanglement. These signatures include the $\sqrt{\bar{N}}$-enhancement of the Rabi frequency between $|\bar{0}\rangle$ and $|\bar r\rangle$, and the amplitude of the Ramsey oscillations. \subsection{Collective Rabi Frequency Enhancement} The interaction of an ensemble with a light field can be written in the basis of individual atom excitations $ |\{0,1\}^{(1)}\rangle \otimes |\{0,1\}^{(2)}\rangle \otimes ... |\{0,1\}^{(N)}\rangle$. The Hamiltonian $\mathbf{H}_{int}$ describing the evolution of the system is a block tridiagonal matrix. The basis states are denoted as $|n_k\rangle$, where $0\le n\le N$ is the eigenvalue of the excitation number operator $\hat{\mathcal N}= \sum_{k=1}^N \hat{S}_z^{(k)} + N/2,$ and the index $k$ labels the degenerate eigenstates, e.g. $|1_1\rangle=|10\cdots 0\rangle,$ $|1_2\rangle=|01\cdots 0\rangle$, etc. . Here $\hat{S}_{z}^{(k)}=\frac{1}{2}\hat\sigma_z^{(k)}$ is the effective spin operator for atom $k$ along $z$. In this basis $\mathbf{H}_{int}$ is given by: \begin{equation} \begin{split} &\mathbf{H}_{int} = \mathbf{A} + \mathbf{\Delta} =\\ &\begin{bmatrix} \mathbf{\Delta}_{0} & \mathbf{A}_{(0,1)} & 0 & \cdots & 0 \\ \mathbf{A}^{T}_{(0,1)} & \mathbf{\Delta}_{1} & \mathbf{A}_{(1,2)} & & \\ \vdots & \ddots & \ddots & \ddots & \vdots \\ & & \mathbf{A}^{T}_{(N-2,N-1)} & \mathbf{\Delta}_{N-1} & \mathbf{A}_{(N-1,N)} \\ 0 & \cdots & 0 & \mathbf{A}^{T}_{(N-1,N)} & \mathbf{\Delta}_{N} \end{bmatrix} \end{split} \end{equation} The matrix $\mathbf{H}_{int}$ has dimensions $2^N\times 2^N$, the dimension of $N$ 2-level systems. The dimension of the block diagonal sub-matrices is given by the binomial coefficient, $\textrm{dim}\left( \mathbf{\Delta}_n \right) = \tbinom N{n} \equiv N_{n}$. The sub-matrices, $\mathbf{\Delta}_{n}$, contain information concerning the sub-systems specific energy levels \begin{equation} \mathbf{\Delta}_{n} = \sum_{k=1}^{N_{n}}\delta^{(n)}_k |n_k\rangle\langle n_k| \end{equation} where $\delta^{(n)}_k$ refers to the energy of the $k^{\mathrm{th}}$ basis state in the subspace with eigenvalue $n$. The matrices on the upper and lower diagonals couple states with excitation numbers differing by $\pm1$, $|n_k\rangle \stackrel{\alpha}{\leftrightarrow} |n\pm 1_j\rangle$ with coupling strength $\alpha$ defined by \begin{eqnarray} \left[\mathbf{A}_{(n\pm 1,n)}\right]_{j,k}&=& \alpha_{jk}|(n\pm1)_j\rangle\langle n_k| \nonumber\\ & =& |(n\pm1)_j\rangle\langle (n\pm1)_j|\hat{A}|n_k\rangle\langle n_k|, \end{eqnarray} where \begin{equation} \hat{A} = \sum_{m=1}^{N} \alpha_m \hat{S}_x^{(m)} \end{equation} and $\alpha_m$ is the strength of the light-atom coupling at atom $m$. In an ideal Rydberg blockaded ensemble states with $n>1$ are not excited and all $\alpha_m$ are equal. Departures from the ideal case are accounted for by allowing for atom specific $\alpha_m$ and double excitations are included by truncating the basis at $n=2$ and adding the doubly excited interaction energies to $\mathbf\Delta_2$. A strong blockade shift, $\delta_m^{(n=2)} = \delta_{dd} \gg \alpha_m$ reduces the available Hilbert space for the problem to $n=\{ 0,1 \}$, and $\mathbf{H}_{int}$ becomes: \begin{equation} \mathbf{H}_{int} = \begin{bmatrix} 0 & \alpha_1 & \alpha_2 & \cdots & \alpha_N \\ \alpha_1 & \delta_1^{(1)} & 0 & \cdots & 0 \\ \alpha_2 & 0 & \delta_2^{(1)} & & 0 \\ \vdots & \vdots & & \ddots & \vdots \\ \alpha_N & 0 & 0 & \cdots & \delta_N^{(1)} \end{bmatrix} \end{equation} The detunings $\delta_m^{(1)}$ are nominally 0, so it makes sense to treat the $\delta^{(1)}$ entries as a perturbation. Under the condition of perfect blockade and no detuning, the energy eigenstates of $ \mathbf{H}_{int}=\mathbf{A}$ are the dressed states $\frac{1}{\sqrt{2}}\left(|\bar{0}\rangle \pm |\bar{1}\rangle\right)$ with total angular momentum $J=N/2$ and $N-1$ orthogonal states with total angular momentum $J = (N/2 -1)$: $\{\frac{1}{\sqrt{2}}\left(|\bar{0}\rangle \pm |\bar{1}\rangle\right), |(\bar{1})_{\perp}\rangle\}$, where $|\bar{1}\rangle \equiv \sum_{k=1}^N \frac{\alpha_k}{\bar \alpha_N}|1_k\rangle$ with $\bar\alpha_N^2 \equiv \sum_{k=1}^N \alpha_k^2$. The eigenvalues determine the speed at which the system evolves, for $\frac{1}{\sqrt{2}}\left(|\bar{0}\rangle \pm |\bar{1}\rangle\right)$ this speed is $\pm \alpha_N$ implying a collective enhancement of $\sqrt{N}\alpha$ when the coupling strengths are homogeneous. Our system has low inhomogeneous coupling contributions as evidenced by the $\alpha_N = 0.96\sqrt{\bar{N}}\alpha$ scaling observed in our previous work , for reference an average scaling of $0.972$ is predicted from experimental parameters. The observation of $\sqrt{\bar N}$ scaling of the coupling strength is a classic signature of Rydberg blockade and $N$ participating wavefunctions, as the $|\bar{1}\rangle$ state is the only state that can evolve with that coupling strength. A state with $k$-partite entanglement consistent with the observed perfect blockade given by, $|\psi^{N}\rangle = |\bar{1}^{k}\rangle \otimes |\bar{0}^{N-k}\rangle$, will still oscillate at the same $\sqrt{\bar{N}}$ frequency, but the amplitude will be reduced to the overlap with $|\bar{1}^{N}\rangle$, $|\langle \bar{1}^N |\psi^{N}\rangle |^2 = k/N$, this is discussed further in the next section. The orthogonal singly-excited states $|(\bar{1})_{\perp}\rangle$ do not couple to the symmetric states $\{|\bar{0}\rangle,|\bar{1}\rangle\}$ under ideal conditions ($\delta_k^{(1)}=0$). This becomes clear when the Bloch picture is invoked, since the symmetric states have total angular momentum $J=N/2$ while the $|(\bar{1})_{\perp}\rangle$ states have $J=N/2-1$ and a rotation on the Bloch sphere conserves angular momentum. Inhomogeneous broadening, including differential AC Stark shifts, Doppler shifts, and finite intermediate state lifetimes, are added perturbatively with $\mathbf{\Delta}$ and provide a mechanism for coupling into the $|(\bar{1})_{\perp}\rangle$ space. This coupling should be negligible and reduce with increasing $N$ and additionally will not display the characteristic $\sqrt{N}$ enhancement. Figure shows simulated projections of $\frac{1}{\sqrt{2}}\left(|\bar{0}\rangle - |\bar{1}\rangle\right)$ along the energy eigenstates of $\mathbf{H}_{int}$ for $N=5$ atoms with our experimental parameters. \subsection{Coherence Amplitude} Since the coupling to the orthogonal subspace is negligible for our experimental parameters, the amplitude of the Ramsey fringe oscillations provide a threshold for entanglement. A thermal sample of singly excited states $|1_{\rm th}\rangle=\sum_{k=1}^{N} e^{\imath\phi_k}|1_k\rangle$, where $\phi_k$ is a random phase factor for the $k^{\mathrm{th}}$ atom, will only couple back to $|\bar{0}\rangle$ by the amount of overlap with the $|\bar{1}\rangle$ state. The projection $|\langle \bar{1}|1_{\rm th}\rangle|^2$ will average to $1/N$, therefore an oscillation with contrast above $1/N$ cannot be a thermal sample. To generate a threshold for $k$-partite entanglement we perform a numerical simulation along the lines of the analysis in . Briefly, the goal is to generate an upper bound on a measurement of $P_{\bar{1}}=|\langle \bar{1}|\psi\rangle|^2$ as a function of $P_{\bar 0}=|\langle\bar{0}|\psi\rangle|^2$ for states $|\psi\rangle$ with a maximum of $k$ entangled particles. We establish bounds in two ways. First, we do not assume Rydberg blocakde so multiple excitations are possible. This is done by creating a random $k$-partite entangled wavefunction \begin{equation} |\psi\rangle = |\psi^{(k)}_1\rangle \otimes ... |\psi^{(k)}_{m-1}\rangle \otimes |\psi^{(k_m)}_m\rangle, \end{equation} where $|\psi^{(k)}_i\rangle = \sin \left( \theta_i/2\right) |\bar{0}^{(k)}\rangle+\cos \left(\theta_i/2\right)e^{i\phi_i} |\bar{1}^{(k)}\rangle$, $\theta_i$ and $\phi_i$ are randomly generated, and $k_m = N-(m-1)k$. We extract the maximum $P_{\bar{1}}$ for a given $P_{\bar 0}$ bin obtained numerically to arrive at the thresholds shown in Fig. a) for $k=3$ particle entanglement with ensemble atom numbers $N=4-8$. Any state above the threshold must have at least $k$-partite entanglement. The black cross is an experimental data point recorded for a sample with $\bar N=8.8$ atoms, verifying the presence of entanglement. Rydberg blockade limits the Hilbert space to $n\leq 1$ excitations, which simplifies the calculation and enables an analytical bound for the $k$-partite entanglement threshold. The state in Eq. () includes kets with multiple excitations. To remove these we impose the blockade condition $P_{(n>1)}=0$ and write the state as \begin{equation} |\psi\rangle=\left(a_1|\bar{0}^{(k)}\rangle+b_1 |\bar{1}^{(k)}\rangle\right) \otimes |\bar{0}^{(N-k)}\rangle. \end{equation} Maximization of $P_{\bar{1}}$ for a given $P_{\bar 0}$ can be readily accomplished analytically to give \begin{equation} P_{\bar{1}} = \frac{k}{N}(1-P_{\bar 0}). \end{equation} Note that this agrees with the limiting case of $P_{\bar 0}=0$ from . Rearranging () to give $\frac{k}{N} \leq \frac{P_{\bar{1}}^{\mathrm{max}}}{1-P_{\bar 0}^{\mathrm{max}}}$, and given our extreme value $(P_{\bar0},P_{\bar{1}})=(0.44\pm0.02,0.46\pm0.03)$ we can show that we meet the threshold for creation of the $W$-state with $\frac{k}{N}=82\pm 6\ \section{Summary} In summary we have shown evidence for $N$ particle W-state entanglement on the basis of the following three arguments. First, the excellent agreement of the observed collectively enhanced Rabi frequency with theory reported in our previous work using the same experimental setup and procedures as are used here implies an $N$ component wavefunction. Second, the amplitude of the Ramsey-style oscillations for $\bar N=8.8$ is four times larger than the $1/\bar N$ limit expected from a thermal sample of singly excited states. Our data shows entanglement without making the assumption of perfect blockade. Third, with the assumptions of perfect blockade, entanglement percentage independent of $N$, and negligible coupling to $\{|(\bar 1)_\perp\rangle \}$, which is justified by Fig. 1, then $\frac{k}{N}=82\pm 6\ $82\pm 6\ |
1501.04086 | Title: Synthetic gauge fields stabilize a chiral spin liquid phase
Abstract: We calculate the phase diagram of the SU($N$) Hubbard model describing
fermionic alkaline earth atoms in a square optical lattice with on-average one
atom per site, using a slave-rotor mean-field approximation. We find that the
chiral spin liquid predicted for $N\ge5$ and large interactions passes through
a fractionalized state with a spinon Fermi surface as interactions are
decreased before transitioning to a weakly interacting metal. We also show that
by adding an artificial uniform magnetic field with flux per plaquette
$2\pi/N$, the chiral spin liquid becomes the ground state for all $N\ge 3$ at
large interactions, persists to weaker interactions, and its spin gap
increases, suggesting that the spin liquid physics will persist to higher
temperatures. We discuss potential methods to realize the artificial gauge
fields and detect the predicted phases.
Body: \title{Synthetic gauge fields stabilize a chiral spin liquid phase} \author{Gang Chen} \altaffiliation[Current address: ]{chggst@gmail.com, Department of Physics, University of Toronto, Toronto, Ontario M5S 1A7, Canada} \affiliation{Department of Physics, University of Colorado-Boulder, Boulder, Colorado 80309-0440, USA} \author{Kaden R.~A. Hazzard} \affiliation{Department of Physics, Rice University, Houston, Texas 77005, USA} \author{Ana Maria Rey} \affiliation{JILA and Department of Physics, University of Colorado-Boulder, NIST, Boulder, Colorado 80309-0440, USA} \affiliation{Center for Theory of Quantum Matter, University of Colorado, Boulder, Colorado 80309, USA} \author{Michael Hermele} \affiliation{Department of Physics, University of Colorado-Boulder, Boulder, Colorado 80309-0440, USA} \affiliation{Center for Theory of Quantum Matter, University of Colorado, Boulder, Colorado 80309, USA} \date{\today} \begin{abstract} We calculate the phase diagram of the SU($N$) Hubbard model describing fermionic alkaline earth atoms in a square optical lattice with on-average one atom per site, using a slave-rotor mean-field approximation. We find that the chiral spin liquid predicted for $N\ge5$ and large interactions passes through a fractionalized state with a spinon Fermi surface as interactions are decreased before transitioning to a weakly interacting metal. We also show that by adding an artificial uniform magnetic field with flux per plaquette $2\pi/N$, the chiral spin liquid becomes the ground state for all $N\ge 3$ at large interactions, persists to weaker interactions, and its spin gap increases, suggesting that the spin liquid physics will persist to higher temperatures. We discuss potential methods to realize the artificial gauge fields and detect the predicted phases. \end{abstract} \maketitle \begin{figure*} \caption{ Phase diagram, calculated with a slave-rotor mean-field approximation, as a function of spin degrees of freedom $N$ and interaction strength $U/t$ in the (a) absence and (b) presence of an artificial uniform magnetic field with flux per plaquette $\Phi=2\pi/N$, illustrated in panel (d) for $N=3$. Thin black lines are second order phase transitions, while thick black lines are first order phase transitions. The states found are the valence bond solids (VBS), chiral spin liquid (CSL), spinon Fermi surface (SFS), Fermi liquid (FL), and integer quantum Hall (IQH) states. These are described in the text and illustrated in panel (c). } \end{figure*} \textit{Introduction.}---The experimental realization of a topologically ordered phase of matter other than the fractional quantum Hall effect that occurs in two-dimensional electron gases is a major goal in both condensed matter and atomic physics. Phases with intrinsic topological order~ are of fundamental interest, as they exist outside of the standard symmetry-breaking framework for classifying phases of matter and display exotic phenomena such as fractionalized excitations and edge states that are robust to local perturbations~; in some cases these phases have been predicted to be useful for topological quantum computation~. Ultracold atomic systems are uniquely tunable and clean systems that offer a platform to realize exotic phases. However, so far, reaching the required low temperatures remains a challenge. Previous work predicted a topologically ordered chiral spin liquid (CSL) ground state in fermionic alkaline earth atoms (AEA) in a deep square optical lattice~. In this Letter we show, within a slave-rotor approximation, that by applying a synthetic gauge field to this system it is possible to enhance the parameter space where the CSL exists, to increase the corresponding spin gap, and in turn to increase the temperatures at which CSL physics manifests. In addition, without a synthetic gauge field, away from the strongly insulating limit we find a gapless quantum spin liquid with a spinon Fermi surface. Recently, experiments have trapped and cooled AEA to quantum degeneracy and loaded them in an optical lattice . Moreover, experiments have confirmed the predicted SU($N$) spin symmetry in the collisional properties of fermionic AEA . This SU($N$) symmetry generalizes the usual SU($2$) symmetry, and $N$ can be controllably varied by initial state preparation up to $2I+1$, with $I$ the nuclear spin (as large as $N=10$ for $^{87}{\rm Sr}$ with $I=9/2$). The low temperatures reached in recent experiments~\cite{taie:su6_2012,cazalilla:ultracold_2014, stellmer:annual_2014}, at which short range spin correlations should begin to develop, makes it particularly timely to study quantum magnetism in these systems. Several theory works have addressed questions related to the expected SU($N$) magnetic phases in the strongly interacting limit~\cite{assaraf:sun_1999,hermele:mott_2009,toth:sun_2010,manmana:sun_2011,corboz:sun_2011, rapp:sun_2011,nonne:sun_2011,hermele:topological_2012,bonnes:adiabatic_2012, messio:sun_2012, corboz:sun2_2012,corboz:sun3_2012,bauer:sun_2012, cai:quantum_2013,wu:pomeranchuk_2013,bluemer:mott_2013,song:mott_2013, wang:competing_2014,zhou:quantum_2014}. In parallel, other ultracold atom experiments have realized synthetic gauge fields~\cite{lin:synthetic_2009,dalibard:artificial_2011, aidelsburger:experimental_2011,struck:tunable_2012, goldman:light-induced_2013,aidelsburger:realization_2013, miyake:realizing_2013,jotzu:experimental_2014}. In these experiments, the atoms behave as if they were charged particles in external electromagnetic fields despite their neutrality. Although many schemes in principle can create the gauge field that we study in this paper, we focus on methods utilizing laser-induced tunneling~. \textit{AEA in optical lattices with synthetic gauge fields.}---AEA in a sufficiently deep optical lattice are described by an SU($N$) generalization of the usual ($N=2$) Hubbard model, \be H &=& -t\!\sum_{\langle i,j\rangle, \alpha}\! e^{ i \phi_{ij} } c^\dagger_{\alpha,i}c_{\alpha,j}^{\phantom \dagger}+\frac{U}{2}\sum_{i} (n_{i}-1)^2 \ee where $c_{\alpha,i}$ is the fermionic annihilation operator for nuclear spin state $\alpha$ at lattice site $i$, $\sum_{\langle i,j\rangle}$ indicates a sum over nearest neighbors $i$ and $j$; $\phi_{ij}=-\phi_{ji}$ is the (externally imposed) lattice gauge field. We define $n_i=\sum_\alpha c^\dagger_{\alpha,i} c_{\alpha,i}^{\phantom \alpha}$, and $t$ and $U$ are the hopping energy and on-site interaction energy, whose ratio can be tuned by modifying the optical lattice depth. In this Letter, we take the average fermion number per site to be one. The gauge field $\phi_{ij}$ depends both on the artificial electromagnetic field as well as the gauge choice. We are interested in the physics of a two-dimensional square lattice with a spatially uniform, time-independent artificial magnetic field, and use the Landau gauge where \be \phi_{ij} &=& \begin{cases} \Phi x_j \delta_{ y_j -1, y_i} & \text{if $\{i,j\}$ bond is vertical}\\ 0 & \text{otherwise}, \end{cases} \ee $x_j$ is the x coordinate of site $j$ measured in lattice units, and $\Phi$ is the flux penetrating a single square plaquette of the lattice~. We focus on the case $\Phi=2\pi/N$, because this choice of $\Phi$ is favorable for the existence of the chiral spin liquid. We note that the magnetic unit cell associated with the translational invariance of the Hamiltonian is enlarged from the one imposed by the optical lattice potential. Figure~(d) shows the system with this flux and gauge choice, and the enlarged magnetic unit cell, for $N=3$. We calculate the phase diagram and properties of this system within a slave rotor mean-field approximation , which we describe briefly. This technique is designed to match on to the previous large-$N$ solution in the large $U/t$ limit, and is well-suited for describing non-magnetic ground states in proximity to the Mott transition. First we expand the Hilbert space to include a U(1) bosonic rotor degree of freedom on each site, $\theta_j$, and new fermionic spinon degrees of freedom associated with operators $f_{\alpha,j}$, which are defined by \be c_{\alpha,j} &=& e^{-i\theta_j}f_{\alpha,j}. \ee In order to reproduce the original Hilbert space, we must impose the constraint \be L_j &=& \sum_\alpha f^\dagger_{\alpha,j}f^{\phantom\dagger}_{\alpha,j}-1 \ee that the rotor angular momentum $L_j$ is uniquely determined by the particle number. Here, $L_j$ satisfies $[\theta_j,L_j]=i$. We rewrite the Hamiltonian in terms of these new degrees of freedom, giving \be H=-t\sum_{\langle i,j \rangle, \alpha} e^{i\phi_{ij}} e^{i(\theta_i-\theta_j)}f^{\dagger}_{\alpha,i} f^{\phantom\dagger}_{\alpha,j} + \frac{U}{2} \sum_i L_i^2. \ee Although the rewritten Hamiltonian Eq.~\eqref{eq:SUN-slave-rotor} together with the constraint Eq.~\eqref{eq:ang-mom-constraint} is exactly equivalent to Eq.~\eqref{eq:SUN-Hubbard}, to make further progress we make a mean-field approximation to decouple the rotor and spinon degrees of freedom. We then obtain the coupled mean-field Hamiltonians for the rotors and the spinons, \begin{eqnarray} H_r &=& -\sum_{\langle i,j \rangle} J_{ij} e^{i\theta_i - i\theta_j} + \sum_i \frac{U}{2} L_i^2 + h_i (L_i + 1), \\ H_f &=& - \sum_{\langle i,j \rangle, \alpha} \tilde{t}_{ij} e^{i\phi_{ij}} f^{\dagger}_{\alpha,i} f^{\phantom\dagger}_{\alpha,j} - \sum_{i,\alpha} h_i f^{\dagger}_{\alpha,i} f^{\phantom\dagger}_{\alpha,i}, \end{eqnarray} where $h_i$ is a Lagrange multiplier that enforces on average the constraint Eq.~\eqref{eq:ang-mom-constraint}, $\tilde{t}_{ij} \equiv t \langle e^{i\theta_i - i\theta_j} \rangle_r$, and $J_{ij} \equiv t e^{i\phi_{ij}} \sum_{\alpha} \langle f^{\dagger}_{\alpha,i} f^{\phantom\dagger}_{\alpha,j} \rangle_f$. Here the sub-index $r$ ($f$) refers to taking the expectation value in the rotor (spinon) mean-field ground state $|\psi\rangle_r$ ($|\psi\rangle_f$). The Hamiltonians $H_r$ and $H_f$ are invariant under a U(1) gauge transformation, $f^\dagger_{\alpha,i} \rightarrow f^\dagger_{\alpha,i} e^{-i \chi_i}, \theta_i \rightarrow \theta_i + \chi_i $, and $\tilde{t}_{ij} \rightarrow \tilde{t}_{ij} e^{i \chi_i -i \chi_j}, J_{ij} \rightarrow J_{ij}e^{-i \chi_i + i \chi_j}$. We solve $H_r$ and $H_f$ self-consistently for several variational ansatz and find the ground state by optimizing the total energy $\langle \psi | H | \psi \rangle$ where $H$ is given by Eq.~\eqref{eq:SUN-slave-rotor} and $|\psi\rangle \equiv |\psi\rangle_r | \psi\rangle_f $ is the mean-field state. \textit{Results.}--- Figure~(a, b) shows the slave-rotor mean-field phase diagram as a function of $U/t$ and $N$; the top panel shows the phase diagram in the absence of a gauge field and the bottom shows the phase diagram for a gauge field with flux $\Phi=2\pi/N$. We find five phases: Fermi liquid (FL), integer quantum Hall (IQH), valence bond solids (VBS), a gapless spin liquid with a spinon Fermi surface (SFS) , and a chiral spin liquid (CSL) . Thin black lines indicate second order transitions and thick black lines indicate first order phase transitions. Generically, the role of the Hubbard $U$ interaction is to localize the atom on lattice sites. Such Mott localization is signalled in the rotor sector; when the bosonic rotor is gapped and uncondensed with $\langle e^{i\theta}\rangle = 0$, the system is in a Mott insulating state. The mean-field parameters and some key properties of the different phases are listed in Table~. As we show in the table, the rotor and the spinon may experience different, even opposite, gauge fluxes in their mean-field Hamiltonians for different phases. Since the rotor and the spinon must form a whole atom, the total gauge flux experienced by the rotor and the spinon should be equal to the synthetic gauge flux that is externally imposed on the atom. The FL phase is very similar to the usual SU(2) Fermi liquid, and its structure and instabilities are essentially those described in the absence of a lattice~. The VBS are translation-symmetry breaking phases with repeating units of SU($N$) singlets spread across multiple sites. In particular, as we plot in Figure~(c), the system is decoupled into 6-site rectangular (4-site square) clusters in the SU(3)-VBS [SU(4)-VBS] state. The SFS spin liquid state is characterized by a gapless spinon Fermi surface with a gapped bosonic rotor in the mean-field theory. Going beyond the mean-field description, we need to include the U(1) phase fluctuation of the spinon hopping $\tilde{t}_{ij}$. This is the internal gauge fluctuation ; it is dynamically generated and is unrelated to the synthetic gauge field that is imposed externally. At low energies, the SFS spin liquid is described by the spinon Fermi surface coupled by a fluctuating internal U(1) gauge field . Due to the spinon-gauge coupling, the overdamped U(1) gauge fluctuation scatters the spinons on the Fermi surface and destroys the coherence of the spinon quasi-particles. The resulting state is a non-Fermi liquid of fermionic spinons. The CSL is distinct from the SFS in that the spinons form an integer quantum Hall state in the CSL. Upon coupling to U(1) gauge fluctuations, this leads to a chiral topologically ordered phase with anyon excitations, and gapless chiral edge states that carry spin but no charge . \begin{table} \begin{tabular}{C{1.7cm}C{1.5cm}C{1.5cm}C{1.6cm}C{1.6cm}N} Phases & $\langle e^{i\theta} \rangle $ & rotor flux & spinon gap & spinon flux & \\[1.4ex] \hline FL & $\neq 0$ & 0 & 0 & 0 & \\ [1.4ex] SFS & 0 & 0 & 0 & 0 & \\ [1.4ex] CSL & 0 & $-2\pi/ N$ & $\neq 0$ & $2\pi/N$ & \\ [1.4ex] SU(3)-VBS & 0 & $-\pi$ & $\neq 0$ & $\pi$ & \\[1.4ex] SU(4)-VBS & 0 & 0 & $\neq 0$ & 0 & \\[1.4ex] \hline IQH & $\neq 0$ & 0 & $\neq 0$ & $2\pi/N $ & \\[1.4ex] CSL & 0 & 0 & $\neq 0 $ & $2\pi/N $& \\[1.4ex] SU(3)-VBS & 0 & $\pi/3$ & $\neq 0$ & $\pi$ & \\[1.4ex] SU(4)-VBS & 0 & $\pi/2$ & $\neq 0$ & 0& \\[1.4ex] \hline \end{tabular} \caption{Parameters that characterize the obtained phases. The upper five (lower four) rows describe phases in the absence (presence) of the synthetic gauge field. The rotor (spinon) flux refers to the flux that is experienced by the rotor (spinon) in the mean-field Hamiltonian $H_{r}$ ($H_f$). For the FL, SFS, IQH, and CSL states, the flux is defined for the elementary square plaquette. For SU(3)-VBS [SU(4)-VBS] state, the flux is defined through the 6-site [4-site] cluster .} \end{table} To understand the global structure of the phase diagram, it is useful to consider the two limits $U/t=0$ and $U/t\rightarrow \infty$. The FL and IQH states are simply the non-interacting ground states occurring at $U/t=0$. In the strongly interacting limit, the Hubbard model reduces to an SU($N$) Heisenberg model, and the phase diagram coincides with previous slave-fermion mean-field calculations of the Heisenberg model~: for $N=3,4$ the ground state is a VBS, while for $N\ge 5$ the ground state is a CSL. This is true both with and without a synthetic gauge field, as in the the $U/t\rightarrow \infty$ limit the physics is governed by two-site nearest neighbor superexchange, which is insensitive to the gauge flux. In the intermediate $U/t$ regime, the gauge field causes more significant differences. Without a gauge field, we find that an SFS phase intervenes between the non-interacting FL and Heisenberg-limit CSL or VBS for all $N$ except $N=4$, in which case there is a direct transition between the FL and VBS ground states. The FL-SFS transition is second order and is expected to remain continuous beyond mean-field theory , while the SFS-CSL and FL-VBS are first order phase transitions. In contrast, in the presence of the $\Phi=2\pi/N$ gauge flux, a direct second order transition occurs between the non-interacting IQH phase and the CSL phase within our mean-field theory, and the CSL exists at intermediate $U/t$ even for $N=3$ and $4$. The gauge field increases the parameter space for which the CSL occurs: in addition to persisting down to $N=3,4$, the CSL occurs for a broader range of $U/t$ values. In particular, the minimum $U/t$ for which the CSL exists decreases from about $U/t\approx5.5$ to $U/t\approx3.5$ (the exact values depend on $N$). In the CSL, both the spinon sector and the rotor sector are gapped. Figure~ illustrates the excitation gap $\Delta$'s dependence on $U/t$, $N$, and the gauge flux in the CSL where $\Delta$ is the smaller of the spin gap and the rotor gap. In the slave-rotor mean-field approximation, the spin gap is simply the band gap of the spinon spectrum, and the rotor gap is set by the Hubbard $U$ interaction and thus stays much larger than the spin gap in the Mott insulating regime except near the Mott transition. For a given $U/t$, the spin gap slightly increases when the gauge field is turned on. An even more favorable effect of the gauge field for the spin gap occurs because the CSL persists to lower $U/t$. Since $\Delta$ increases as $U/t$ decreases, the gauge field increases the maximum $\Delta$ by about a factor of 1.5. Because $\Delta$ sets the temperature to which the CSL's characteristics remain, we therefore expect the gauge field to increase the temperature range over which the CSL behavior is accessible. \textit{Gauge field implementation.}---Many proposals to implement artificial gauge fields exist. Here we suggest one scheme, which uses Raman-induced tunneling in deep lattices subject to a uniform potential gradient~\cite{aidelsburger:realization_2013, miyake:realizing_2013}. A Raman process is on resonant with the energy splitting between adjacent lattice sites, and the atoms acquire a phase kick each time they hop, imprinting the phase $\phi_{ij}$ in Eq.~\eqref{eq:SUN-Hubbard}. This scheme is natural for our current considerations, since it utilizes the optical lattice and generates the Hamiltonian Eq.~\eqref{eq:SUN-Hubbard} with strong gauge fluxes. Gauge fields have been recently demonstrated in bosonic alkali atoms using this technique~\cite{aidelsburger:realization_2013, miyake:realizing_2013}, although we note that these experiments have observed unexplained heating, which could be problematic for realizing low temperature phases. We also mention the alternative scheme proposed in Ref.~ that seems natural for the present work with AEA: rather than using Raman lasers, one traps the ${}^1$S$_0$ ground ($g$) and ${}^3$P$_0$ excited ($e$) states in, for example, a checkboard pattern in an optical lattice by using an appropriate, ``anti-magic," wavelength~. The $e$ state has a $\sim 100$s natural lifetime, and is therefore stable on the timescale of the system. Because a single laser can directly drive tunneling of a $g$ atom to an $e$ atom at an adjacent lattice site while imprinting a phase $\phi_{ij}$, one avoids the complexity of driving Raman processes. However, when this proposal is implemented in the context of interacting quantum phases additional considerations arise that were not accounted for in the prior analysis. First, two $e$-state atoms on the same site can inelastically collide and be lost from the trap. We have found that this problem can be largely mitigated when using a checkerboard $g$-$e$ pattern . Second, the interactions are inhomogeneous, being different for the sites occupied by $g$ atoms and $e$ atoms. This issue can modify the discussed phase diagram. Third, the flux generated in the simplest implementation of this proposal is staggered and thus requires rectification techniques to make it homogeneous. \textit{Preparation and detection.}---Reaching the temperature regimes to observe the phase diagram Figure~ is challenging. However, the expected advantage of the SU($N$) symmetry for cooling~ together with the less stringent temperature requirements to observe CSL phases in the presence of the synthetic gauge field might help achieve the required conditions. Other potentially favorable aspects of the gauge field are the absence of an intermediate SFS phase and that all transitions are second order in the mean-field analysis. Consequently, adiabatically going from weak to strong interactions may be easier than in the absence of the gauge field. On the other hand, the gauge field itself introduces further constraints such as the requirement to use a deep lattice potential and a complex band structure even in the weakly interacting regime. Consequently, determining optimal preparation is beyond the scope of this work. To conclude, we briefly outline methods to detect the CSL and SFS. Although it is premature to analyze protocols in detail, as these will depend substantially on the specific experimental implementation, it is useful to describe the basic ingredients that would be required. To detect the CSL Ref.~ suggests methods to probe two characteristic properties of topological phases: looking for topologically protected, chiral edge currents and introducing a weak attractive optical potential that is localized to a few lattice sites, which should bind the anyonic quasiparticles. Braiding or interfering these quasiparticles can manifest their anyonic nature. To detect the SFS state, one can perform spin-dependent Bragg spectroscopy to detect the 2-spinon continuum in the dynamic spin structure factor; the most basic signature of the exotic nature of this phase is the lack of order and existence of gapless excitations. More details of the state and its excitations could be revealed by considering more structure of the spectrum, similar to that considered in Refs.~\onlinecite{PhysRevB.65.165113,PhysRevB.87.045119}. \emph{Acknowledgements.} This work was supported by the AFOSR, AFOSR-MURI, NSF JILA-PFC-1125844, NSF-PIF-1211914, NIST and ARO (AMR) and the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award number DE-FG02-10ER46686 (G.C. and M.H.). G.C. acknowledges NSF grant no.~PHY11-25915 for supporting the visitor program at the Kavli institute for theoretical physics during the workshop ``Frustrated Magnetism and Quantum Spin Liquids'' in 2012, when and where part of the current work was done. \bibliography{CSL-gauge} \vspace{1cm} \appendix \begin{center} {\bf SUPPLEMENTARY MATERIAL} \end{center} \section{Slave rotor mean-field theory: no translational symmetry breaking} \emph{Synthetic gauge field case.}---Here we give a detailed description of the slave rotor mean-field theory in the presence of the synthetic gauge flux. To study the energetics as well as the phase transition from the IQH to the CSL in Figure~(b), we first choose the variational ansatz for the IQH and the CSL such that $\tilde{t}_{ij} = \tilde{t}, J_{ij} = J $. Moreover, we assume an uniform Lagrange multiplier such that $h_i \equiv h$. This is equivalent to replacing the local constraint at every site with a global constraint. This simplification is justified by the fact that both the IQH and the CSL preserve the lattice translation symmetry. The rotor and spinon mean-field Hamiltonians are then given by \begin{eqnarray} H_r &=& -J \sum_{\langle i,j \rangle} \Phi^{\dagger}_i \Phi^{\phantom\dagger}_j + \sum_i \frac{U}{2} L_i^2 + h (L_i + 1), \\ H_f &=& - \tilde{t} \sum_{\langle i,j \rangle,\alpha } e^{i\phi_{ij}} f^{\dagger}_{\alpha,i} f^{\phantom\dagger}_{\alpha,j} - h \sum_{i,\alpha} f^{\dagger}_{\alpha,i} f^{\phantom\dagger}_{\alpha,i}, \end{eqnarray} where we have replaced the rotor variable $e^{i\theta_i}$ by a uni-modular operator $\Phi_i$ such that $| \Phi_i | \equiv 1 $. Since the operator $L_i = \sum_{\alpha} f^\dagger_{\alpha,i} f^{\phantom\dagger}_{\alpha,i} -1$, $h$ is then thought as a chemical potential. Because $\langle\sum_i L_i \rangle = 0$, $h$ must vanish for the mean-field solutions. With a $2\pi/N$ flux per square plaquette and one fermion per site, the Hamiltonian $H_f$ gives a spinon band structure with $N$ bands. Only the lowest band is filled (for each species) and separated from the others by a gap. Moreover, the lowest spinon band has Chern number $C=1$ for each fermion flavor $\alpha$. As we shown in Table~, whether the system is in the IQH or the CSL is determined by the behavior of the rotor sector. To solve the rotor Hamiltonian $H_r$, we implement a coherent state path integral formalism in imaginary time. We integrate out the conjugate variable $L_i$ and obtain the partition function that is written as a functional integration over the $\Phi$ variable, \begin{equation} {\mathcal Z} \simeq \int {\mathcal D} \Phi^{\dagger} {\mathcal D} \Phi {\mathcal D} \lambda e^{- S - \int_\tau \lambda_i (| \Phi_i |^2 -1 )}. \end{equation} Here $\lambda_i$ is the Lagrange multiplier introduced to the unimodular condition for the rotor variable at every lattice site. The effective action is given by \begin{equation} S = \int_{d \tau} \frac{1}{2U} \sum_{{\bf k} \in \text{BZ} } |\partial_{\tau } \Phi_{\bf k}|^2 - 2J \sum_{{\bf k} \in \text{BZ} }( \cos k_x + \cos k_y) |\Phi_{\bf k}|^2, \end{equation} where ``BZ'' refers to the Brioullin zone of the square optical lattice and we have set the lattice constant to unity. In a standard spherical approximation for a mean-field (or saddle point) analysis, we assume a uniform Lagrange multiplier $\lambda_i \equiv \lambda $. We integrate out the variable $\Phi$ and obtain the saddle point equation for $\lambda$, \begin{equation} \frac{1}{N_s}\sum_{{\bf k}\in \text{BZ}} \frac{U}{\omega_{\bf k}} = 1, \end{equation} where $\omega_{\bf k} = [2U(\lambda - 2J (\cos k_x + \cos k_y) )]^{1/2}$ is the band dispersion of the rotor and $N_s$ is the number of lattice sites. We solve the saddle point equation Eq.~\eqref{eqsadd} self-consistently with the spinon mean-field Hamiltonian $H_f$. When the rotor band touches zero energy, the rotor is condensed and the internal U(1) gauge field picks up a mass due to the Anderson-Higgs mechanism. The rotor and the spinon are then bound together and form a fermionic atom. The resulting phase is the IQH. When the rotor band is gapped and the rotor is not condensed, the internal U(1) gauge field is gapped out by the Chern-Simons term and the resulting phase is the CSL. In the IQH, the system has $N$ chiral edge modes that transport spin quantum numbers as well as atoms. The CSL, however, is a Mott insulating state. The atoms are localized by the interaction in the CSL. The chiral edge states in the CSL only carry spin quantum number and cannot transport charge. The effect of the phase transition from the IQH to the CSL on the edge states is to gap out the mode that transports atoms. \emph{No synthetic gauge field case.}---With no gauge flux the Hamiltonian is modified by putting $\phi_{ij} \rightarrow 0$. For the FL and the SFS, we choose the spinon mean-field ansatz such that $\tilde{t}_{ij} \equiv \tilde{t}, J_{ij} \equiv J$ and $h_i \equiv h$. The spinons partially fill the bands and give rise to the spinon Fermi surface. Again, whether the system is in the FL or the SFS is determined by the rotor sector. Since the FL and the SFS only occur for the model without the synthetic gauge flux, the rotor sector Hamiltonian is identical to Eq.~\eqref{eqhr}. When the rotor is condensed, the system falls into the FL. When the rotor is gapped, the system is in the SFS whose low energy property is described by the spinon Fermi surface coupled with a fluctuating U(1) gauge field. \section{Variational ansatz: VBS states} As we described in the main text, VBS states become favorable in the strongly interacting limit for the SU(3) and SU(4) models. As expected, a similar conclusion was found in the previous slave-fermion study, {\it i.e.} the ground state of the SU(3) [SU(4)] Heisenberg model -- the $U/t\rightarrow \infty$ limit of the Hubbard model -- on the square lattice favored the VBS state shown in Figure~ [Figure~]. In our slave rotor mean-field calculation, we have chosen $H_f$ such that the spinons have the same hopping and feel the same mean-field gauge fluxes as those ones shown in Figure~ and Figure~, while the rotor sector has the same cluster structure as the spinon sector but experiences a different flux (see Table.~). |
1501.04091 | Title: A Hierarchy of Linear Threshold Models for the Spread of Political
Revolutions on Social Networks
Abstract: We study a linear threshold agent-based model (ABM) for the spread of
political revolutions on social networks using empirical network data. We
propose new techniques for building a hierarchy of simplified ordinary
differential equation (ODE) based models that aim to capture essential features
of the ABM, including effects of the actual networks, and give insight in the
parameter regime transitions of the ABM. We relate the ABM and the hierarchy of
models to a population-level compartmental ODE model that we proposed
previously for the spread of political revolutions [1], which is shown to be
mathematically consistent with the proposed ABM and provides a way to analyze
the global behaviour of the ABM. This consistency with the linear threshold ABM
also provides further justification a posteriori for the compartmental model of
[1]. Extending concepts from epidemiological modelling, we define a basic
reproduction number $R_0$ for the linear threshold ABM and apply it to predict
ABM behaviour on empirical networks. In small-scale numerical tests we
investigate experimentally the differences in spreading behaviour that occur
under the linear threshold ABM model when applied to some empirical online and
offline social networks, searching for quantitative evidence that political
revolutions may be facilitated by the modern online social networks of social
media.
Body: \maketitle \begin{abstract} We study a linear threshold agent-based model (ABM) for the spread of political revolutions on social networks using empirical network data. We propose new techniques for building a hierarchy of simplified ordinary differential equation (ODE) based models that aim to capture essential features of the ABM, including effects of the actual networks, and give insight in the parameter regime transitions of the ABM. We relate the ABM and the hierarchy of models to a population-level compartmental ODE model that we proposed previously for the spread of political revolutions , which is shown to be mathematically consistent with the proposed ABM and provides a way to analyze the global behaviour of the ABM. This consistency with the linear threshold ABM also provides further justification a posteriori for the compartmental model of . Extending concepts from epidemiological modelling, we define a basic reproduction number $R_0$ for the linear threshold ABM and apply it to predict ABM behaviour on empirical networks. In small-scale numerical tests we investigate experimentally the differences in spreading behaviour that occur under the linear threshold ABM model when applied to some empirical online and offline social networks, searching for quantitative evidence that political revolutions may be facilitated by the modern online social networks of social media. \end{abstract} \section{Introduction} In this paper, we propose a linear threshold agent-based model (ABM) for the spread of political revolutions on social networks in dictatorial regimes. In our ABM, nodes of the network can be in two states: active, i.e., participating in the revolution, or inactive. Transitions from the inactive to the active state are governed by a growth process that uses a traditional linear threshold mechanism: an individual $v$ may change from the inactive to the active state if the fraction of neighbours in the active state exceeds the linear threshold $\theta$. Linear threshold models have been studied before in the context of influence maximization , enforcement of unpopular norms , and general ``complex contagion'' processes . The development and analysis of linear threshold processes has been highly influenced by more established epidemiology and rumour spreading models, with one key difference: whereas a contact with a single ``infected'' individual is enough to spread a contagion in most epidemiology or rumour spreading applications, i.e. in ``simple contagion'' processes, multiple contacts with infected individuals are generally necessary to spread a contagion in a linear threshold process. Indeed, epidemiological and rumour spreading models can be considered to be a special case of linear threshold models where the threshold parameter is chosen sufficiently small such that one infected neighbour enables propagation. Most of the linear threshold models that have been studied in the context of social spreading or contagion processes consider evolution in discrete time. In contrast, our ABM is chosen to evolve in continuous time since that facilitates the comparison of the ABM with simplified population-level models, which is a major goal of this paper. The main justification for our linear threshold modelling approach in the context of political revolutions is that the decision to join a political revolution can be assumed to be a collective action problem : if individuals act unilaterally they are subject to retaliation by the regime, whereas if they act collectively then the regime loses the ability to punish due to a lack of resources. The linear threshold modelling approach is consistent with the collective action principle, since an individual transitions from an inactive to an active state only after the number of neighbours in the active state has been observed to reach the critical fraction $\theta$, and the individual deems the revolution of sufficient size to consider participation. In order to incorporate communication network structure explicitly, we represent interactions between individuals $v$ by edges $e$ in an undirected graph $G=G(V,E)$. This contrasts with previous attempts to model political revolutions using linear threshold models that assume homogeneous mixing , i.e. that assume that $G=G(V,E)$ is the complete graph. The goals and contributions of this paper are as follows. We start by presenting the linear threshold ABM for the spread of political revolutions on social networks, and then derive a hierarchy of simplified ordinary differential-equation (ODE) models of varying degree of sophistication that characterize the solutions of the linear threshold ABM, see Table~. Although the linear threshold ABM allows us to explicitly model the effects of the communication network on the dynamics of the political revolution it models, the inherent complexity of modelling the state of all individuals simultaneously makes this approach difficult to analyze and expensive to simulate. For these reasons there is significant interest in simplified aggregate or population-level dynamical models, which can provide more cost-effective dynamical simulations, and can give insight in the qualitative dynamics of the ABM and its parameter regimes, through dynamical analysis of the simplified model. It is a significant challenge to incorporate actual network structure in the ODE models, and this is often important since details of the network structure may determine the qualitative behaviour of ABM solutions as the ABM parameters are changed. In this paper, we present two new effective ways to incorporate network structure into the one-compartmental ODE that approximates the dynamical evolution of the expected fraction of the population that participates in the revolution in the linear threshold ABM model. These approaches make use of the degree distribution of the graph or samplings of the network, and we call the resulting population-level ODE models the \emph{binomial visibility function} (BVF) and the \emph{empirical visibility function} (EVF) models. The EVF model is significantly less expensive than the BVF, but we demonstrate its equivalence with the BVF in the limit of large network and sample size. Extensive numerical tests show that these ODE models perform remarkably well to predict qualitative behaviour of the ABM; in fact, they often perform as well as or better than a previously specified higher-order model that has many compartments and is much more expensive computationally , and that we refer to as the \emph{degree approximation} (DA) model. \begin{table}[hp] \centering \caption{Summary of hierarchy of models from highest (network-level) to lowest (population-level) complexity.} \begin{tabular}{llllc} \hline Model & Abbreviation & Section & Equation(s) & Reference(s) \\ \hline Linear Threshold Agent-Based Model & ABM & & \eqref{eq:0to1} and \eqref{eq:1to0} & \\ Degree Approximation Model & DA & & \eqref{eq:ApproxDA} & \\ Binomial Visibility Function Model & BVF & & \eqref{eq:Approx} and \eqref{eq:BVF} & \\ Empirical Visibility Function Model & EVF & & \eqref{eq:Approx} and \eqref{eq:EVF} & \\ Step Visibility Function Model & SVF & & \eqref{eq:SVF} & \\ \hline \end{tabular} \end{table} Similar to the population-level one-compartmental ODE model that we proposed in for the spread of a political revolution, the linear threshold ABM proposed in this paper applies to dictatorial regimes that employ censorship and police repression. We make the simplifying assumption that the population is uniform in its dislike of the regime (such that all nodes can potentially become active in the revolution), and we assume that the regime censors the communication between individuals except for the links in the social network. A decay process models police repression that is limited by the regime's finite police capacity. The model features parameters that describe timescales of growth and decay, a linear threshold parameter for the growth process, and a parameter for the finite police capacity. The dynamics of our ABM is governed by stochastic transitions in continuous time, because this allows us to relate the ABM mathematically to simplifying time-dependent ODE models in limits of expectation over the aggregate population. These simplified ODE-based models aim to capture essential features of the linear threshold ABM at the population level, including effects of the actual networks, and give insight in the parameter regime transitions of the ABM. We also relate the ABM to the population-level one-compartmental ODE model for the spread of political revolutions that was proposed in and that indirectly takes network structure into account by a single parameter that describes from which fraction of participants the growth term is switched on discontinuously. As such, it is a precursor of the BVF/EVF models introduced in this paper. In this single parameter was called the visibility parameter of the revolution, with the interpretation that people will only join the revolution when the revolution is of sufficiently large size to be visible to the population, while repressive regimes attempt to make unrest invisible through censorship. We call the model from the \emph{step-visibility function} (SVF) model in this paper. This ODE model was justified in by a simplified network model with significant further assumptions; in this paper, we show that our ABM is mathematically consistent with the model from and we test the assumptions from with the BVF/EVF models and the ABM model on real networks, corroborating a posteriori the assumptions from , and showing that the model from reproduces the qualitative behaviour of the ABM and the spread of a revolution under a linear threshold model. This provides further justification for the ODE model of , and at the same time provides inexpensive ways to predict the parameter regime behaviour of the ABM presented in this paper using simple compartmental models like the SVF and the new BVF/EVF models. We emphasize that the broad applicability of the linear threshold process implies that the framework, methods, and results presented in this paper are relevant to the study of many related social spreading processes. The final contribution of this paper relates to an interesting application of our linear threshold ABM. We investigate experimentally the differences in spreading behaviour that occur under the linear threshold model when applied to some empirical online and offline social networks, searching for quantitative evidence that political revolutions may be facilitated by the online social networks of social media. Indeed, it is often assumed that the connectivity of modern online social networks has greatly facilitated the spread of political revolutions in the past decade, e.g., in the Arab Spring revolutions of 2011 \cite{eltantawy2011arab,khondker2011role,tufekci2012social, KhamisVaughn11, LangDeSterck14}, while the traditional offline social networks of the pre-Internet era (using in-person physical contact, or mail or phone interaction for safe communication) had a different connectivity structure that was often severely restricted by regime censorship. As a starting point to investigate differences in propagation properties that may arise between online and offline social networks within the linear threshold propagation model, we investigate propagation on two empirical networks of modest size using our linear threshold ABM: we consider a small Facebook network as a representative of an online social network, and a small physical contact network between individuals as a representative of an offline social network. These two networks are also used for all validation and comparison simulations for the ABM, DA, BVF/EVF and SVF models throughout the paper. We also extend the concept of basic reproduction number $R_0$ from epidemiological modelling to characterize the potential of networks to spread the revolution. The basic reproduction number for our ABM with a linear threshold process is easy to compute, and we show that it gives useful predictions. Our results indicate that the offline social network indeed is less conducive to spreading the revolution than the online social network: it has a smaller basic reproduction number, and in simulations a larger initial population of revolutionaries is required to spread the revolution. This provides some initial quantitative evidence that the spread of revolutions under a linear threshold process may occur more easily on modern online social networks than on traditional offline networks, but we also comment on limitations in our approach and further investigations that are required to address this intriguing but complex question more comprehensively. The rest of this paper is structured as follows. Section describes the Facebook and physical contact empirical networks on which we test our ABM and approximations. Section specifies the linear threshold ABM in detail and briefly reviews the SVF model presented in . Section then establishes the consistency between these two approaches through mathematical derivation and numerical simulation on the networks described in Section , i.e., we show that the SVF model can be considered to approximate the aggregate behaviour of the ABM. Section describes the BVF and EVF methods to improve the SVF model by changing the functional form of the visibility function, explicitly incorporating the structure of the network without increasing the dimensionality of the SVF model. Section confirms the usefulness of the BVF and EVF models by showing that they perform no worse than the higher order degree approximation model of , while being much less expensive to evaluate. Section extends the concept of basic reproduction number from epidemiology to the linear threshold process specified in Section and illustrates how it can be interpreted in terms of the BVF/EVF. Finally, Section explores how our methods can be used to study the differences in spreading behaviour of political revolutions on online and offline social networks under the linear threshold mechanism. \section{Network Data} In this section we describe the two empirical social networks that will be used in this paper to validate and compare the ABM, DA, BVF/EVF and SVF models for the spread of political revolutions on social networks under the linear threshold model. The first network we consider is the physical contact network presented in . It was constructed by distributing wireless sensors to students, teachers, and staff at a U.S. high school during a one day period. When two wireless sensors are in proximity of one another, i.e. when they are less than approximately 3m apart, they register an interaction with a temporal resolution of 20s. Therefore, the communication network we extract from this data is referred to as the \emph{physical contact network $G_P$}. The second network we consider is the Facebook subnetwork presented in . It was constructed by combining the \emph{ego-networks} of an individual $v\in V$ is a subnetwork of the overall network composed of individual $v$ (called the \emph{ego-node}), the neighbours of individual $v$, and all of the connections between these individuals.} of ten individuals and then taking the largest connected component of the resulting network. We refer to this network as the \emph{Facebook social network} $G_F$. For details on the network extraction protocol for $G_P$ we refer to Appendix . To facilitate the comparison of experimental results, the minimal contact duration to register an edge in $G_P$ is chosen such that $G_P$ and $G_F$ have approximately the same average degree, see Appendix . The physical contact network $G_P$ has $N=776$ nodes with sparsity $S=0.06$, and is visualized in Fig.~. The Facebook network $G_F$ has $N=3,963$ nodes with sparsity $S=0.01$, and is visualized in Fig.~. The cumulative degree distributions^k \rho_j$. The cumulative degree distribution is displayed in place of the degree distribution because the degree distribution is subject to significantly more noise than the cumulative degree distribution. This is a common approach for empirical networks.} are displayed in Fig.~. The distributions of local clustering coefficients are displayed in Fig.~. \begin{table}[h] \centering \caption{Summary of common network measures for physical contact and Facebook social networks.} \begin{tabular}{lcc} \hline Quantity & Physical Contact Network & Facebook Social Network \\ \hline Number of Nodes ($N$) & 776 & 3,963 \\ Average Degree ($\mu_k\pm\sigma_k$) & $44.8\pm18.2$ & $44.5\pm52.4$\\ Minimum/Maximum Degree ($k_{\min}/k_{\max}$) & 1/109 & 2/1034\\ Sparsity ($S$) & 0.06 & 0.01\\ Diameter ($D$) & 8 & 8\\ Average Path Length($\mu_l\pm\sigma_l$) & $2.37\pm0.82$ & $3.77\pm1.29$\\ Acerage Local Clust. Coeff. ($\mu_c\pm\sigma_c$) & $0.29\pm0.12$ & $0.62\pm0.20$\\ \hline \end{tabular} \end{table} It is interesting to consider how the physical contact and Facebook networks differ on some common network measures, see Table . Comparing the physical contact and Facebook social networks along these admittedly limited measures nevertheless highlights substantial structural differences between these networks. The physical contact network appears more homogeneous than the Facebook social network, in the sense that the Facebook social network can be visually grouped into distinct communities, whereas the physical contact network cannot (Fig.~). The network measures presented in Figs.~- and Table~ are generally supportive of this observation. For example, the cumulative degree distributions displayed in Fig.~ show that the physical contact network has an approximately normal distribution with a relatively small standard deviation, and hence, a thin tail. In contrast, the Facebook social network has a much broader distribution (possibly scale-free or exponential, although this cannot be determined conclusively with such a small network size) with a relatively large standard deviation, and hence, a fat tail. Similarly, it is known that online social networks often have larger clustering coefficients than offline social networks. In Section we use the Facebook and physical contact networks as representatives of online and offline social networks, respectively, to investigate differences in propagation properties that may arise between online and offline social networks within the linear threshold propagation model. \section{Agent-Based Model} \subsection{Specification of the Agent-Based Model} Consider a population of individuals who are represented by nodes $V = \{v_i\}_{i=1}^N$, and whose interactions are represented by edges $E = \{e_i\}_{i=1}^M$, of the graph $G=G(V,E)$ with degree distribution $\rho_k$. An individual $v\in V$ at time $t$ can be in one of two states, i.e. $s_v(t) = 0$ (\emph{inactive} in the revolution) or $s_v(t) = 1$ (\emph{active} in the revolution). We assume that the network $G$ is static so that the dynamics of the ABM can be fully specified by providing rules for the transition of individuals from an inactive state to an active state and vice versa, together with parameter values and an initial condition. For the moment, we set aside the issue of the choice of parameters and initial conditions and restrict the remainder of this section to the specification and justification of the transition rules. \subsubsection{Growth Process: Inactive ($s_v=0$) to Active ($s_v=1$)} To specify the growth process of the ABM, we use the standard linear threshold model that has been used before in the context of social spreading or contagion processes such as opinion formation, technology adoption, marketing, rioting, and political movements . As mentioned in Section , the choice of individuals to join a revolution is a collective action problem : individuals are averse to unilateral action against the regime for fear of severe retaliation, but are willing to take action collectively in the belief that the regime will lack sufficient resources to punish the entire collective. Thus, an individual $v\in V$ will decide to join the revolution if he or she believes that it has grown sufficiently large to reduce the risk of retaliation from the regime to an acceptable level. It is reasonable, therefore, to assume that individuals will decide to join the revolution only after a large enough fraction of their neighbours in the social network have done so. This behaviour is captured by the following transition rule: if individual $v\in V$ has $k$ neighbours $\{w_j\}_{j=1}^k$, if $s_v(t) = 0$, and if \begin{equation} \sum_{j=1}^k s_{w_j}(t) \geq \theta_v k, \end{equation} i.e., if $v$ is inactive at time $t$ and has at least a fraction $\theta_v$ of its neighbours that are active, then node $v$ transitions from state 0 to state 1 at time $t' = t + \xi_1$, where $\xi_1>0$ is the first arrival time of a Poisson process with rate $c_1$. We say that nodes that satisfy \eq{0to1} can ``see'' the revolution, i.e. the revolution is visible to them. Alternatively, we say that these nodes are ``considering joining the revolution''. Parameter $c_1$ determines the timescale of the growth process. While most of the linear threshold models that have been studied in the context of social spreading or contagion processes consider evolution in discrete time by choosing $\xi_1$ constant, we have specified $\xi_1$ so that \eq{0to1} evolves in continuous time. Specifying $\xi_1$ in this way is often considered in biology and epidemiology , since it (a) facilitates the comparison of the ABM with population-level models (as the SVF model described in Section ), and (b) eliminates the problem of choosing a suitable time-step for iterating the discrete-time process. Furthermore, we note that by choosing $\xi_1$ to be the first arrival time of a Poisson process, we are assuming that the decision making process is a Markov, or memoryless, process. Specifically, conditioned on the state of the system at time $t$, we assume that the likelihood of an individual joining the revolution in the time interval $[t, t+\Dt]$ is independent of the state of the system at any time $\tilde{t}<t$. In words, the future is independent of the past, given the present. \subsubsection{Decay Process: Active ($s_v=1$) to Inactive ($s_v=0$)} We assume that once an individual $v\in V$ has become active the regime will attempt to arrest or disperse him or her, thus returning $v$ to an inactive state. As in the previous section, we assume that this is a memoryless process. We further assume that the regime can only arrest or disperse protesters so long as the total fraction of active protesters remains less than the regime's finite police capacity $\beta\in(0,1)$. The transition rule can then be characterized by the following: if $s_v(t) = 1 $ and \begin{equation} \frac{1}{N}\sum_{w\in V} s_w(t) < \beta, \end{equation} i.e., if $v$ is active at time $t$ and the fraction of active individuals is less than the regime's police capacity $\beta$, then the node $v$ transitions from state 1 to state 0 at time $t'=t+\xi_2$, where $\xi_2>0$ is the first arrival time of a Poisson process with rate $c_2$, which determines the timescale of the decay process. This fully specifies the evolution of the ABM, which we simulate using Gillespie's algorithm, see Appendix . \subsection{Brief Review of the SVF Population-Level Compartmental Model of } Before deriving a population-level ODE for the linear threshold ABM and establishing the consistency of the SVF model from with the ABM in the following section, it is useful to first review some basic properties of the SVF model. For example, we will show later that the parameter regimes of the SVF are predictive of the parameter regime transitions in the ABM. The SVF model is given by the equation \begin{equation} \dot{r} = \underbrace{c_1 \spc (1-r) \spc v_s(r;\alpha)}_{g(r)} - \underbrace{c_2 \spc r \spc p(r;\beta)}_{d(r)}, \end{equation} with parameters $\alpha,\beta\in(0,1)$ and $c_1,c_2>0$. The growth and decay of the fraction of protesters $r(t)$ are modelled by the \emph{growth} and \emph{decay} functions $g,d:[0,1]\rightarrow\mathbb{R}^+$, respectively. Given their roles in determining the rate of growth and decay, the parameters $c_1$ and $c_2$ are called the \emph{protesters' enthusiasm} and \emph{police efficiency} parameters in . The growth and decay in the number of protesters are modulated by the \emph{visibility} and \emph{policing} terms that are taken to be step functions $v_s(r;\alpha) = \mathbb{I}_{\{r > 1-\alpha\}}$ and $p(r;\beta) = \mathbb{I}_{\{r<\beta\}}$, where the indicator function is defined as $$ \mathbb{I}_{\{X\}} = \left\{\begin{array}{ll} 1 & \mbox{if $X$ is true} \\ 0 & \mbox{otherwise} \end{array}\right.. $$ These functions are parametrized by \emph{visibility} and \emph{police capacity} parameters $\alpha$ and $\beta$, respectively. In words, the fraction of protesters can grow only when the protest movement is sufficiently large to be ``visible'' to the general population, i.e. when protests are large enough that their existence cannot be masked by the regime through censorship. Conversely, the fraction of protesters can decay only so long as the fraction of protesters is less than the regime's finite capacity to disperse or arrest protesters. This is consistent with the idea that individuals who consider joining the revolution are faced with a collective action problem : if they choose to act unilaterally then they face severe retaliation from the regime, so they wait to act until the revolution becomes visible and if they act collectively then the regime loses the ability to punish due to a lack of resources. Depending on the values of the parameters $\alpha$, $\beta$, $c_1$, and $c_2$, the dynamics of the SVF fall into one of four main regimes, see Fig.~. The equilibrium of \emph{total state control} $r=0$ and the equilibrium of the \emph{realized revolution} $r=1$ are present in all parameter regimes and are always locally asymptotically stable. Different parameter regimes are distinguished by the behaviour of \eq{SVF} when $r$ lies between $1-\alpha$ and $\beta$. Region II ($\beta<1-\alpha$) is characterized by an interval of stable equilibria $r\in(\beta,1-\alpha)$ bounded by two unstable equilibria $r\in\{\beta,1-\alpha\}$. Since parameters in Region II are characterized by a relatively weak regime (low $\beta$) and relatively small visibility (low $\alpha$), Region II is interpreted in as the \emph{failed state} parameter region. In contrast, Regions III0, IIIe, and III1 ($1-\alpha<\beta$) are characterized by a relatively strong regime (large $\beta$) and relatively large visibility (large $\alpha$) and differ in the value of the quantity $c^* = c_1/(c_1+c_2)$, i.e. in the ratio of protester's enthusiam to the sum of protester's enthusiasm and police efficiency. In Region III0, i.e. when $c^*\leq 1-\alpha <\beta$, the interval $(1-\alpha, \beta)$ lies in the basin of attraction of the equilibrium of total state control $r=0$. Analogously, in Region III1, i.e. when $1-\alpha <\beta\leq c^*$, the interval $(1-\alpha, \beta)$ lies in the basin of attraction of the equilibrium of the realized revolution $r=1$. Finally, in Region IIIe, i.e. when $1-\alpha < c^* < \beta$, the interval $(1-\alpha,\beta)$ becomes the basin of attraction of a new equilibrium $r=c^*$, which is called the \emph{equilibrium of civil unrest} in . Because of the contribution of the interval $(1-\alpha, \beta)$ to the stability of the equilibrium of total state control, of the realized revolution, or of civil unrest, Regions III0, III1, and IIIe are called the \emph{stable police state}, \emph{unstable police state}, and \emph{meta-stable police state}, respectively. For a full description, interpretation, and analysis of the SVF model we refer the reader to . \subsection{Population-Level ODE for ABM and Relation to SVF Model} \subsubsection{Mathematical Analysis} To show consistency of the ABM proposed in Section with the SVF model of we begin by defining $r_a(t) = r_a(t|t_0)$ to be the fraction of nodes (in the ABM model) that are expected to be in state 1, i.e. the expected fraction of active nodes at time $t$ conditioned on information at time $t_0$: $$ r_a(t) = \frac{1}{N}\sum_{i=1}^N \ex[ s_{v_i}(t) | t_0]. $$ The fraction of nodes that are expected to be in state 0 (inactive) at time $t$, conditioned on information at time $t_0$, is then $1-r_a(t|t_0)$. We now write the change in $r_a$ from time $t_0$ to time $t$ as \begin{equation} \Delta r_a(t) = r_a(t) - r_a(t_0) = g(t|t_0) - d(t|t_0), \end{equation} where $\Delta t = t-t_0$, and the expected growth and decay of the fraction of active nodes are modelled by the non-negative growth and decay functions $g(t) = g(t|t_0)$, and $d(t) = d(t|t_0)$, respectively. In order to obtain the one-compartment model we will need to approximate the quantities $g(t)$ and $d(t)$ in terms of $r_a(t_0)$, $\{\theta_{v_i}\}_{i=1}^N$, and $\beta$. For notational convenience we begin by considering the case where $\forall v\in V:\theta_v= \theta$. The pool of individuals that can go from active to inactive at time $t_0$ is $r_a(t_0)$. Active nodes become inactive at the first arrival time of a Poisson process with rate $c_2$, provided that the fraction of active nodes does not exceed the regime's police capacity, i.e. $r_a(t_0)<\beta$. Thus, \begin{equation} d(t|t_0) = [c_2\spc r_a(t_0) \spc\Delta t + o(\Delta t) ] \spc p(r_a;\beta) = c_2 \spc r_a(t_0) \spc p(r_a(t_0); \beta) \spc \Delta t+ o(\Delta t), \end{equation} where we take $p(r;\beta) = \mathbb{I}_{\{r<\beta\}}$, as in \eq{SVF}. Next, let \begin{equation} \nu(r_a(t_0);\theta) = \frac{1}{N} \left[\sum_{v\in V} \mathbb{I}_{\left\{\sum_{j=1}^{k_v} s_{w_{v,j}}(t_0) \geq \theta k_v \right\}}\right], \end{equation} where $\br{w_{v,j}}{j=1}{k_v}$ denotes the neighbours of individual $v\in V$. In words, we let $\nu(r_a(t_0);\theta)$ be the fraction of the total population at time $t_0$ that can see the revolution. We call $\nu(r_a(t_0);\theta)$ the visibility function of the linear threshold ABM. In the chosen notation for $\nu$ it is emphasized that $\nu$ depends on $r_a(t_0)$, i.e. on the fraction of nodes that are active at time $t_0$. At time $t=t_0$, the pool of individuals that can go from inactive to active, i.e. that are considering joining the revolution, is therefore approximately $(1-r_a(t_0))\spc \nu(r_a(t_0);\theta)$. Since inactive nodes that can see the revolution become active at the first arrival time of a Poisson process with rate $c_1$, we have \begin{equation} g(t|t_0) = c_1 \spc (1 - r_a(t_0)) \spc \nu(r_a(t_0);\theta) \spc \Delta t+ o(\Delta t). \end{equation} Combining Equations \eqref{eq:preApprox}-\eqref{eq:A0} and dividing by $\Delta t$ gives $$ \frac{\Delta r_a}{\Delta t} = c_1\mbox{ }(1 - r_a(t_0)) \spc \nu(r_a(t_0);\theta) - c_2 \mbox{ } r_a(t_0) \mbox{ }p(r_a(t_0);\beta) + o(1). $$ Taking the limit as $\Delta t\rightarrow 0$ yields \begin{equation} \frac{dr_a}{dt} = c_1 \mbox{ }(1 - r_a)\mbox{ }\nu(r_a;\theta)- c_2 \mbox{ }r_a \mbox{ }p(r_a;\beta). \end{equation} This establishes a general population-level ODE model that approximates the linear threshold ABM, with visibility function $\nu(r_a;\theta)$. To close the model, the visibility function has to be specified, guided by Equation (). Observe that if we substitute $\nu(r_a;\theta)$ by the step function $v_s(r_a;\alpha(\theta))$, then this exactly recovers the SVF model of , see \eq{SVF}. So, provided that the step function $v_s(r;\alpha(\theta))$ is a suitable choice for $\nu(r;\theta)$, which we argued in the appendix of based on considering a simplified averaged threshold process that ignores correlations, this confirms the consistency of the ABM and the SVF. The crux of the argument used in , which we discuss further in Sec.\ , is that, for real networks $\nu(r;\theta)$ can be expected to be a steep sigmoidal function of $r$, which can be approximated by a step function. In this paper, we test these assumptions explicitly on the empirical physical contact and Facebook social networks discussed above using ABM simulations, and we derive new improved ODE models that form a more accurate representation of $\nu(r;\theta)$ than a step function, taking into account the actual network structure of the empirical networks. The step function assumption was chosen in because it is useful when little information is known about the underlying network structure. For example, without having to specify an underlying communication network, the SVF model provides a potential mechanism by which networks with increased visibility $\alpha$ are more susceptible to political revolution. The following section explores the relationship between the ABM and the SVF model numerically for empirical networks by using a fitting procedure to attempt to estimate the network-dependent $\alpha = \alpha(\theta)$ for which $v_s(r;\alpha(\theta))$ best approximates $\nu(r;\theta)$. Given this estimate for $\alpha=\alpha(\theta)$, we compare ABM simulations to solutions of \eq{SVF} in order to verify the consistency of the ABM and the SVF numerically and to investigate the ABM behaviour in the parameter regimes of the SVF model as described in Fig.\ . Section then expands on this by proposing and analyzing a more sophisticated approximation for the visibility function $\nu(r;\theta)$. We conclude this section by noting that the above derivation can easily be modified for the case where the $\theta_v$ are not uniform. In this case the growth and decay functions $g$ and $d$, respectively, can be computed by averaging over the distribution of the $\theta_v$'s. \subsubsection{Numerical Simulation} In this section we simulate the ABM on the physical contact and Facebook networks for a fixed set of model parameters and find the optimal $\alpha$ for \eq{SVF} that minimizes the difference between the ABM realizations and the output of \eq{SVF}. The closeness of the output of \eq{SVF} and the ABM realizations for broad ranges of parameter values demonstrates that the SVF model does indeed approximate the ABM, at least for non-pathological parameter choices. We begin by fixing the initial condition $r_a(0) = r_0$ and the parameters $\theta_v\equiv\theta$, $\beta$, $c_1$, and $c_2$. To compare realizations of the ABM with the solution to \eq{SVF} we use Gillespie's Algorithm (summarized in Appendix ) to simulate $rep=100$ realizations of the ABM, recording $r_a$ at each $t\in\{0,0.01,0.02,\ldots,1\}$, and then find the $\alpha$ that minimizes the difference (in $L_2$ norm) between the ABM realizations and the solution to \eq{SVF} (see Appendix ). The results of this procedure for ABM simulations performed on the physical contact and Facebook social networks with parameters $r_0=0.25$, $\theta=0.15$, $\beta=0.3$, $c_1=1$, and $c_2=9$ are given in Fig.~. For these parameters we find the optimal $\alpha$ is given by $\hat{\alpha} = 0.835$ and $\hat{\alpha} = 0.834$ for the Facebook and physical contact networks, respectively. Equivalently, the fitted value for the visibility threshold $1-\alpha$ is given by $1-\hat{\alpha} = 0.165$ and $1-\hat{\alpha} = 0.166$, respectively. Holding the initial condition $r_0=0.25$ and parameters $\beta=0.3$, $c_1=1$ and $c_2=9$ constant, we are able to repeat the fitting procedure to find fitted values of $\alpha$ for $\theta\in\{0.105,0.11,0.115,\ldots,0.25\}$. These results are displayed in Fig.~ and illustrate that for a large range of individual thresholds $\theta$ the Facebook social network has a slightly lower fitted visibility threshold $1 - \hat{\alpha}$ than does the physical contact network. Observe that, for the parameters employed in Figs.~- ($\theta=0.15$, $\beta=0.3$, $c_1=1$, and $c_2=9$), the SVF model lies in the stable police state dynamical regime (Region III0, $c^*=c_1/(c_1+c_2) \leq 1-\hat{\alpha} < \beta $). We note that this is critical to the fitting procedure, since the parameter $\alpha$ appears in the solution to \eq{SVF} only when the parameters $\alpha$, $\beta$, $c_1$, and $c_2$ lie in Region III0, see Appendix . We then also attempt to approximate the ABM by the SVF model in Regions IIIe and III1 by using the fitted $\hat{\alpha}$ displayed in Fig.~. Figures - show some representative results. As illustrated in Fig.~, for many combinations of parameters the SVF model is able to approximate the behaviour of the ABM in parameter Region IIIe (left panels of Fig.~) and parameter Region III1 (right panels of Fig.~). However, we observe that under certain circumstances stochastic effects and details of the network structure become important and the SVF ODE model fails to approximate, even in a qualitative sense, the behaviour of the ABM. Specifically, the SVF model fails to capture the behaviour of the ABM when the initial condition is small ($r_0<<1$), when the initial condition is near the boundary between the basins of attraction of $r=0$ and $r=1$, and when parameters $c_1$, $c_2$, $\alpha$, and $\beta$ are close to the boundary between parameter Regions III0, IIIe, or III1 in Fig.\ , see Figs.~, , and , respectively. These are well-known limitations of population-level ODE models that attempt to approximate complex dynamics . Note that these plots also provide comparisons with the binomial visibility function (BVF) model, which we introduce in Section . In the previous section we observed that setting $\nu(r_a;\theta) = v_s(r_a;\alpha(\theta))$ in \eq{Approx} exactly recovers the SVF model. Therefore, a natural strategy for improving on the ability of the SVF model to approximate the aggregate behaviour of the ABM is to specify an alternative visibility function that better captures the structure of the underlying network. Section proposes two such visibility functions and compares their ability to approximate the ABM to the ability of the SVF model to approximate the ABM. \FloatBarrier \section{Alternative Visibility Functions for Population-Level ODE} Due to the difficulty of analyzing and computational cost of simulating the ABM, as well as the limitations of the SVF in approximating the ABM, we propose two alternative choices for visibility functions (different from the step-visibility function $v_s(r;\alpha)$) that can lend insight into how the ABM process behaves on different networks. Specifically, Sections - introduce the binomial and empirical visibility functions, respectively (BVF and EVF). These two visibility functions can be seen to be equivalent in the limit of large network and sample size, as shown in Appendix . Section demonstrates numerically that the use of the BVF/EVF in \eq{Approx} represents an improvement over the use of the step visibility function $v_s(r;\alpha)$ (SVF) when approximating the ABM for real networks. \subsection{Binomial Visibility Function} Consider the following derivation. Suppose that $v\in V$ is a node in the network $G=G(V,E)$ with degree $k$, and that the fraction of nodes active in the revolution is $r$. Also, assume that the states of $v$'s neighbours are active with probability $r$ independently so that the probability of $v$ having $j$ neighbours active in the revolution is $$ \binom{k}{j}r^j(1-r)^{k-j}. $$ The probability that the fraction of $v$'s neighbours exceeds the linear threhsold $\theta$ is, therefore, \begin{equation} \sum_{j = \cl{\theta k}}^k \binom{k}{j} r^j(1-r)^{k-j} = \mbox{BinCDF}(k-\cl{\theta k};k, 1-r), \end{equation} where $\mbox{BinCDF}(x;n,p)$ is the cumulative distribution function for the binomial distribution with $n$ trials and probability of success $p$ evaluated at $x$. This quantity can be used to approximate the function $\nu(r; \theta)$ (the expected fraction of the population that can see the revolution). In we considered $k$ in \eq{ProtoBVF} to be the average degree over all nodes and argued that \eq{ProtoBVF} is a sigmoidal function that can be approximated by step-visibility function $v_s(r;\alpha)$ for some appropriate visibility parameter $\alpha$ (see Appendix for a brief summary of this argument). One possible improvement on adopting $\nu(r;\theta) = v_s(r;\alpha(\theta))$ in \eq{Approx} (resulting in the SVF) is to actually adopt $\nu(r;\theta) = \mbox{BinCDF}(\fl{k}-\cl{\theta k};\fl{k}, 1-r)$ instead, where $k$ is the average degree of the network. However, we can improve on this even further in terms of using more relevant information about the real graph by choosing $\nu(r;\theta)$ to be equal to \begin{equation} v_b(r;\theta,\rho_k) = \sum_k \rho_k \mbox{BinCDF}(k - \lceil \theta k \rceil; k, 1-r) = \sum_k \sum_{j=0}^{k-\cl{\theta k}} \rho_k \binom{k}{j} (1-r)^jr^{k-j}. \end{equation} We call the function $v_b(r;\theta,\rho)$ in \eq{BVF} the \emph{binomial visibility function} (BVF), see Fig.~. Note the steep sigmoidal shape of $v_b(r;\theta,\rho_k)$ for the two empirical networks of Fig.~, further justifying the choice of a step visibility function as an appropriate approximation in . We note that computing the binomial visibility function on some grid of $\mu+1$ discrete $r$-values $r\in\{r_i\}_{i=1}^{\mu+1}$ requires calculating a double sum for every $i=1,\ldots\mu+1$. In practice, this can be very expensive, especially for degree distributions $\rho_k$ with fat tails. For this reason, the following section introduces a complementary but less expensive approximation to the visibility function that is equivalent to the binomial visibility function of \eq{BVF} in the limit of large network and sample size. \subsection{Empirical Visibility Function} We now propose to reconstruct the binomial visibility function $v_b(r;\theta,\rho_k)$ empirically by sampling the underlying network directly. Specifically, for fixed $\theta$ we approximate $v_b$ by a discrete function $v_e\in \mathbb{R}^{\mu+1}$ where $\forall j=1,\ldots, \mu+1: v_{e,j} \approx v_b(\frac{j-1}{\mu})$. Since $v_b(0)=0$ and $v_b(1)=1$, we set $v_{e,1}=0$ and $v_{e, \mu+1}=1$. We calculate $v_{e,j}$ for $j=2,\ldots,\mu$ as follows. \begin{enumerate} \item Seed the network with $r_j=\frac{j-1}{\mu}$ active nodes in expectation. (Generate $N$ random numbers drawn from the uniform distribution on $[0,1]$: $rand_i$ for $i=1,\ldots,N$. If $rand_i < \frac{j-1}{\mu}$ then node $i$ is active, otherwise node $i$ is inactive.) \item Determine the fraction of nodes that can see the revolution. \item Repeat 1-2 a total of $rep$ times and set $v_{e,j}$ to be the average over the realizations. (We choose $rep=100$ in our numerical experiments.) \end{enumerate} We denote the linear interpolation between the $v_{e,j}$ by the function $v_e(r;\theta)$ \begin{equation} v_e(r;\theta) = \left\{ \begin{array}{ll} v_{e,j} & \mbox{if } r = \frac{j-1}{\mu} \mbox{ for some } j=1,\ldots,\mu+1 \\ v_{e,j}\frac{j/\mu - r}{1/\mu} + v_{e,j+1}\frac{r - (j-1)/\mu}{1/\mu} & \mbox{if } r\in [\frac{j-1}{\mu}, \frac{j}{\mu}] \mbox{ for some } j=1,\ldots,\mu\end{array}\right. \end{equation} which we call the \emph{empirical visibility function}, see Fig.~. We now make several brief remarks about the procedure for calculating the empirical visibility function, outlined above. Firstly, we remark that this procedure is easily extended to the case where $\theta_v$ are specified for each individual node. However, to keep our analysis as simple as possible we continue to restrict ourselves to the case where $\theta_v\equiv\theta$. In Appendix we show that the empirical and binomial visibility functions are equivalent in the limit of large network and sample size. Figure shows that they can also be expected to give similar results for finite network and sample size. Finally, we remark that compared to the binomial visibility function, the procedure for calculating the empirical visibility function is in practice significantly less costly to implement, especially for networks with fat-tailed degree distributions. \subsection{Comparison of ABM and BVF/EVF Models} In this section we compare the ability of the BVF model to approximate the ABM with that of the SVF model. We note that since the BVF and EVF are equivalent in the limit of large network and sample size and are very close in practice (see Fig.~), the observations we make about the behaviour of the BVF model also apply to the behaviour of the EVF model. We further note that the behaviour of the BVF/EVF model has parameters regimes that induce behaviour analogous to those observed in the SVF model. Since the BVF/EVF have sigmoidal shape, this follows directly from the analysis presented in Appendix B of for visibility and policing functions of general sigmoidal shape (the results from this analysis are summarized in Appendix ). We observe four types of outcomes when comparing the ability of the BVF model to approximate the ABM to the ability of the SVF model to approximate the ABM: \begin{enumerate} \item The BVF model may succeed better in approximating the dynamics of the ABM than the SVF model, even if the improvement is only qualitative in nature, see Figs.~a and b (right panel), , and (left panel), \item The BVF model may complement the dynamics predicted by the SVF model, see Fig.~, \item The BVF model may produce qualitatively similar predictions to the SVF model, see Figs.~, b (left panel) and (right panel), or \item Both the BVF and the SVF models fail to approximate the dynamics of the ABM either qualitatively or quantitatively, see Figs.~b (left panel) and . \end{enumerate} For each of these outcomes, the approximation of the ABM by the BVF model is no worse than the approximation by the SVF model, and, in many cases the approximation of the ABM by the BVF model is much better than the approximation by the SVF model. Therefore, we conclude that the BVF model, and hence the EVF model, represents an improvement over the SVF model with respect to their ability to approximate the behaviour of the ABM. This is no surprise, since the BVF/EVF model takes important information of the real network into account. \section{Higher Order ODE Models} In the previous section we introduced an extension to the SVF model. In particular, we showed that the BVF/EVF model is better able to approximate the ABM than the SVF model. However, both the BVF/EVF and SVF models are one-dimensional ordinary differential equation. This section now compares the performance of the BVF/EVF models as approximations to the ABM to the performance of a higher order (i.e., higher dimensional) compartmental model which we modify from . We call this model the \emph{degree approximation} (DA) because it compartmentalizes individuals based on their state as well as their degree. We demonstrate numerically that the BVF model, and hence the EVF model, is no worse at approximating the aggregate behaviour of the ABM than the DA. This highlights the usefulness of the BVF/EVF model, especially since the DA is in practice much more computationally expensive to solve and much harder to analyze than the BVF/EVF model. \subsection{Degree Approximation Model} Before we detail the approach which we modify from , it is useful to introduce some notation. We define $V_k$ to be the set of nodes with degree $k$ and $N_k$ to be the set of nodes that have at least one neighbour of degree $k$, \begin{align*} V_k &= \{v\in V: d_v = k\}, \mbox{ and}\\ N_k &= \{w\in V:\exists v\in V_k \mbox{ such that } (v,w)\in E\}, \end{align*} respectively, where we denote the degree distribution of $N_k$ by $\rho_{k,j}$, i.e. $$ \rho_{k,j} = \frac{\mbox{number of nodes of degree $j$ that have at least one neighbour of degree $k$}}{\mbox{number of nodes that have at least one neighbour of degree $k$}} = \frac{|V_j \cap N_k|}{|N_k|}. $$ As mentioned above, the degree approximation aggregates individuals according to their state and degree. Thus, conditioning on the state of the system at time $t_0$, we define $$ r^{(k)}(t) =r^{(k)} (t|t_0) = \frac{1}{|V_k|}\sum_{v\in V_k}\ex[ s_v(t)|t_0] $$ to be the fraction of nodes with degree $k$ that are expected to be in state 1 at time $t$. Since nodes and edges are neither created nor destroyed, the fraction of nodes with degree $k$ that are expected to be in state 0 at time $t$ is given by $1-r^{(k)}(t)$. Analogously, we define $$ r_d(t) = r_d(t|t_0) = \frac{1}{N} \sum_k |V_k|r^{(k)}(t) $$ to be the fraction of nodes expected to be in state 1 at time $t$. Applying the notation introduced above with the rules characterizing the dynamics of the process we find an equation analogous to \eq{preApprox} \begin{equation} \Delta r^{(k)}(t) = r^{(k)}(t) - r^{(k)}(t_0) = g^{(k)}(t|t_0) - d^{(k)}(t|t_0), \end{equation} where the expected growth and decay of the fraction of active nodes is modelled by the nonnegative growth and decay functions $g^{(k)}(t)=g^{(k)}(t|t_0)$ and $d^{(k)}(t)=d^{(k)}(t|t_0)$, respectively. In order to close the degree approximation we need to approximate the quantities $g^{(k)}(t)$ and $d^{(k)}(t)$ in terms of $r^{(k)}(t)$, $\{\theta_v\}_{v\in V}$, and $\beta$. For notational convenience we first suppress the time argument, and consider the case where $\theta_v\equiv\theta$. In order to approximate $g^{(k)}(t)$ or $d^{(k)}(t)$ in terms of $r^{(k)}$, $\theta$, and $\beta$, we are now required to make an assumption about how the states of nodes sharing a common neighbour are correlated. Specifically, as for the BVF and EVF models, for any fixed $v\in V_k$ we assume that the states of any two neighbours of $v$ are independent. This assumption implies that the probability of $v$ having exactly $j$ active neighbours is $\binom{k}{j}\bar{n}_k^j(1-\bar{n}_k)^{k-j}$, where $\bar{n}_k = \sum_l \rho_{k,l}r^{(l)}$ is the fraction of nodes in $ N_k$ with state 1. So, the probability of $v$ having at least $\lceil \theta k\rceil$ active neighbours is $$ \sum_{j=\lceil \theta k \rceil}^k \binom{k}{j} \bar{n}_k^j (1 - \bar{n}_k)^{k-j} = 1 - \Binom(\lceil \theta k \rceil -1;k,\bar{n}_k) = \Binom(k - \lceil \theta k \rceil; k,1-\bar{n}_k), $$ It follows that \begin{equation} g^{(k)}(t) = c_1 \spc (1- r^{(k)}(t_0)) \spc \Binom(k - \lceil \theta k \rceil; k,1-\bar{n}_k) \Dt + o(\Dt). \end{equation} As in \eq{A1}, we choose \begin{equation} d^{(k)}(t) = c_2 \spc r^{(k)}(t_0) \spc p(r_d(t_0);\beta) \Dt + o(\Dt). \end{equation} Combining equations \eqref{eq:preApproxDA}-\eqref{eq:A1DA} and dividing by $\Delta t$ gives $$ \frac{\Delta r^{(k)}}{\Delta t} = c_1 \spc (1 - r^{(k)}(t_0) ) \spc \Binom(k - \lceil \theta k \rceil; k,1-\bar{n}_k) - c_2 \spc r^{(k)}(t_0) \spc p(r^{(k)}(t_0);\beta) + o(1). $$ Now, taking the limit as $\Delta t\rightarrow 0$ yields \begin{equation} \frac{dr_k}{dt} = c_1 (1 - r^{(k)}) \Binom(k - \lceil \theta k \rceil; k,1-\bar{n}_k) - c_2 \spc r^{(k)} \spc p(r^{(k)}; \beta). \end{equation} \subsection{Comparison of ABM and Degree Approximation} In Section we showed that the approximation of the ABM by the BVF/EVF model is no worse, and usually much better, than the approximation of the ABM by the SVF model. In this section we show that the approximation of the ABM by the BVF/EVF model is no worse than the approximation of the ABM by the DA given in \eq{ApproxDA}. Specifically, we observe that the BVF and DA model produce similar qualitative predictions for most choices of parameters $\theta$, $\beta$, $c_1$, and $c_2$, see for example Fig.~. Also, although this is an atypical result, we present one set of parameters where the BVF model outperforms the DA on the Facebook social network $G_F$, and where the BVF model complements the DA on the physical contact network $G_P$, see Fig.~. We observe that, for the specific parameters and networks used in the simulations of Figs.\ -, the DA does not significantly outperform the BVF model. It was confirmed in extensive additional simulations that this occurs generically for large parts of the parameter space of the models. This is worth emphasizing once more, since the BVF model is much easier to analyze and much less costly to solve than the DA. This supports our proposition that the BVF (or EVF) model is a powerful tool for approximating and analyzing the ABM and a general technique for formulating tractable models for spreading processes on social networks that take real network characteristics into account. \section{The Basic Reproduction Number ($R_0$) and the Initial Slope of the BVF/EVF} In epidemiological models the \emph{basic reproduction number} $R_0$ is defined to be the number of secondary infections caused by a single infected individual introduced into a population that is entirely susceptible . Note that in this context exposure to one infected individual is sufficient for a new infection. If $R_0 < 1$ then an infected individual is, on average, unable to replace himself and the outbreak terminates on its own in short order. On the other hand, if $R_0>1$ then an infected individual is, on average, more than able to replace himself and the outbreak may spread into a full-blown epidemic. We define the basic reproduction number for our linear threshold model of political revolution analogously: the basic reproduction number is the average number of individuals that become active in the revolution due directly to the introduction of a single active individual into a population that is otherwise completely inactive. In this section, we show that the basic reproduction number in our linear threshold ABM is related to the initial slope of the BVF/EVF by the equation \begin{equation} R_0 = \underbrace{\frac{c_1}{c_1+c_2} }_{=c^*} v_b'(0;\theta,\rho_k), \end{equation} and discuss efficient ways to compute $R_0$. We begin by deriving an expression for $R_0$. Suppose that we activate individual $v\in V$ in the population, and suppose that individual $v$ has degree $k$. Further, suppose that this individual has $\mathcal{V}$ neighbours who can see the revolution (i.e., the linear threshold criterion is satisfied) once individual $v$ is activated. Let $w\in V$ be a neighbour of $v$'s that can see the revolution, and let $\tau_1$ and $\tau_2$ be the first arrival time of Poisson processes with rates $c_1$ and $c_2$, respectively. The probability that individual $w$ becomes active before individual $v$ becomes inactive is \begin{align*} \pr{\tau_1 < \tau_2} &= \int_0^\infty \pr{\tau_1 < k | \tau_2 = t}\pr{\tau_2 = t}dt\\ &= \int_0^\infty [1-\exp(-c_1 t)] c_2\exp(-c_2 t) dt\\ &= \frac{c_1}{c_1+c_2} = c^*. \end{align*} Since we only consider activations that result directly from $v$'s activation (and not activations caused by a combination of the activation of $v$ and the subsequent activation of $v$'s neighbours), for our purposes the behaviour of each of $v$'s neighbours is independent. Thus, the expected number of $v$'s neighbours which become active is $\mathcal{V}c^*$. Recall that the social network $G = G(V,E)$ has $N$ nodes, $M$ edges, degree distribution $\rho_k$ and secondary degree distribution $\rho_{k,j}$. For the single activated individual $v$ with degree $k$ the expected number of neighbours who can see the revolution is $$ \mathbb{E}[\mathcal{V}] = k \sum_j \rho_{k,j}\ind{ j\theta \leq 1}. $$ Averaging over all possible degrees $k$ now yields the basic reproduction number $$ R_0 =c^*\ex[\mathcal{V}] = \frac{c_1}{c_1+c_2} \sum_{k,j} k \rho_k \rho_{k,j}\ind{ j\theta \leq 1} = c^*\sum_{k=1}^\infty \sum_{j=1}^{\lfloor \theta^{-1} \rfloor} k \rho_k \rho_{k,j}. $$ It is now possible to prove the relationship given in \eq{R0visp}. Writing $$ R_0 =\frac{c_1}{c_1+c_2} \sum_{j=1}^{\lfloor \theta^{-1} \rfloor} \sum_{k=1}^\infty k \rho_k \rho_{k,j}, $$ it suffices to show that $$ j \rho_j = \sum_{k=1}^\infty k \rho_k \rho_{k,j}, $$ since \begin{align} v_b(r;\theta,\rho_k) &= \sum_{k=1}^\infty \rho_k \mbox{BinCDF}(k - \lceil \theta k \rceil; k, 1-r) \nonumber \\ & = \sum_{k=1}^\infty \sum_{j=0}^{k-\lceil \theta k \rceil} \rho_k \binom{k}{j} (1-r)^j r^{k-j} \nonumber \\ & = \sum_{k=1}^\infty \rho_k \left[ \binom{k}{0} r^k + \ldots +\binom{k}{k-\lceil \theta k \rceil} (1-r)^{k-\lceil \theta k \rceil}r^{\lceil \theta k \rceil}\right] \nonumber \\ &\implies v_b'(0;\theta,\rho_k) = \sum_{k=1}^\infty k \rho_k \spc \ind{\theta k \leq 1} = \sum_{k=1}^{\lfloor \theta^{-1}\rfloor} k \rho_k. \end{align} We note that together with \eq{R0visp}, \eq{vbprime} gives an exact and cheap way to compute $R_0$. From the undirected social network $G=G(V,E)$ form the directed social network $G' = G(V,E')$, where $ E' = \{(v,w): \{v,w\}\in E\}. $ In this case, the number of edges emanating from all nodes of degree $k$ is $k \rho_k N$. So, the number of edges emanating from nodes of degree $k$ and incident on nodes of degree $j$ is $k\rho_k N \rho_{k,j}$. Thus, the number of edges incident on nodes of degree $j$ is $$ N \sum_{k=1}^\infty k \rho_k \rho_{k,j}. $$ Equivalently, the number of edges incident on nodes of degree $j$ is $$ j\rho_j N. $$ Putting these two expressions together, we find $$ jp_j = \sum_{k=1}^\infty k\rho_k\rho_{k,j}, $$ which completes the proof. Intuitively, this makes sense since the number of nodes that can see the revolution when one node is activated can also be approximated by $$ N \, v_b\left(\frac{1}{N}; \theta, \rho_k\right) = \frac{v_b\left(\frac{1}{N}; \theta, \rho_k\right) - \overbrace{v_b\left(0; \theta, \rho_k\right)}^{=0}}{\frac{1}{N}} \approx v_b'\left(0; \theta, \rho_k\right). $$ \section{Application: Linear Threshold Propagation on Online versus Offline Social Networks} In this section we investigate experimentally the differences in spreading behaviour that occur under the linear threshold model for the spread of political revolutions when applied to some empirical online and offline social networks, searching for some initial quantitative evidence that political revolutions may be facilitated by the network structure of online social networks of social media. \subsection{Network Structure: Online versus Offline Social Networks} It is often assumed that the connectivity of modern online social networks was an important factor in the spread of political revolutions in the past decade, e.g., in the Arab Spring revolutions of 2011 , while traditional offline social networks (using in-person physical contact, or phone or mail interaction for safe communication) featured a different connectivity structure that was often severely restricted by censorship of the regime. Unfortunately, representative samples of the offline, traditional communication networks that were in existence prior to the adoption of new media technologies are unavailable at sufficient scale for countries affected by the Arab Spring, or indeed for any country. As opposed to the online networks of social media which are, by their nature, digitally stored and available, the offline social networks of pre-Internet societies have not been recorded at scale, simply because it was impractical in terms of cost and effort. This is a serious roadblock when investigating the effects of the structure of new media networks on the dynamics of political revolution, and if one wants to compare with propagation on pre-Internet social networks, it is necessary to identify proxy networks that are likely to be reasonable approximations to pre-Internet social networks, in terms of network structure. In this section, we use simulations of our ABM model on two empirical networks as a starting point to investigate differences in propagation properties that may arise between online and offline social networks within the linear threshold propagation model. We choose the small physical contact network $G_P$ between individuals from as a representative for offline social networks, and we choose the Facebook network $G_F$ as a representative for online social networks. \subsection{ABM on Facebook $G_F$ and Physical Contact $G_P$ Networks} We now briefly examine the differences in how the political revolution spreads on the Facebook social and physical contact networks via direct simulation of the ABM. In particular, choosing parameters $\theta=0.1$, $\beta=0.3$, $c_1=1$, and $c_2=0.2$ in the unstable police state region (Region III1, $1-\hat{\alpha} < \beta \leq c^*$), we simulate $rep=100$ realizations of the ABM for each initial condition $r_0\in\{0, 0.0015, 0.003, \ldots, 0.0495\}$ from time $t=0$ until time $t=20$. The average final size (at $t=20$) for each initial condition $\langle r_{final}|r_0\rangle = \left.\langle r_a(20) \rangle\right|_{r_0}$ is recorded and displayed in Fig.~, which demonstrates that for these parameters the political revolution described by our ABM propagates to a much greater extent on the Facebook social network, i.e. the online proxy network, than on the physical contact network, i.e. the offline proxy network. As an important illustration of the usefulness of the basic reproduction number $R_0$ we defined in Sec.\ , we find that this is consistent with a difference in $R_0$ for the two networks: the basic reproduction number is $R_0=1.12$ (above the value of 1 which is expected to be required for propagation) for the Facebook network, and $R_0=0.35<1$ for the physical contact network. These results indicate that the offline social network is less conducive to spreading the revolution in the ABM than the online social network: it has a smaller basic reproduction number, and in simulations a larger initial population of revolutionaries is required to spread the revolution. This provides some initial quantitative evidence that the spread of revolutions under a linear threshold process may occur more easily on modern online social networks than on traditional offline networks. While this is an interesting first observation, the next section discusses limitations and further investigations that are required to address this intriguing but complex question more comprehensively. \FloatBarrier \section{Discussion and Conclusion} In this paper we developed a linear threshold agent based model (ABM) to model the spread of a political revolution in a dictatorial regime. We showed that this model is consistent with previous simple step visibility function (SVF) ODE model developed in . Using the relationship between these two models as a template we developed a hierarchy of models of varying complexity that approximate the behaviour of the ABM, see Table~. Of all the models we have identified, we find that the BVF and EVF models (models of moderate complexity) offer the optimal combination of low computational complexity (cost), ease of analysis, and ability to approximate the behaviour of the ABM. Specifically, we find that for most parameters and initial conditions the BVF/EVF model is better able to approximate average ABM behaviour than the SVF. Also, for most networks the BVF/EVF model is much less costly to solve and much easier to analyze than the degree approximation from . Importantly, the analysis of the simple ODE models in terms of stability of solutions for various parameter regimes directly gives insight in the qualitative dynamics of the linear threshold ABM for the spread of political revolutions. We extended the concept of the basic reproduction number $R_0$ from epidemiology to the linear threshold ABM, we showed how it is related to the slope of the empirical or binomial visibility functions at $r=0$, and we provided efficient ways to compute or estimate it. Analogously to epidemiological models, when $R_0>1$ we expect the political revolution to spread, and when $R_0<1$ we expect it to die out. Thus, computing this quantity for a network can give an indication of how the ABM will behave on that network without the need to perform simulations, as we have demonstrated for empirical networks. The Facebook and physical contact networks we consider as case studies, i.e. the online and offline proxy networks, provide initial support to the hypothesis that the adoption of online social media may facilitate the spread of political revolutions by effectively changing the connectivity structure of the population in a way that makes linear threshold spreading more effective. Specifically, we find that for certain parameters the online proxy network is more susceptible to the linear threshold spreading process than the offline proxy network. Moreover, we find that the different behaviour of these two networks is consistent with the basic reproduction number calculated for these two networks. In addition to studying the ABM and its approximations on more and larger online and offline networks, much work remains to be done in the actual modelling of the political revolution process. For example, throughout this manuscript we have assumed that the linear threshold of individuals is constant for the population, that the underlying communication networks are static, that the graph is undirected, and that the nodes that are initially activated are chosen uniformly at random. Each of these assumptions represents a major simplification of reality that needs to be addressed to study further aspects of the spread of political revolutions on online social networks. For example, consider that in our model individuals estimate the current participation in the revolution by sampling their neighbours. Since individuals with greater sample size, i.e. larger degree, can form more accurate estimates of the current participation in the revolution, they should be willing to join the revolution at a lower linear threshold $\theta$ than individuals who are more uncertain in their estimate of revolution size. We would therefore expect the linear threshold to vary from individual to individual as a function of their network degree. Or, consider that one of the principal strengths of new media is the ability of individuals to search for both content and like-minded individuals. Thus, we would expect that the underlying communication network should be changing on the same time scale as the revolution. Finally, one might consider that the nodes that are predisposed to be active at the initial stages of a revolution, e.g. nodes that represent activists, may also have larger degree (if, for example, they are charismatic and allowed to accrue followers) or smaller degree (if, for example, they are the specific target of regime censorship). There is significant value in the kind of parsimonious ABM model on static networks that we have considered in this paper, because much can be learned from this type of model and it is easier to analyze and interpret than more complex models. Nevertheless, extensions along the lines sketched above are important avenues for further study. The simulations in this paper were performed in Matlab. All Matlab code and data files will be made publicly available online upon publication of this paper. \begin{thebibliography}{10} \bibitem{LangDeSterck14} J.C. Lang and H.~{De Sterck}. \newblock {The Arab Spring: A Simple Compartmental Model for the Dynamics of a Revolution}. \newblock {\em Mathematical Social Sciences}, 69:12--21, 2014. \bibitem{KempeKleinbergTardos03} D.~Kempe, J.~Kleinberg, and E.~Tardos. \newblock Maximizing the spread of influence through a social network. \newblock In {\em Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining}, pages 137--147, 2003. \bibitem{CentolaEtAl05} D.~Centola, R.~Willer, and M.~Macy. \newblock {The Emperor's Dilemma: A Computational Model of Self-Enforcing Norms}. \newblock {\em American Journal of Sociology}, 110(4):1009--1040, 2005. \bibitem{CentolaEguiluzMacy07} D.~Centola, V.M. Eguiluz, and M.W. Macy. \newblock Cascade dynamics of complex propagation. \newblock {\em Physica A}, 374:449--456, 2007. \bibitem{CentolaMacy07} D.~Centola and M.W. Macy. \newblock {Complex Contagions and the Weakness of Long Ties}. \newblock {\em American Journal of Sociology}, 113(3):702--734, 2007. \bibitem{Watts02} D.J. Watts. \newblock A simple model of global cascades on random networks. \newblock {\em Proceedings of the National Academy of Sciences}, 99(9):5766--5771, 2002. \bibitem{AmesEtAl11} G.M. Ames, D.B. George, C.P. Hampson, A.R. Kanarek, C.D. McBee, D.R. Lockwood, J.D. Achter, and C.T. Webb. \newblock {Using network properties to predict disease dynamics on human contact networks}. \newblock {\em Proceedings of the Royal Society B: Biological Sciences}, 278(1724):3544--3550, 2011. \bibitem{Bansal07} S.~Bansal, B.T. Grenfell, and L.A. Meyers. \newblock {When individual behaviour matters: homogeneous and network models in epidemiology}. \newblock {\em Journal of The Royal Society Interface}, 4(16):879--891, 2007. \bibitem{Hethcote00} H.W. Hethcote. \newblock The mathematics of infectious diseases. \newblock {\em SIAM Review}, 42(4):599--653, 2000. \bibitem{LindEtAl07} P.~Lind, L.~{da Silva}, J.~Andrade, and H.~Herrmann. \newblock Spreading gossip in social networks. \newblock {\em Physical Review E}, 76(3):036117, 2007. \bibitem{ZhaoEtAl11} L.~Zhao, Q.~Wang, J.~Cheng, Y.~Chen, J.~Wang, and W.~Huang. \newblock {Rumor spreading model with consideration of forgetting mechanism: A case of online blogging LiveJournal}. \newblock {\em Physica A}, 390(13):2619--2625, 2011. \bibitem{Granovetter78} M.~Granovetter. \newblock {Threshold models of collective behavior}. \newblock {\em American Journal of Sociology}, 83(6):1420--1443, 1978. \bibitem{Kuran91} T.~Kuran. \newblock {Now Out of Never: The Element of Surprise in the East European Revolution of 1989}. \newblock {\em World Politics}, 44(1):7--48, 1991. \bibitem{NekoveeEtAl07} M.~Nekovee, Y.~Moreno, G.~Bianconi, and M.~Marsili. \newblock {Theory of rumour spreading in complex social networks}. \newblock {\em Physica A: Statistical Mechanics and its Applications}, 374(1):457--470, January 2007. \bibitem{eltantawy2011arab} Nahed Eltantawy and Julie~B Wiest. \newblock The {Arab Spring} | social media in the {Egyptian} revolution: Reconsidering resource mobilization theory. \newblock {\em International Journal of Communication}, 5:18, 2011. \bibitem{khondker2011role} Habibul~Haque Khondker. \newblock Role of the new media in the {Arab Spring}. \newblock {\em Globalizations}, 8(5):675--679, 2011. \bibitem{tufekci2012social} Zeynep Tufekci and Christopher Wilson. \newblock Social media and the decision to participate in political protest: Observations from {Tahrir Square}. \newblock {\em Journal of Communication}, 62(2):363--379, 2012. \bibitem{KhamisVaughn11} S.~Khamis and K.~Vaughn. \newblock {Cyberactivism in the Egyptian Revolution: How Civic Engagement and Citizen Journalism Tilted the Balance}. \newblock {\em Arab Media and Society}, 14, 2011. \bibitem{McAuleyLeskovec12} Julian McAuley and Jure Leskovec. \newblock {Learning to discover social circles in ego networks}. \newblock In {\em Advances in Neural Information Processing Systems 25}, pages 548--556, 2012. \bibitem{SalatheEtAl10} Marcel Salath{\'e}, Maria Kazandjieva, Jung~Woo Lee, Philip Levis, Marcus~W Feldman, and James~H Jones. \newblock {A high-resolution human contact network for infectious disease transmission}. \newblock {\em Proceedings of the National Academy of Sciences}, 107(51):22020--22025, 2010. \bibitem{gursoy} Attila G{\"u}rsoy and Murat Atun. \newblock Neighbourhood preserving load balancing: a self-organizing approach. \newblock In {\em Euro-Par 2000 Parallel Processing}, pages 234--241. Springer, 2000. \bibitem{KissEtAl06} I.Z. Kiss, D.M. Green, and R.R. Kao. \newblock The effect of contact heterogeneity and multiple routes of transmission on final epidemic size. \newblock {\em Mathematical Biosciences}, 203(1):124--136, 2006. \bibitem{PastorSatorrasVespignani02} R.~Pastor-Satorras and A.~Vespignani. \newblock Epidemic dynamics in finite size scale-free networks. \newblock {\em Physical Review E}, 65(3):035108, 2002. \bibitem{LindquistEtAl10} J.~Lindquist, J.~Ma, P.~Driessche, and F.H. Willeboordse. \newblock Effective degree network disease models. \newblock {\em Journal of Mathematical Biology}, 62(2):143--164, 2010. \end{thebibliography} \appendix \section{Physical Contact Network Data} In this appendix, we briefly discuss the network extraction protocol for the physical contact network $G_P$ from . The network presented in was constructed by distributing wireless sensors to students, teachers, and staff at a U.S. high school during a one day period from approximately 08:00 to 16:30. When two wireless sensors are in proximity of one another, i.e. when they are less than approximately 3m apart, they register an interaction with a temporal resolution of 20s. Data are given for each seperate interaction in CSV format with three columns: ID of first wireless sensor, ID of second wireless sensor, duration of interaction (measured in 20s increments). A weighted undirected network $G_{P,0}$ is formed by connecting each pair of individuals with an edge whose weight is given by the total amount of time they spent in proximity to one another. In order to admit a comparison with the unweighted undirected Facebook subnetwork, we de-weight network $G_{P,0}$ by discarding all edges whose weight is less than the minimum duration $w_P$, and by weighting all remaining edges equally. We denote the largest connected component of the resulting network by $G_{P}$, and for convenience we refer to this network as the \emph{physical contact network}. The choice of the minimum duration $w_P$ is important since it determines, for example, the average degree of the physical contact network $G_P$. We choose $w_P$ such that the average degree of $G_P$ is close to the average degree of $G_F$, for the following reason. In the simulations carried out in this paper to compare propagation under the linear threshold ABM on $G_P$ and $G_F$, we seed the networks with a fixed percentage of active nodes and use the same threshold $\theta$ on both networks. We want to calibrate the average degree of $G_P$ to the average degree of $G_F$, such that, on average, nodes in $G_P$ have the same chance as nodes in $G_F$ to satisfy the linear threshold criterion and see the revolution. After this calibration, differences in propagation between $G_P$ and $G_F$ (for the same $\theta$) are only due to the differences in network structure that go beyond the average degree. This motivates us to choose $w_P = 34$ sensor measurements, equivalent to 11 minutes and 20 seconds, in order to match the average degrees of $G_P$ and the Facebook subnetwork as closely as possible. \section{Justification of Step Visibility Function} As argued in Appendix A of , the expression for the visibility function given in \eq{ProtoBVF}, with $k$ taken to be the average degree of the network, can be expected to have a steep sigmoidal shape, which can be approximated by a step function. Indeed, Fig.\ shows that $\mbox{BinCDF}(\lfloor k\rfloor - \lceil\theta k\rceil; \lfloor k\rfloor,1-r)$ has a clear sigmoidal shape, with a steep transition from 0 to 1. In it was also argued that increased average degree $k$ can be expected to co-occur with decreased threshold $\theta$, since the more neighbors one has the more certain one can be about the true extent of the revolution because large samples are more reliable. Under this assumption, $\mbox{BinCDF}(\lfloor k\rfloor - \lceil\theta k\rceil; \lfloor k\rfloor,1-r)$ can be considered a single-parameter sigmoidal function, which can be approximated by the single-parameter step visibility function (SVF) $v_s(r;\alpha)$, with parameter $\alpha$ indicating where the step transition occurs. We conclude this section with a few remarks regarding the assumption made in Sec.\ and in that the states of an individual's neighbors are uncorrelated. While this assumption is applied in many models of spreading processes, for example rumour spreading processes and epidemiological processes , this assumption was not verified in . In this paper we show that this assumption results in a valid approximation by comparing simulation results of the SVF model, and also the BVF/EVF models (which were derived in Sec.\ under the same assumption), with ABM simulations for empirical networks (which do not make this assumption), finding consistent results for large parts of parameter space. \section{Analysis of BVF/EVF Model} In this section, we summarize the parameter regime and solution stability analysis that was developed in Appendix B of for the one-dimensional ODE model \begin{equation} \frac{dr}{dt} = \dot{r} = \underbrace{c_1\spc (1-r) \spc \nu(r)}_{\gamma(r)} - \underbrace{c_2 \spc r\spc \rho(r)}_{\delta(r)}, \end{equation} with generic sigmoidal visibility function $\nu(r)$ as in Fig.~. Since the newly derived BVF/EVF in this paper have sigmoidal shape, the stability analysis of applies to the BVF/EVF models. As is shown in , the dynamics of model \eqref{eq:Ext} with generic sigmoidal visibility and policing functions closely follows the corresponding dynamics of model \eqref{eq:SVF} in Region II (an open interval of equlibria $\subset (0,1)$), Regions III0 and III1 (one equilibrium $\in (0,1)$), and Region IIIe (three equilibria $\in (0,1)$), thus establishing the equivalence in terms of dynamic behaviour of models \eqref{eq:SVF} (with step functions) and \eqref{eq:Ext} (with sigmoidal functions, including the BVF/EVF models). We summarize the phase portraits of the different regions for model \eqref{eq:Ext} in Fig.\ , which are analogous to the phase portraits of the SVF model shown in Fig.\ . See Appendix B of for a detailed derivation and further explanation. For the BVF/EVF models, the values of equilibrium quantities like $r^*$ and $r^{**}$ depend on $\theta$, $\beta$ and the network structure. It is also important to point out how the SVF stability region plot of Fig.\ generalizes to the BVF/EVF models. For a given empirical network, the BVF/EVF model does not feature a range of visibility parameters $\alpha$ like the SVF model, but each choice of $\theta$ results in a different BVF/EVF function. For a fixed $\theta$, the different phase portrait behaviours of Fig.\ occur depending on the values of $\beta$, $c_1$ and $c_2$. As a result, for a given empirical network, the parameter space for the BFV/EVF models as a function of $\theta$, $\beta$, $c_1$ and $c_2$ can be divided in regions as for the SVF model in Fig.\ , where $\theta$ replaces $\alpha$ as the parameter on the horizontal axis. The precise shape of the BVF/EVF parameter regions would depend on the actual empirical network considered. For example, the boundary between the II and III regions is in general not a straight line, but needs to be determined numerically. \section{Gillespie's Algorithm} Following , we numerically simulate the ABM by implementing Gillespie's Algorithm. In this appendix we give a brief overview of this algorithm. We first introduce some notation. As above, let $r_a$ denote the fraction of individuals in the population that are expected to be active in the revolution at time $t$. For each node $v \in V$ we let $\gamma_v(t)=1$ if $v$ can see the revolution at time $t$, i.e. if $v$ satisfies \eq{0to1}, and let $\gamma_{v}(t)=0$ otherwise. Furthermore, we let $\xi_{1,v}$ and $\xi_{2,v}$ denote the first arrival times of independent Poisson processes with rates $c_1$ and $c_2$, respectively. Gillespie's Algorithm is based on the fact that the sum of two independent Poisson variables is also a Poisson variable with rate equal to the sum of the rates of the original processes. It follows that $$ \sum_{v\in V} \left[ \xi_{1,v}\spc (1-s_{v}(t)) \spc \gamma_{v}(t)+ \xi_{2,v} \spc s_{v}(t) \spc p(r_a(t); \beta)\right] $$ is a Poisson process with rate $$ \Lambda = \sum_{v\in V} \left[ c_1 \spc(1-\spc s_{v}(t)) \spc \gamma_{v}(t) + c_2 \spc s_{v}(t) \spc p(r_a(t); \beta) \right]. $$ The first arrival time of this process, therefore, is an exponential random variable $\tau$ with rate $\Lambda$. At time $t+\tau$ the state of exactly one of the nodes will change. Moreover, since $\xi_{1,v}$ and $\xi_{2,v}$ are independent, the probability that the state of node $v$ will change is $$ \mathbb{P}_v = \frac{ c_1 \spc (1-\spc s_{v}(t)) \spc \gamma_{v}(t)+ c_2 \spc s_{v}(t) \spc p(r_a(t);\beta) }{\Lambda} $$ The Gillespie Algorithm then proceeds iteratively in three steps. \begin{enumerate} \item Find the time $\tau$ of the next event by drawing $\tau$ from an exponential distribution with rate $\Lambda$. \item Determine which node changes state by drawing one node from $V=\{v_i\}_{i=1}^N$, where node $v\in V$ is drawn with probability $\mathbb{P}_v$. \item Update $t \leftarrow t + \tau$ and re-calculate $s_{v}(t)$, $r(t)$, $\gamma_{v}(t)$. \end{enumerate} \section{Complete Analytic Solution to SVF Model} Since the complete analytic solution to the SVF model was not provided in but is useful in the current paper, we provide it here. The solution to \eq{SVF} is as follows. In Region I ($1-\alpha = \beta$) and Region II ($1-\alpha > \beta$) $$ r(t) = \left\{ \begin{array}{ll} r_0 e^{-c_2(t-t_0)} & \mbox{if } r_0 < \beta\\ r_0 & \mbox{if } r_0 \in [\beta, 1-\alpha]\\ 1 - (1-r_0)e^{-c_1(t-t_0)} & \mbox{if } r_0 > 1-\alpha \end{array} \right. . $$ In Regions III0 ($c^* \leq 1-\alpha < \beta$), IIIe ($1-\alpha < c^* < \beta$), and III1 ($1-\alpha < \beta \leq c^*$) $$ r(t) = \left\{ \begin{array}{ll} r_0 e^{-c_2(t-t_0)} & \mbox{if } r_0 < 1-\alpha\\ \\ \left[ c^*+ (r_0-c^*)e^{-(c_1+c_2)(t-t_0)} \right] \mathbb{I}_{\{t<t_{\alpha}\}} \\ \hspace{5mm} + (1- \alpha) e^{-c_2(t-t_\alpha)} \mathbb{I}_{\{t\geq t_{\alpha}\}} & \mbox{if } r_0 \in [\beta, 1-\alpha] \mbox{ and } c^* \leq 1-\alpha < \beta\\ \\ c^*+ (r_0-c^*)e^{-(c_1+c_2)(t-t_0)} & \mbox{if } r_0 \in [\beta, 1-\alpha] \mbox{ and } 1-\alpha< c^* < \beta\\\\ \left[ c^*+ (r_0-c^*)e^{-(c_1+c_2)(t-t_0)} \right] \mathbb{I}_{\{t<t_{\beta}\}} \\ \hspace{5mm} + \left[ 1 - (1- \beta) e^{-c_1(t-t_\beta)} \right] \mathbb{I}_{\{t\geq t_{\beta}\}} & \mbox{if } r_0 \in [\beta, 1-\alpha] \mbox{ and } 1-\alpha < \beta \leq c^*\\\\ 1 - (1-r_0)e^{-c_1(t-t_0)} & \mbox{if } r_0 > \beta \end{array} \right. , $$ where \begin{align*} t_\alpha & = t_0 - \frac{1}{c_1+c_2}\log\left(\frac{1-\alpha-c^*}{r_0-c^*}\right), \mbox{ and}\\ t_\beta & = t_0 - \frac{1}{c_1 + c_2}\log\left(\frac{\beta-c^*}{r_0-c^*}\right). \end{align*} \section{Equivalence of Binomial and Empirical Visibility Functions} We briefly argue that the empirical and binomial visibility functions are equivalent in the limit of large network and sample size, i.e. as $N, rep \rightarrow\infty$, see also Fig.~. Let $v\in V$ be a node in the network $G=G(V,E)$ with degree $k$. Now, fix $\theta$ and suppose that we are calculating $v_{e,j} \approx v_b(\frac{j-1}{\mu}) = v_b(r_j)$ via the algorithm presented in Sec.~. The $l$\textsuperscript{th} iteration of Step 1 of this algorithm selects $\mathcal{N}_l$ nodes uniformly at random to be active and results in $\mathcal{V}_l$ nodes that can see the revolution. Note that by the Law of Large Numbers \begin{align*} \frac{\mathcal{N}_l}{N} &\stackrel{p}{\rightarrow} r_j = \frac{j-1}{\mu} \mbox{ as $N\rightarrow\infty$, and }\\ v_{e,j} = \frac{1}{rep}\sum_{l=1}^{rep} \frac{\mathcal{V}_l}{N} &\stackrel{p}{\rightarrow} \mathbb{E}\left[\frac{\mathcal{V}_l}{N}\right] = \mathbb{E}\left[\frac{\mathcal{V}_1}{N}\right] \mbox{ as $rep\rightarrow\infty$,} \end{align*} where $\stackrel{p}{\rightarrow}$ denotes convergence in probability. Since the $\mathcal{N}_l$ activated nodes are chosen uniformly at random, the statuses of the neighbours of $v$ are independent. Therefore, the probability that $v$ can see the revolution is $$ \sum_{l = \cl{\theta k}}^k \binom{k}{l} \left( \frac{\mathcal{N}_l}{N}\right )^l \left(1-\frac{\mathcal{N}_l}{N} \right)^{k-j} = \mbox{BinCDF}\left(k-\cl{\theta k};k, 1-\frac{\mathcal{N}_l}{N}\right). $$ It follows that the expected fraction of nodes that have degree $k$ and can see the revolution is $$ \mathbb{E}\left[ \rho_k \mbox{BinCDF}\left(k-\cl{\theta k};k, 1-\frac{\mathcal{N}_l}{N}\right) \right], $$ and hence, the expected fraction of nodes that can see the revolution is $$ \mathbb{E}\left[ \frac{\mathcal{V}_l}{N}\right] = \mathbb{E} \left[\sum_k\rho_k \mbox{BinCDF}\left(k-\cl{\theta k};k, 1-\frac{\mathcal{N}_l}{N}\right)\right]. $$ We conclude the proof by observing that by the Continuous Mapping Theorem $\frac{\mathcal{N}_l}{N}\stackrel{p}{\rightarrow}r_j$ and $N\rightarrow\infty$ implies $$ \mathbb{E}\left[ \frac{\mathcal{V}_l}{N}\right] \stackrel{p}{\rightarrow} \sum_k\rho_k \mbox{BinCDF}\left(k-\cl{\theta k};k, 1-r_j\right) = v_b(r;\theta,\rho_k)\mbox{ as }N\rightarrow\infty, $$ and hence, $$ v_{e,j} \rightarrow v_b(r_j;\theta,\rho_k)\mbox{ as }N\rightarrow\infty, rep\rightarrow\infty. $$ |
1501.04093 | Title: Analytical Theory of Neutrino Oscillations in Matter with CP violation
Abstract: We develop an exact analytical formulation of neutrino oscillations in matter
within the framework of the Standard Neutrino Model assuming 3 Dirac Neutrinos.
Our Hamiltonian formulation, which includes CP violation, leads to expressions
for the partial oscillation probabilities that are linear combinations of
spherical Bessel functions in the eigenvalue differences. The coefficients of
these Bessel functions are polynomials in the neutrino CKM matrix elements, the
neutrino mass differences squared, the strength of the neutrino interaction
with matter, and the neutrino mass eigenvalues in matter. We give exact
closed-form expressions for all partial oscillation probabilities in terms of
these basic quantities. Adopting the Standard Neutrino Model, we then examine
how the exact expressions for the partial oscillation probabilities might
simplify by expanding in one of the small parameters {\alpha} and sin{\theta}13
of this model. We show explicitly that for small {\alpha} and sin{\theta}13
there are branch points in the analytic structure of the eigenvalues that lead
to singular behavior of expansions near the solar and atmospheric resonances.
We present numerical calculations that indicate how to use the small-parameter
expansions in practice.
Body: \title{ Analytical Theory of Neutrino Oscillations in Matter with CP violation } \author{Mikkel B. Johnson \\ Los Alamos National Laboratory, Los Alamos, NM 87545 \\ Ernest M. Henley\\ Department of Physics, University of Washington, Seattle, WA 98195\\ Leonard S. Kisslinger\\ Department of Physics, Carnegie-Mellon University, Pittsburgh, PA 15213\\} \begin{abstract} We develop an exact analytical formulation of neutrino oscillations in matter within the framework of the Standard Neutrino Model assuming 3 Dirac Neutrinos. Our Hamiltonian formulation, which includes CP violation, leads to expressions for the partial oscillation probabilities that are linear combinations of spherical Bessel functions in the eigenvalue differences. The coefficients of these Bessel functions are polynomials in the neutrino CKM matrix elements, the neutrino mass differences squared, the strength of the neutrino interaction with matter, and the neutrino mass eigenvalues in matter. We give exact closed-form expressions for all partial oscillation probabilities in terms of these basic quantities. Adopting the Standard Neutrino Model, we then examine how the exact expressions for the partial oscillation probabilities might simplify by expanding in one of the small parameters $\alpha$ and $\sin\theta_{13}$ of this model. We show explicitly that for small $\alpha$ and $\sin\theta_{13}$ there are branch points in the analytic structure of the eigenvalues that lead to singular behavior of expansions near the solar and atmospheric resonances. We present numerical calculations that indicate how to use the small-parameter expansions in practice. \end{abstract} \maketitle \noindent PACS Indices: 11.3.Er,14.60.Lm,13.15.+g \vspace{1mm} \noindent Keywords: \vspace{1mm} \noindent \section{Introduction} In this paper, we develop an exact analytical representation of neutrino oscillations~ in matter within the framework of the Standard Neutrino Model (SNM)~ with 3 Dirac Neutrinos. The exact closed-form expressions we give for the time-evolution operator $S(t,t')$ are obtained from $H_\nu$ using the Lagrange interpolation formula given in Ref.~. The resulting expressions are easily evaluated without any approximations. The paper is divided into two main parts. In the first, we summarize the main results of our theory. Details underlying the derivation are given in Appendices. We also retrieve the well-known two-neutrino flavor results as a special case of our general results. In the second part we address other analytical formulations found in the literature. The expansion of the neutrino oscillation probability in one of the small parameters $\alpha$ and $\sin^2\theta_{13}$ of the standard neutrino model (SNM) for $H_\nu$ is of particular interest. The seminal work along these lines is found in Refs.~. This work underlies many of the analyses and exploratory studies of experiments at present and future neutrino facilities, including our earlier work~. The present paper was undertaken, and used, for the purpose of independently confirming the results of Refs.~. We find that the accuracy of the expanded oscillation probability is restricted by the presence of branch points in the analytic structure of the eigenvalues of neutrinos propagating in matter. We also show that the regions where the expanded results are reliable is different for expansions in $\alpha$~ and $\sin^2\theta_{13}$~. We then map out regions where the expanded results are reliable by comparing numerical results to the exact results of our Hamiltonian formulation. Another recent study~ takes a complementary approach and finds that the predictions of Refs.~ can be improved in certain regions using an exact evaluation of the integral $I_{\alpha *}$ rather than the approximate one found there. It concludes that within these regions, predictions of $(\mu,~e)$ oscillations improve for certain values of the experimental parameters. The dimensionalities of the neutrino Hamiltonian $H_\nu$ and the parameter space characterizing the mixing of three neutrino pairs are sources of difficulty for finding a tractable representation of the oscillation probability. The Lagrange interpolation formula~ is enormously helpful, providing an exact and formally elegant expression for the exponentiation of an $n\times n$ matrix. The description of two-flavor neutrino oscillations is elementary by comparison. In that case, $H_\nu$ is a $2\times 2$ matrix, and the mixing is described by a single real parameter. \section { Neutrino Dynamics } We will be interested in the dynamics of the three known neutrinos and their corresponding anti-neutrinos in matter. This dynamics is determined by the time-dependent Schr$\ddot{o}$dinger equation, \begin{eqnarray} i\frac{d}{dt}|\nu(t)> &=& H_\nu |\nu(t)> ~, \end{eqnarray} where the neutrino Hamiltonian, \beq H_\nu=H_{0v}+H_{1} ~, \eeq consists of a piece $H_{0v}$ describing neutrinos in the vacuum and a piece $H_{1}$ describing their interaction with matter. The solutions of Eq.~() may be expressed in terms of stationary-state solutions of the eigenvalue (EV) equation \beq H_\nu|\nu_{mi}> &=& E_i |\nu_{mi}> ~, \eeq where the label ``$m$" indicates neutrino mass eigenstates, as distinguished from their flavor states sometimes denoted the label ``$f$". In operator form, this dynamics may be expressed in terms of the time-evolution operator $S (t',t)$, which describes completely the evolution of states from time $t$ to $t'$ and also satisfies the time-dependent Schr$\ddot{o}$dinger equation. We will examine neutrinos propagating in a uniform medium for interactions constant not only in space but also time. Because the Hamiltonian is then translationally invariant, attention may be restricted to states, both in the vacuum and in matter, characterized by momentum $\vec p$ and therefore having the overall $r$-dependence $e^{i\vec p\cdot\vec r}$. In this case expressions may be simplified by suppressing the overall plane wave, a convention we adopt. For time-independent interactions, $S(t',t)$, \beq S(t',t) &=& e^{-iH_\nu (t' - t)} ~, \eeq depends on time only through the time {\it difference} $t'-t$. Then, written in terms of the stationary state solutions $|\nu_{mi}>$ of Eq.~(), \beq S(t',t) &=& \sum_i |\nu_{mi}> e^{-iE_i (t' - t ) }<\nu_{mi}| ~. \eeq With the momentum dependence factored out, three basis states $|M(i)>,~i=(1,2,3)$ are then required to describe three neutrinos. The basis states correspond to a specific representation, as in descriptions of a spin-1 object. The basis should, of course, be orthogonal, \beq <M(i)|M(j)> &=& \delta_{ij} \eeq and complete, \beq \sum_{k} |M(k)><M(k)| = \bf 1 ~. \eeq Once the basis is chosen, wavefunctions for a neutrino state are naturally introduced as the components of this state in the chosen basis. For example, with the eigenstates of Eq.~() expanded in the basis, \beq |\nu_{mj}> &=& \sum_i |M(i)> m^{i}_{j} ~, \eeq the wave functions of $|\nu_{mj}>$ would be the set $m^{i}_{j},~i = (1,2,3)$. With the plane wave factored out, the wave function is just a set of three numbers. Additionally, introduction of a basis makes it possible to represent neutrino states and operators such as $H_\nu$ in matrix form, with each entry in the matrix corresponding to a projection of the object being described onto the basis. In this paper we take the Hamiltonian in Eq.~() to be expressed in the {\it standard representation}, where the mass basis states $|M(i)>$ are taken as the set of states that diagonalize the neutrino vacuum Hamiltonian $H_{0v}$, {\it i.e.} $|M(i)> \equiv |\nu^0_{mi}> = |{\bar \nu}^0_{mi}>$, \beq H_{0v}|\nu^0_{mi}> = E^0_i |\nu^0_{mi}> ~. \eeq In matrix form \beq H_{0v} &=& \left( \begin{array}{ccc} E^0_3 & 0 & 0 \\ 0 & E^0_2 & 0 \\ 0 & 0 & E^0_1 \end{array} \right ) ~, \eeq with the EV's taken to be ordered $E^0_1 \leq E^0_2 \leq E^0_3$ as in the normal mass hierarchy. In the literature, the Hamiltonian is often expressed in a different basis obtained by rotating to one in which the complete neutrino Hamiltonian is diagonal as in Ref.~. We assume here that that neutrinos and anti-neutrinos represented by $|\nu^0_{mi}>$ and $|{\bar \nu}^0_{mi}>$, respectively, are the structureless elementary Dirac fields of the the Standard Neutrino Model~. For this reason the theory is invariant under CPT, so the mass of an anti-neutrino in the vacuum is the same as that for its corresponding neutrino. \subsection {Flavor and Mass States} Neutrinos are produced and detected in states of good flavor, $|\nu_{fi}>$. The three flavors, electron ($e$), muon ($\mu$), and tau ($\tau$) correspond, respectively, to the index values $i=(1,2,3)$. In the vacuum, each flavor state is a specific linear combination of the three mass eigenstates $|M(i)>$ of the neutrino vacuum Hamiltonian $H_{0v}$. This linear combination is expressed in terms of the same set of numbers $U_{ij}$ for both neutrinos and anti neutrinos \beq |\nu^0_{fi}> &=& \sum_j U^*_{ij} |M(j)> \nonumber \\ |{\bar \nu}^0_{fi}> &=& \sum_j U_{ij} |M(j)> ~, \eeq where $U_{ij}$ are the elements of a unitary operator $U$, the neutrino analog of the familiar CKM matrix. It is standard to express $U_{ij}$ in terms of three mixing angles $(\theta_{12},\theta_{13},\theta_{23})$ and a phase $\delta_{cp}$ characterizing $CP$ violation, \beq &\left( \begin{array}{ccc} c_{12} c_{13} & s_{12} c_{13} & s_{13} e^{-i\delta_{cp}} \\ U_{21} & U_{22} & s_{23} c_{13} \\ U_{31} & U_{32} & c_{23} c_{13} \end{array} \right ) ~, \eeq where \beq U_{21}&=& -s_{12} c_{23} - c_{12} s_{23} s_{13} e^{i\delta_{cp}} \nonumber \\ U_{22}&=& c_{12} c_{23} - s_{12} s_{23} s_{13} e^{i\delta_{cp}} \nonumber \\ U_{31}&=& s_{12} s_{23} - c_{12} c_{23} s_{13} e^{i\delta_{cp}} \nonumber \\ U_{32}&=& - c_{12} s_{23} - s_{12} c_{23} s_{13} e^{i\delta_{cp}} ~. \eeq We use here the standard abbreviation $s_{12}\equiv\sin{\theta_{12}}$, $c_{12}\equiv\cos{\theta_{12}}$, {\it etc}. The parameters $\theta$ and $\delta_{cp}$ are determined from experiment. Because $U_{ij}\rightarrow U^*_{ij}$ with $\delta_{cp}\rightarrow -\delta_{cp}$ it follows that the relationship in Eq.~() between flavor and mass states for anti-neutrinos and neutrinos in the vacuum is equivalent to $\delta_{cp} \leftrightarrow -\delta_{cp}$. \subsection{Neutrino Interacting Hamiltonian} The interaction $H_1$, determined by taking the electron flavor states scattering from the electrons of the medium to mediate the interaction, is then expressed as an operator in the standard representation, \beq H_{1} &=& U^{-1} \left( \begin{array}{ccc} V & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array} \right ) U ~, \eeq with $V = \pm\sqrt{2} G_F n_e $ and $n_e$ the electron number density in matter. For electrically neutral matter consisting of protons, neutrons, and electrons, the electron density $n_e$ is the same as the proton density $n_p$, \beq n_e &=& n_p \nonumber \\ &=& R N ~, \eeq where $N= n_n + n_p$ is the average total nucleon number density and $R= n_p/N$ is the average proton-nucleon ratio. In the earth's mantle, the dominant constituents of matter are the light elements so $R \approx 1/2$; in the surface of a neutron star $R<<1$. Matrix elements of $H_1$ are thus \beq <M(k)|H_{1}|M(k')> &=& U^*_{1k} V U_{1k'} ~. \eeq Matrix elements of $H_1$ are thus \beq <M(k)|H_{1}|M(k')> &=& U^*_{1k} V U_{1k'} ~. \eeq \subsection{Dimensionless variables} The results are most naturally expressed in dimensionless variables. We first take advantage of the global phase invariance to express all energies relative to the vacuum EV $E^0_1$ of the same momentum. We indicate that a quantity is expressed relative to $E^0_1$ by placing a bar over it, {\it e.g.}, \beq \bar E^0_i &\equiv& E^0_i - E^0_1 ~. \eeq We follow the same convention for the Hamiltonian, \beq {\bar H}_\nu &\equiv& H_\nu - {\bf 1} E^0_1 ~, \eeq so the EV equation Eq.~() becomes \beq ( \bar H_{0v} + H_1) |\nu_{mi}> &=& \bar E_i |\nu_{mi}>, \eeq where \beq \bar H_{0v} &\equiv& H_{0v} -{\bf 1} E^0_1 ~. \eeq Then, to express the theory in dimensionless variables we divide all energies, including the Hamiltonian, by $\bar E^0_{3} = E^0_3 - E^0_1$. The stationary-states $|\nu_{mi}>$ are also be determined from the dimensionless Hamiltonian ${\hat {\bar H}}_\nu $, \beq {\hat {\bar H}}_\nu &=& {\hat {\bar H}}_{0v} + {\hat H}_1 ~, \eeq {\it i.e.}, from the solutions of \beq {\hat {\bar H}}_\nu |\nu_{mi}> &=& {\hat {\bar E}}_i |\nu_{mi}> ~, \eeq where the ``hat" placed over a quantity indicates it is dimensionless. Thus \beq {\hat {\bar E}}_i &\equiv& \frac{ \bar E_i }{ \bar E^0_{3} } \nonumber \\ {\hat {\bar H}}_{0v} &\equiv& \frac{ {\bar H}_{0v} } { \bar E^0_{3} } ~, \eeq and ${\hat H}_1 $ is obtained from $H_1$ by replacing \beq V \rightarrow \hat A &\equiv& \frac{V }{\bar E^0_3} ~. \eeq The quantity ${\hat A}$ is the same as that defined in Refs.~. The connection of the Hamiltonian ${\hat {\bar H}}_\nu $ to the full Hamiltonian $ H_\nu = H_{0v}+H_{1} $ is then \beq H_\nu &=& {\bf 1} E^0_1 + \bar E^0_{3} {\hat {\bar H}}_\nu ~. \eeq \subsection { Neutrino Vacuum Hamiltonian ${\hat {\bar H}}_{0v}$ } The case of main interest for many situations is the ultra-relativistic limit, $\vec |p| >> m^2$ (we take the speed of light $c=1$). For ultra-relativistic neutrinos in the laboratory frame, the energy of a neutrino in the vacuum becomes \beq E^0_i &\approx& |\vec p| + \frac{m_i^2}{2E} ~, \eeq where $m_i$ is its mass the vacuum. Similarly, $E_i$ appearing in Eq.~() may be written \beq E_i &\approx& |\vec p| + \frac{M_i^2}{2E} ~, \eeq where $M_i$ is its mass in the medium. Thus, in this limit, \beq {\hat {\bar E}}_i &\to& \frac{M_i^2 - m^{2}_1}{m^{2}_3 - m^{2}_1} \eeq and \beq {\hat {\bar H}}_{0v} &\to & \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & \alpha & 0 \\ 0 & 0 & 1 \end{array} \right ) \eeq with \beq \alpha \equiv \frac{m_2^2-m_1^2}{m_3^2-m_1^2} ~. \eeq In this limit, the distance $L$ from the source to the detector corresponding to $S(t',t)$ in Eq.~() is \beq L &=& t'-t ~. \eeq The time-evolution operator, Eq.~(), expressed in dimensionless variables is, \beq S(L) &=& e^{-i H_\nu (t' - t)} \nonumber \\ &=& e^{2i {\hat {\bar E}}^0_1 \Delta_L} e^{-2i {\hat {\bar H}}_\nu \Delta_L} ~, \eeq where ${\hat {\bar H}}_\nu $ is given in Eq.~(), and where \beq \Delta_L &\equiv& \frac{L(m^2_3-m^2_1)}{4E} ~. \eeq [The similar quantity $\Delta_L$ as defined in Ref.~ is exactly one-half of that appearing in Eq.~().] \subsection{Neutrino Mass Eigenvalues} The neutrino mass eigenstates in a medium are solutions to the EV equation for ${\hat {\bar H}}_\nu$, Eq.~(). In many familiar formulations~ the full solutions, including both the eigenstates $|\nu_i>$ and EV's ${\hat {\bar E_i}}$, are required to find the neutrino oscillation probabilities. \subsubsection{Diagonalization of Neutrino Hamiltonian} The energies ${\hat {\bar E_i}}$ are solutions of the cubic equation~ \beq &&{\hat {\bar E_i}}^3 +a {\hat {\bar E_i}}^2 + b {\hat {\bar E_i}} +c = 0 ~, \eeq where \beq a &=& -(1+ \alpha + \hat A ) \nonumber \\ b &=& \alpha + \hat A \cos^2{ \theta_{13}} +\hat A \alpha C^{(+)}_2 \nonumber \\ c &=& -\hat A \alpha \cos^2{ \theta_{12}} \cos^2{ \theta_{13}} ~. \eeq We have expressed $b$ in terms of a frequently recurring combination of mixing angles, \beq C_2^{(\pm)} &\equiv& \cos^2{ \theta_{12}} \pm \sin^2{ \theta_{12}} \sin^2{ \theta_{13}} ~. \eeq Note that the mass eigenstate energies are independent of $\delta_{cp}$ and $\theta_{23}$ for both neutrinos and antineutrinos. The solutions of Eq.~() are expressed conveniently in terms of the quantity $d$, \beq d &=& \psi+ \sqrt{\psi^2-4\gamma^3} \nonumber \\\ \gamma &\equiv& a^2-3 b \nonumber \\ \psi &\equiv& a^3 - 27 c - 3 a \gamma ~. \eeq These solutions are real when \beq |d^{1/3}|^2 &=& 2^{2/3} \gamma ~ > 0 ~, \eeq which requires \beq \psi^2 < 4 \gamma^3 ~, \eeq and, thus, that $d$ be complex. Because having real energies is required by Hermiticity of the neutrino Hamiltonian, Eqs.~(,) amount to conditions on all parameter sets in terms of which $H_\nu$ is defined. We find \beq {\hat {\bar E_1}} &=& -\frac{a}{3} - \frac{1 }{ 3 \cdot 2^{1/3} } ( \sqrt{3} Im[ d^{1/3}] + Re[ d^{1/3}]) \nonumber \\ {\hat {\bar E_2}} &=& -\frac{a}{3} + \frac{1 }{ 3 \cdot 2^{1/3} } (\sqrt{3} Im[ d^{1/3}] - Re[ d^{1/3}]) \nonumber \\ {\hat {\bar E_3}} &=& -\frac{a}{3} + \frac{2^{2/3} }{ 3} Re[ d^{1/3}] ~. \eeq The masses are ordered so that $ m_3 > m_2 > m_1$. Because EV do not cross, $ {\hat {\bar E}}_3 > {\hat {\bar E}}_2 >{\hat {\bar E}}_1$ for all $|{\hat A}|$. A simple constraint among $ {\hat {\bar E}}_i$ is found from the trace of Eq.~(), \beq Tr{\hat {\bar H}}_\nu &=& {\hat {\bar E_1}} + {\hat {\bar E_2}} + {\hat {\bar E_3}} \nonumber \\ &=& Tr {\hat {\bar H}}_{0v} + Tr {\hat H}_1 \nonumber \\ &=& 1 + \alpha + {\hat A} \equiv -a ~. \eeq \subsubsection{Using neutrino mass eigenvalues in our Hamiltonian Formulation} In our formulation, the entire dependence of the time evolution operator on the neutrino eigenvalues occurs through three eigenvalue combinations, \beq \Delta {\hat {\bar E}}_{\ell\ell'} &\equiv& {\hat {\bar E_\ell}} - {\hat {\bar E_{\ell'}}} \nonumber \\ \Sigma_{\ell\ell'} &\equiv& {\hat {\bar E_\ell}} + {\hat {\bar E_{\ell'}}} \nonumber \\ \Pi_{\ell\ell'} &\equiv& {\hat {\bar E_\ell}} {\hat {\bar E_{\ell'}}} ~, \eeq with $\ell > \ell'$ (and powers thereof). We denote such quantities using a bracket notation, For example, \beq \Delta {\hat {\bar E }}[1] &=& {\hat {\bar E}}_3 - {\hat {\bar E}}_2 \nonumber \\ \Delta {\hat {\bar E }}[2] &=& {\hat {\bar E}}_3 - {\hat {\bar E}}_1 \nonumber \\ \Delta {\hat {\bar E }}[3] &=& {\hat {\bar E}}_2 - {\hat {\bar E}}_1 ~, \eeq in the case of $\Delta {\hat {\bar E}} $. We will generally use this bracket notation also for other quantities in our formulation that depend on two indices $(\ell,\ell')$, such as $\Sigma_{\ell\ell'} $ and $\Pi_{\ell\ell'} $. An expression for $ \Sigma [\ell] $, \beq \Sigma[\ell] &=& -a - {\hat {\bar E_\ell}} ~, \eeq follows from Eq.~(). An equivalent expression for $\Pi[\ell] $ in terms of ${\hat {\bar E_\ell}}$ is found by subtracting Eq.~() for ${\hat {\bar E_\ell}} $ and that for ${\hat {\bar E_{\ell'}}} $ and dividing through by $ \Delta {\hat {\bar E}}_{\ell\ell'} $. We find \beq 0 &=& ( {\hat {\bar E_\ell}}^2 + {\hat {\bar E_\ell}} {\hat {\bar E_{\ell'}}} + {\hat {\bar E_{\ell'}}}^2 ) \nonumber \\ &+& a ( {\hat {\bar E_\ell}} + {\hat {\bar E_{\ell'}}} ) + b \nonumber \\ &=& ( {\hat {\bar E_\ell}} + {\hat {\bar E_{\ell'}}} )^2 - {\hat {\bar E_\ell}} {\hat {\bar E_{\ell'}}} \nonumber \\ &+& a ( {\hat {\bar E_\ell}} + {\hat {\bar E_{\ell'}}} ) + b ~, \eeq giving \beq \Sigma[\ell]^2 - \Pi[\ell] + a \Sigma[\ell] + b &=& 0 ~. \eeq Then, using Eq.~(), \beq \Pi[\ell] &=& \Sigma_\ell ( \Sigma_\ell + a) + b \nonumber \\ &=& b + a {\hat {\bar E_\ell}} + {\hat {\bar E_\ell}}^2 ~. \eeq Finally, having observed that powers of the quantities given in Eq.~() will appear in various expressions, we note that $\Pi[\ell]^p$ and $\Sigma[\ell]^q$ with $p \geq 2$ and $q \geq 3$ involve linear combinations of eigenvalues ${\hat {\bar E_\ell}}^n$ with powers $n \geq 3$. Such terms are equivalently represented by a linear combination of three terms, one proportional to ${\hat {\bar E_\ell}}^2$, one proportional to ${\hat {\bar E_\ell}}$, and one independent of ${\hat {\bar E_\ell}}$, obtained by using the equation of motion repeatedly. We later make use of this fact to simplify various expressions. \section{ The S-Matrix in Our Hamiltonian Formulation} The probability $\mathcal{P}(\nu_a \rightarrow \nu_b)$ for neutrinos to oscillate from the initial state of flavor $a$ to a final state of flavor $b$ is found from the time-evolution operator $ S(t',t) $ as \beq \mathcal{P}(\nu_a \rightarrow \nu_b) &=& | S^{ab}(t',t)|^2 \nonumber \\ &\equiv& P^{ab}(t'-t) ~, \eeq where \beq |S^{ab}(t',t)|^2 &=& | <\nu^0_{fb}| S(t',t) |\nu^0_{fa}> |^2 ~. \eeq We accordingly determine here $ P^{ab}(t'-t) $ from $ S(t',t) $ defined in Eq.~(). In this section we review the formulation of neutrino oscillations based on the Lagrange interpolation formula as used in Ref.~. This formulation leads to exact, closed-form expressions for the time-evolution operator and the partial oscillation probabilities that are linear combinations of spherical Bessel functions in the eigenvalue differences whose coefficients are polynomials in the neutrino CKM matrix elements, the neutrino mass differences squared, the strength of the neutrino interaction with matter, and the neutrino mass eigenvalues in matter. We are led quite naturally to such expressions for all the partial oscillation probabilities in terms of these basic quantities. The numerical results given later in this paper are based on this formulation. \subsection{ Time-Evolution Operator } The overall phase in Eq.~() does not contribute to $| S^{ab}(t',t)|^2$, so for the purpose of calculating the oscillation probability, we may take \beq S(L) &\to& e^{-i {\hat {\bar H}}_\nu \Delta_L} ~. \eeq Then, with neutrinos created and detected in flavor states, which are coherent linear combinations of the neutrino vacuum mass eigenstates given in Eq.~(), \beq |\nu^{0}_{fa}> &=& \sum_{j} U^*_{a j} |M(j)> ~, \eeq we see that the mass eigenstate components of the flavor states contribute coherently to the time-evolution operator. Thus, \beq &&<\nu^0_{fb}| e^{-i {\hat {\bar H}}_\nu \Delta_L} |\nu^0_{fa}> \nonumber \\ &=& <M(b)| U e^{-i {\hat {\bar H}}_\nu \Delta_L} U^\dag |M(a)> ~. \eeq This coherence leads to the oscillation phenomenon. The elegant formulae for $ S(L) \equiv e^{-i {\hat {\bar H}}_\nu \Delta_L} $ are obtained from the Lagrange interpolation formula, Eqs.~(9,11) of Ref.~, \beq U e^{-i {\hat {\bar H}}_\nu \Delta_L} U^{-1} &=& \sum_{\ell} F_\ell \exp^{-i {\hat {\bar E}}_\ell \Delta_L } ~, \eeq where $T = L = t' - t$ and \beq F_\ell &\equiv& \Pi_{j\neq \ell} \frac{ U {\hat {\bar H}}_\nu U^{-1} - {\bf 1} {\hat {\bar E}}_j }{ {\hat {\bar E}}_{\ell} - {\hat {\bar E}}_j} ~. \eeq For three neutrinos, the sum in Eq.~() runs over three values of $\ell$ and the product in Eq.~() over two values of $j$. Using the convention that $O^{ab}$, written without parentheses enclosing $ab$, denotes the matrix elements of the operator $O$, \beq O^{ab} &\equiv& <M(b)| O |M(a)> ~, \eeq the matrix elements $ F^{ab}_{\ell} $ of $ F_{\ell} $ given in Eq.~() may be compactly written \beq F^{ab}_{\ell} &=& \frac{ <M(b)| {\hat {\bar W}}[\ell] |M(a)> } { {\hat {\bar D}}[\ell] } ~, \eeq where~, \beq {\hat {\bar W}}[1] &\equiv& (U {\hat {\bar H}}_\nu U^{-1} - {\bf 1}{\hat {\bar E}}_3 ) (U {\hat {\bar H}}_\nu U^{-1} - {\bf 1}{\hat {\bar E}}_2 ) \nonumber \\ {\hat {\bar W}}[2] &\equiv& (U {\hat {\bar H}}_\nu U^{-1} - {\bf 1}{\hat {\bar E}}_3 ) (U {\hat {\bar H}}_\nu U^\dag - {\bf 1}{\hat {\bar E}}_1 )\nonumber \\ {\hat {\bar W}}[3] &\equiv& (U {\hat {\bar H}}_\nu U^{-1} - {\bf 1}{\hat {\bar E}}_2 ) \nonumber \\ &\times& (U {\hat {\bar H}}_\nu U^\dag - {\bf 1}{\hat {\bar E}}_1 ) \eeq and \beq {\hat {\bar D}}[1] &=& ({\hat {\bar E}_{3}} - {\hat {\bar E}_1} )({\hat {\bar E}_{2}} - {\hat {\bar E}_1}) \nonumber \\ {\hat {\bar D}}[2] &=& ({\hat {\bar E}_{1}} - {\hat {\bar E}_2}) ({\hat {\bar E}_{3}} - {\hat {\bar E}_2}) \nonumber \\ {\hat {\bar D}}[3] &=& ({\hat {\bar E}_{1}} - {\hat {\bar E}_3}) ({\hat {\bar E}_{2}} - {\hat {\bar E}_3}) ~. \eeq Equations~(,) use the same bracket notation introduced in Eq.~(). The result in Eqs.~(,,) is immediately verified to be correct by inserting a complete set of intermediate eigenstates of $H_\nu$ in Eq.~(). It follows from the unitarity of $U$ that $ {\hat {\bar W}}[\ell] $ is Hermitian, \beq {\hat {\bar W}}[\ell]^\dagger &=& {\hat {\bar W}}[\ell] \eeq and that the two factors in Eqs.~() commute with each other. We find from Eq.~() that \beq Tr {\hat {\bar W}}[\ell] &=& {\hat {\bar D}} [\ell] ~. \eeq Equation~() establishes the reflection symmetry \beq F^{ab*}_{\ell} &=& F^{ba}_{\ell} ~. \eeq Explicit expressions for ${\hat {\bar W}}[\ell]$ are easily found in terms of $H_\nu$. The entire dependence of ${\hat {\bar W}}[\ell]$ on $\delta_{cp}$ occurs through three operators independent of $\delta_{cp}$, $W_0[\ell]$, $W_{cos}[\ell]$ and $W_{sin}[\ell]$, \beq {\hat {\bar W}}[\ell] &=& {\hat {\bar W}}_0[\ell] + \cos{\delta_{cp}} {\hat {\bar W}}_{cos}[\ell] \nonumber \\ &+& i \sin{\delta_{cp}} {\hat {\bar W}}_{sin}[\ell] ~, \eeq with ${\hat {\bar W}}_0[\ell]$, ${\hat {\bar W}}_{cos}[\ell]$ and ${\hat {\bar W}}_{sin}[\ell]$ real and independent of $\delta_{cp}$. Details are given in Appendix~. \subsection{Total Oscillation Probability } Expressions for $\mathcal{P}(\nu_a \rightarrow \nu_b)$ may be obtained directly from $S(L)$, \beq \mathcal{P}(\nu_a \rightarrow \nu_b) &=& | S^{ab}(t',t)|^2 \nonumber \\ &=& Re[S^{ab}(L)]^2 + Im[S^{ab}(L)]^2~. \eeq Convenient expressions for $Re[S^{ab}(L)]$ and $Im[S^{ab}(L)]$ defined by Eq.~() are presented in Appendix~. In our Hamiltonian formulation, the dependence of $ S(T)$ on the CP violating phase $\delta_{cp}$ is very simple and follows from Eqs.~(,) noting that $ F^{ab}_{\ell} = {\hat {\bar W}}^{ab}[\ell] / {\hat {\bar D}}[\ell] $, Eq.~(). We thus find, \beq Re[ S^{ab}(t',t) ] &=& \delta_{ab} - 2 \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_0[\ell]} { {\hat {\bar D}}[\ell] }\sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& 2 \cos{\delta_{cp}} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{cos}[\ell] } { {\hat {\bar D}}[\ell] }\sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &+& \sin\delta_{cp} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{sin}[\ell] } { {\hat {\bar D}}[\ell] }\sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~, \eeq and \beq Im[ S^{ab}(t',t) ] &=& - 2 \sin\delta_{cp} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{sin}[\ell] } { {\hat {\bar D}}[\ell] } \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_0[\ell] } { {\hat {\bar D}}[\ell] } \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& \cos{\delta_{cp}} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{cos}[\ell] } { {\hat {\bar D}}[\ell] } \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~, \eeq where $\Delta_L$ is defined in Eq.~(). Approximate expressions for $ \mathcal{P}(\nu_a \rightarrow \nu_b) $ in terms of the parameters of $ H_\nu$ were obtained from $S(L)$ in Refs.~ by an expansion in $\sin\theta_{13}$. \subsection{Partial Oscillation Probabilities } Using somewhat different techniques, the oscillation probability may be expressed through a set of functions that express how $\mathcal{P}(\nu_a \rightarrow \nu_b) \equiv P^{ab}$ depends on the CP violating phase $\delta_{cp}$~. In our Hamiltonian formulation there are four such terms, \beq P^{ab} &=& \delta(a,b) + P_0^{ab} + P_{\sin\delta}^{ab} + P_{\cos\delta}^{ab} \nonumber \\ &+& P_{\cos^2\delta}^{ab} ~, \eeq with $P_{\sin\delta}^{ab}$ linear in $\sin{\delta_{cp}}$, $P_{\cos\delta}^{ab}$ linear in $\cos{\delta_{cp}}$, $P_{\cos^2\delta}^{ab}$ quadratic in $\cos{\delta_{cp}}$, and $P_0^{ab}$ independent of $\delta_{cp}$. Although only the overall oscillation probability is a true probability, guaranteed to be strictly positive everywhere, we find it convenient to refer to these four terms as ``partial oscillation probabilities." Approximate expressions for the partial oscillation probabilities expanded in the small parameter $\alpha$ of the SNM in Ref.~. Obtaining expressions for the partial oscillation probabilities from the time-evolution operator requires additional analysis, given in Appendix~. In terms of spherical Bessel functions, we find there, \beq P_{sin\delta}^{ab} (\Delta_L,\hat A) &=& \sin { \delta_{cp} } \frac{4 \Delta_L } {{\hat D}} \sum_\ell (-1)^\ell w_{sin}^{ab} [\ell] \nonumber \\ &\times& j_0 ( 2 {\hat \Delta}[\ell] ) ~, \eeq where $ w_{sin}^{ab} [\ell] $, and therefore $ P_{sin\delta}^{ab} $, is anti-symmetric under $a\leftrightarrow b$. The other three partial oscillation probabilities are individually symmetric under $a\leftrightarrow b$. We find \beq P_{cos\delta}^{ab}(\Delta_L,\hat A) &=& - \cos {\delta_{cp} } \frac{ 4 \Delta_L^2} { {\hat {\bar D}} } \sum_\ell (-1)^\ell w_{cos}^{ab} [\ell] \nonumber \\ &\times& \Delta \hat {\bar E}}[\ell] j_0^2{ ( {\hat \Delta}[\ell] ) \nonumber \\ P_{cos^2\delta}^{ab}(\Delta_L,\hat A) &=& - \cos^2 { \delta_{cp} } \frac{ 4 \Delta_L^2} { {\hat {\bar D}} } \sum_\ell (-1)^\ell w_{cos^2}^{ab} [\ell] \nonumber \\ &\times& \Delta \hat {\bar E}}[\ell] j_0^2{ ( {\hat \Delta}[\ell] ) \nonumber \\ P_{0}^{ab}(\Delta_L,\hat A) &=& - \frac{ 4 \Delta_L^2} { {\hat {\bar D}} } \sum_\ell (-1)^\ell w_{0}^{ab} [\ell] \nonumber \\ &\times& \Delta \hat {\bar E}}[\ell] j_0^2{ ( {\hat \Delta}[\ell] ) ~, \eeq where all sums run over $\ell=1~,2~,3$, ${\hat \Delta}[\ell] $ is defined as \beq {\hat \Delta}[\ell] &\equiv& \Delta {\hat{\bar E}}[\ell] \Delta_L ~, \eeq with $\Delta {\hat {\bar E}[\ell]} $ defined in Eq.~(), ${\hat {\bar D} }$ is defined as \beq {\hat {\bar D} } &\equiv& \Delta {\hat {\bar E}[1]} \Delta {\hat {\bar E}[2]} \Delta {\hat {\bar E}[3]} ~, \eeq and the matrix elements $ w_{i}^{ab} [\ell] $ are given in terms of the mixing angles and ${\hat A}$ in Appendix~. Since we order the energies so that $ {\hat {\bar E}}_3 > {\hat {\bar E}}_2 >{\hat {\bar E}}_1 $, $\Delta {\hat {\bar E}[\ell]}$ as well as ${\hat {\bar D} }$ are all positive. We begin our derivation with the expression for the oscillation probability written in terms of $S(T)$, Eq.~(), \beq &&\mathcal{P}(\nu_a \rightarrow \nu_b) \equiv P^{ab} \nonumber \\ &=& |<M(b)| U e^{-iH_{\nu}(t' -t)} U^\dag |M(a)> |^2 \nonumber \\ &=& \sum_{\ell \ell' } F^{ab}_{\ell \ell' } \exp^{-i ( {\bar E}_{\ell} - {\bar E}_{\ell'} )L} ~. \eeq Here $ F^{ab}_{\ell \ell'} $ is defined as \beq F^{ab}_{\ell \ell'} &\equiv& F^{ab}_{\ell} F^{ab*}_{\ell'} = \frac{ w^{ab}_{\ell\ell'} } { {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } \eeq with ${\hat {\bar D}}[\ell] $ given in Eq.~(), and \beq w^{ab}_{\ell \ell'} &\equiv& <M(b)| {\hat {\bar W}}[\ell] |M(a)> \nonumber \\ &\times& <M(b)|{\hat {\bar W}}[\ell'] |M(a)>^* ~. \eeq All results needed for determining exact, closed-form expressions for the partial oscillation probabilities are found in Appendix~. As we explain in Appendix~, $ w^{ab}_{\ell \ell'}$ may be expressed through four operators, \beq w^{ab}_{\ell\ell'} &=& w_{0\ell\ell'}^{ab} + \cos{\delta_{cp}} w_{cos~\ell\ell'}^{ab} + \cos^2{\delta_{cp}} \nonumber \\ &\times& w_{cos^2\ell\ell'}^{ab} + i \sin{\delta_{cp}} w_{sin\ell\ell'}^{ab} ~. \eeq found from the decomposition of ${\hat {\bar W}}[\ell] $ given in Eq.~(). The quantities $ w_{i}^{ab} [\ell] $ appearing in Eqs.~(,) are the same as $ w_{i\ell\ell'}^{ab} $ written in the streamlined notation, \beq w_{i}^{ab} [1] &=& w_{i\ell\ell'}^{ab} ~~{\rm for} ~~(\ell,\ell')=(3,2) \nonumber \\ w_{i}^{ab} [2] &=& w_{i\ell\ell'}^{ab}~~{\rm for} ~~(\ell,\ell')=(3,1) \nonumber \\ w_{i}^{ab} [3] &=& w_{i\ell\ell'}^{ab}~~{\rm for} ~~(\ell,\ell')=(2,1) ~, \eeq which takes advantage of $\ell > \ell'$. An analytic expression for $ w^{ab}_{sin}[\ell]$, \beq w^{ab} _{sin}[\ell] &=& K \sin{2\theta_{23}} \Delta {\hat{\bar E}}[\ell] \epsilon^{ab}_{\sin} ~, \eeq follows directly from Eq.~(). Here, $\epsilon_{\sin}$ is the anti-symmetric matrix \beq \epsilon_{sin} &\equiv& \left( \begin{array}{ccc} 0 & 1 & -1 \\ -1 & 0 & 1 \\ 1 & -1 & 0 \end{array} \right) \eeq and \beq K &=& -\frac{\alpha(1-\alpha) }{8} \nonumber \\ &\times& \cos{\theta_{13}} \sin{2\theta_{12}} \sin{2\theta_{13}} ~. \eeq Equation~() is one of the more striking results. Analytic formulae for the other $ w_{i}^{ab}[\ell] $ follow from Eq.~(). These are all given in Appendix~, where they are expressed in terms of the parameters specifying ${\hat {\bar H}}_{0v}$ and ${\hat H}_1 = U^{-1} {\hat V} U$. These are the same parameters defining SNM. It follows from Eq.~() that the coefficients of $ \Delta {\hat{\bar E}}[\ell] j_0 ( 2 {\hat \Delta}[\ell] ) $ in Eq.~() are all proportional, leading to a simple expression for $P_{sin\delta}^{ab}$, \beq P_{sin\delta}^{ab}(\Delta_L,\hat A) &=& \Delta_L^3 \alpha (1-\alpha) \sin {\delta_{cp} } \epsilon_{sin}^{ab} \nonumber \\ &\times& \cos{\theta_{13}} \sin{ 2 \theta_{12}} \sin{2\theta_{13}} \sin{2\theta_{23}} \nonumber \\ &\times& j_0 ({\hat \Delta}[3] ) j_0 ( {\hat \Delta}[2] ) j_0 ( {\hat \Delta}[1] ) ~. \eeq Note that $P^{ab}_{\sin\delta}$ is antisymmetric under $a\leftrightarrow b$. Analytic expressions for all other partial oscillation probabilities follow from Eq.~() using Eq.~(), which expresses $ F^{ab}_{\ell \ell'} $ in terms of $w^{ab}_{\ell\ell'} $. In this fashion, these partial oscillation probabilities are also expressed in terms of the parameters of the SNM and the neutrino eigenvalues, ${\hat{\bar E}}_\ell$. The usefulness of the partial oscillation probabilities can be seen as follows. It is a general result that the exchange of initial and final states in the oscillation probability or neutrinos (antineutrinos) is equivalent to letting $\delta_{cp}\rightarrow -\delta_{cp}$. Thus, the result for the inverse reaction $\mathcal{P}(\nu_b \rightarrow \nu_a)$ is found by exchanging $(a,b)$ in Eq.~(). Since $ P_{\sin\delta}^{ab} $ is antisymmetric under the exchange of $(a,b)$, and $ P_0^{ab}$, $P_{\cos\delta}^{ab}$ and $P_{\cos^2\delta}^{ab}$ symmetric, it follows that $P^{ba}$ is given by \beq \mathcal{P}(\nu_b \rightarrow \nu_a) &=& \delta(a,b) + P_0^{ab} - P_{\sin\delta}^{ab} + P_{\cos\delta}^{ab} \nonumber \\ &+& P_{\cos^2\delta}^{ab} ~. \eeq In analogy to Eq.~(), we may express the oscillation probability for antineutrinos as \beq &&\mathcal{P}({\bar \nu}_a \rightarrow {\bar \nu}_b) \equiv {\bar P}^{ab} \nonumber \\ &=& \delta(a,b) + { \bar P}_0^{ab} + {\bar P}_{\sin\delta}^{ab} + {\bar P}_{\cos\delta}^{ab} + {\bar P}_{\cos^2\delta}^{ab} ~, \eeq where the bared probabilities for anti neutrinos are obtained from the unbarred for neutrinos by replacing $\delta_{cp}\rightarrow -\delta_{cp}$ and ${\hat A} \rightarrow - {\hat A}$. Because the energies of antineutrinos are different from those of the neutrinos in matter, we can expect $P^{ab} \neq {\bar P}^{ab}$ in this situation. Again applying the rule that exchange of initial and final states is accomplished by letting $\delta_{cp}\rightarrow -\delta_{cp}$, the oscillation probability $\mathcal{P}({\bar \nu}_b \rightarrow {\bar \nu}_a)$ is expressed in terms of the same four quantities, \beq &&\mathcal{P}({\bar \nu}_b \rightarrow {\bar \nu}_a) \nonumber \\ &=& \delta(a,b) + {\bar P}_0^ {ab} - {\bar P}_{\sin\delta}^{ab} + {\bar P}_{\cos\delta}^{ab} + {\bar P}_{\cos^2\delta}^{ab} ~. \eeq It is worth noting that the entire dependence of the oscillation probabilities given in Eqs.~(,) on the neutrino beam energy $E$, the baseline $L$, and the medium properties occurs through the variables $\Delta_L$ and ${\hat A}$ defined in Eqs.~(,), respectively. Since we will be most interested in how the neutrino oscillation probability depends on the beam energy, baseline, and medium properties, the partial oscillation probabilities have been expressed as functions of $\Delta_L$ and ${\hat A}$. Because the vacuum result is also easily obtained by less sophisticated arguments, the vacuum limit provides an opportunity to verify our Hamiltonian formulation in a well-known special case. \subsection{ Special Cases } We next examine special cases chosen to highlight specific aspects of our formulation. Because these cases are well known in other contexts, they provide useful cross checks. We first consider two-flavor mixing and then the vacuum limit. \subsubsection{ Two Flavor Mixing} Expressions for two-neutrino oscillations are easily obtained from the Hamiltonian, \beq {\hat {\bar H}}_\nu = {\hat {\bar H}}_{0\nu}+ U^{-1} {\hat {\bar H}}_{1\nu} U ~, \eeq where \beq {\hat {\bar H}}_{0 \nu} &\equiv& \left( \begin{array}{ccc} 0 & 0 & \\ 0 & 1 \end{array} \right) ~, \eeq \beq {\hat {\bar H}}_{1\nu} &\equiv& \left( \begin{array}{ccc} {\hat A} & 0 \\ 0 & 0 \end{array} \right) ~. \eeq The standard mixing matrix $U$ is \beq U &\equiv& \left( \begin{array}{ccc} \cos\theta & \sin\theta \\ -\sin\theta & \cos\theta \end{array} \right ) ~. \eeq The neutrino Hamiltonian is easily diagonalized to find the medium-modified two-flavor neutrino eigenvalues, \beq {\hat {\bar E_2}} &=& \frac{1}{2}\left(1+\hat A + \phi \right) \nonumber \\ {\hat {\bar E_1}} &=& \frac{1}{2}\left(1+\hat A - \phi \right) ~, \eeq where $\phi = \sqrt{1 + {\hat A}^2-2 {\hat A} \cos{2 \theta}}$. Expressed in dimensionless variables, the time-evolution operator $S^{ab}$ for the transition $a \to b$ becomes \beq S^{a b }&=& <M(b)| U \exp^{-2 i {\hat {\bar H}}_\nu \Delta_L}U^{-1} |M(a)> ~, \eeq with $\Delta_L$ given in Eq.~(). The two-flavor oscillation probability $P^{12}_{2f}(\Delta_L,\hat A) $ is then \beq P^{12}_{2 f}(\Delta_L,\hat A) &=& |S_{2f}^{12}|^2 \nonumber \\ &=& \frac{ \sin^2 2 \theta }{ \phi^2 } \sin^2\Delta_L \phi ~, \eeq which we will next compare to our Hamiltonian formulation. To find $S_{2f}^{1 2}=<M(2)|S[L]|M(1)>$, it is easy to overlook that the $2\times 2$ matrix $H_\nu$ in Eq.~() is not diagonal in the flavor basis . The oscillation probability in the three neutrino mixing of our Hamiltonian formulation is matched to the two-flavor case just discussed by setting two of the mixing angles in Eq.~() to zero, say $\theta_{13}\to 0$ and $\theta_{23}\to 0$ and identifying $\theta$ with $\theta_{12}$. Accordingly, we find, \beq U &\equiv& \left( \begin{array}{ccc} \cos\theta & \sin\theta & 0 \\ -\sin\theta & \cos\theta & 0 \\ 0 & 0 & 1 \end{array} \right ) ~. \eeq This is easily recognized as a two-flavor mixing matrix by noting that one of the three neutrinos does not mix with the other two. We see that it not only depends on just one mixing angle $\theta$, but also that the dependence on the CP violating phase $\delta$ has dropped out. Natural choices for a $3 \times 3$ two-flavor neutrino vacuum Hamiltonian and interaction are, \beq {\hat {\bar H}}_{0 \nu} &\equiv& \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 0 \end{array} \right) ~, \eeq and \beq {\hat {\bar H}}_1 &\equiv& \left( \begin{array}{ccc} {\hat A} & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array} \right) ~. \eeq When the $3\times 3$ neutrino Hamiltonian ${\hat {\bar H}}_\nu ={\hat {\bar H}}_{0\nu}+ U^{-1} {\hat {\bar H}}_{1\nu} U$ is diagonalized, two of the three medium-modified neutrino eigenvalues, \beq {\hat {\bar E_3}} &=& \frac{1}{2}\left(1+\hat A + \phi \right) \nonumber \\ {\hat {\bar E_2}} &=& \frac{1}{2}\left(1+\hat A -\phi \right) \nonumber \\ {\hat {\bar E_1}} &=& 0 ~. \eeq are identical to those found above in the two-flavor case. Note that the eigenvalues have no branch points for ${\hat A}$ on the real axis. From these eigenvalues, we find the differences, \beq \Delta {\hat {\bar E}} [3] &=& {\hat {\bar E_2}} -{\hat {\bar E_1}} \nonumber \\ &=& \frac{1}{2}\left(1+\hat A -\phi \right) \nonumber \\ \Delta {\hat {\bar E}} [2] &=& {\hat {\bar E_3}} -{\hat {\bar E_1}} \nonumber \\ &=& \frac{1}{2}\left(1+\hat A + \phi \right) \nonumber \\ \Delta {\hat {\bar E}} [1] &=& {\hat {\bar E_3}} -{\hat {\bar E_2}} = \phi ~. \eeq Finally, consider the oscillation probability in our Hamiltonian formulation. Because $P^{ab}_{\sin\delta}$, $P^{ab}_{\cos\delta}$, and $P^{ab}_{\cos^2\delta}$ are proportional to $\sin 2 \theta_{23}$ (and $\theta_{23}$ has been set to $0$), these terms do not contribute for two-flavor mixing. Thus, \beq P^{ab}(\Delta_L,\hat A)&=& \delta(a,b)+P^{ab}_0(\Delta_L,\hat A) \eeq with $P^{ab}_0(\Delta_L,\hat A)$ taken from Eq.~(). We find after a straightforward calculation, \beq w_0^{11} [\ell] &=& -w_0^{12}[\ell] \nonumber \\ w_0^{21} [\ell] &=& w_0^{12} [\ell] \nonumber \\ w_0^{22} [\ell] &=& -w_0^{12}[\ell] ~, \eeq where, \beq w_0^{12} [\ell] &=& w^{(12)}_{02} ({\hat A} \cos^2\theta - (1+{\hat A} ) {\hat {\bar E}}_\ell \nonumber \\ &+& {\hat {\bar E}}_\ell ^2) ~. \eeq Taking ${\hat {\bar E}}_\ell$ from Eq.~(), we evaluate \beq {\hat A} \cos^2\theta - (1+{\hat A} ) {\hat {\bar E}}_\ell + {\hat {\bar E}}_\ell ^2 ~, \eeq to find \beq w_0^{12} [1] = w_0^{12} [2] = 0 ~, \eeq and \beq w_0^{12} [3] &=& \hat A \cos^2\theta w^{(12)}_{0,2} \nonumber \\ &=& \frac{\hat A}{4} \cos^2\theta \sin^2 2\theta ~. \eeq Thus, for all two-flavor transitions, \beq P^{ab}(\Delta_L,\hat A) &=& \delta(a,b) + \frac{4 \Delta_L^2}{\hat D} w_0^{ab} [3] \Delta {\hat {\bar E[3]}} \nonumber \\ &\times& j_0^2(\Delta {\hat {\bar E[3]}} \Delta_L) ~, \eeq where $\hat D$ is from Eq.~(), \beq \hat D &\equiv& \Delta {\hat {\bar E}} [1] \Delta {\hat {\bar E}} [2] \Delta {\hat {\bar E}} [3] \nonumber \\ &=& \frac{\Delta {\hat {\bar E}} [1] }{4} ((1+\hat A)^2 - 1 - {\hat A}^2 +2 {\hat A} \cos{2 \theta} ) \nonumber \\ &=& \hat A \cos^2\theta \Delta {\hat {\bar E}} [1] ~. \eeq To compare our result to to the well-known two-flavor oscillation probability given in Eq.~(), we evaluate Eq.~() for $1 \to 2$ transitions obtaining, \beq P^{12}(\Delta_L,\hat A) &=& \frac{\sin^2 2\theta }{ \phi^2} \sin^2 \Delta_L \phi ~. \eeq As expected, this is in complete agreement with Eq.~(). \subsubsection{ Vacuum Oscillation Probability} In our Hamiltonian formulation, an expression for the time-evolution operator in the vacuum limit $\hat A \to 0$ is found from Eqs.~(), (), (), and (), \beq &&<M(b)|S^0(t',t)|M(a)> \equiv S^{0ab}(t'-t) \nonumber \\ &=& \sum_{\ell} \frac{<M(b)| {\hat {\bar W^0}}[\ell] |M(a)> }{ {\hat {\bar D^0}}[\ell] } \exp^{-i \bar E^0_\ell \Delta_L } ~. \eeq Here, $ \bar E^0_\ell $ and ${\hat {\bar D^0}}[\ell]$ are the vacuum values of $ \bar E_\ell $ and ${\hat {\bar D}}[\ell]$, respectively. In the vacuum, of course, there is no distinction between the energy differences for neutrinos and anti-neutrinos. Evaluating Eq.~() by inserting a complete set of states intermediate states $|n><n|$ inside ${\hat {\bar W}}^0[\ell]$, we arrive at \begin{widetext} \beq <M(b)|S^0(t',t)|M(a)> &=& \sum_{\ell } \sum_{n} U_{bn} \frac{ ( { \hat {\bar E}}^0_n - {\hat {\bar E}}^0_{a })( {\hat {\bar E}}^0_n - {\hat {\bar E}}^0_{b} ) }{ {\hat {\bar D^0}}[\ell] } U_{na}^* e^ { -i {\bar E}^0_\ell (t' - t) } ~. \eeq \end{widetext} The indices $(a, b, n)$ run over all permutations of the three integers $(1,2,3)$, from which it follows that $( { \hat {\bar E}}^0_n - {\hat {\bar E}}^0_{a[\ell] })( {\hat {\bar E}}^0_n - {\hat {\bar E}}^0_{b[\ell]} ) \equiv {\hat {\bar D^0}}[\ell]$, and, consequently, \beq &&<M(b)|S^0(t',t)|M(a)> \nonumber \\ &=& \sum_n U_{bn} e^{ -i E^0_n (t'-t) } U_{na}^* ~. \eeq This well-known vacuum limit also follows directly from more elementary considerations, using Eqs.~(,), \beq && <\nu^0_{fb}|e^{-i H_{0v} (t' - t)}|\nu^0_{fa}> \nonumber \\ &= & \sum_n U_{bn} e^{ -i E^0_n (t'-t) } U_{na}^* ~. \eeq The equality of Eqs.~(,) verifies our exact Hamiltonian formulation in the vacuum limit. Finally, we find that $F^{ab}_{\ell\ell'} $ appearing Eq.~() takes a familiar form in the vacuum. Using Eq.~(), or Eq.~(), \begin{widetext} \beq && \mathcal{P}(\nu^0_a \rightarrow \nu^0_b) = \left(\sum_\ell U_{b\ell} e^{-i {\bar E^0_\ell} (t' -t)} U^*_{\ell a} \right) \left(\sum_{\ell'} U_{b\ell'} e^{-i {\bar E^0_{\ell'} } (t' -t)} U^*_{\ell' a}\right)^* ~. \eeq Expressing this in terms of Freund's $J^{ab}_{\ell\ell'}$~, \beq J^{ab}_{\ell\ell'} &\equiv& U^*_{b\ell} U_{a\ell} U^*_{a\ell'} U_{b\ell'} ~, \eeq and using Eq.~(), \beq {\bar E}^0_{\ell} - {\bar E}^0_{\ell'} &=& \frac{m_\ell^2-m_{\ell'}^2}{2E} ~, \eeq we find that Eq.~() may be written, \beq && \mathcal{P}(\nu^0_a \rightarrow \nu^0_b) = \delta_{ab} - 4\sum_{i>j} Re J^{ab}_{ij} \sin^2{ ( ( {\bar E}^0_{\ell} - {\bar E}^0_{\ell'} )L )} - 2 \sum_{i>j} Im J^{ab}_{ij} \sin ( {2 ( {\bar E}^0_{\ell} - {\bar E}^0_{\ell'} )L ) } ~. \eeq \end{widetext} Comparing Eqs.~() and (), it is clear that $F^{ab}_{\ell\ell'} \rightarrow J^{ab*}_{\ell\ell'}$ in the vacuum limit. \section{The Standard Neutrino Model} We adopt the Standard Neutrino Model~ as our description of neutrino physics. The parameters defining the model include a (dimensionless) interaction strength ${\hat A}$ of neutrinos and anti-neutrinos with matter, the three neutrino mass differences, the three mixing angles, and the $CP$-violating phase. Most of the parameters of the SNM are consistent with global fits to neutrino oscillation data with relatively good precision. The neutrino mass differences of the SNM are taken to be~ \beq m^2_2-m^2_1 &\equiv& \delta m_{21}^2 \nonumber \\ &=& 7.6\times 10^{-5} ~{\rm eV}^2 \nonumber \\ m^2_3-m^2_1 &\equiv& \delta m_{31}^2 \nonumber \\ &=& 2.4\times 10^{-3} ~{\rm eV}^2 ~, \eeq corresponding to \beq \alpha &\equiv& \frac{\delta m_{21}^2}{\delta m_{31}^2} \nonumber \\ &=& 3.17\times 10^{-2} ~. \eeq In Ref.~, $\delta \equiv (m_2^2 - m_1^2)/(2E) $ and $\Delta = (m_3^2 - m_1^2)/(2E) $ whereas in Ref.~, ${\hat \Delta} = (m_3^2 - m_1^2)L/(4E)$ and $\Delta = m_3^2 - m_1^2$. The mixing angle $\theta_{23}$, \beq \theta_{23}&=& \pi/4 ~, \eeq is the best-fit value from Ref.~, and $\theta_{12}$, \beq \theta_{12}&=& \pi/5.4 ~, \eeq is consistent with the recent analysis of Ref.~. The mixing angle $\theta_{13}$ is known to be small ($\theta_{13}<0.18$ at the 95\ \beq \theta_{13}&=& 0.151 ~. \eeq This fixes $R_p \equiv \sin^2\theta_{13}/\alpha \approx 0.711$. The CP violating phase $\delta_{cp}$ is not known at all and will one of the major interests at future neutrino facilities. Parametrizing the interaction strength ${\hat A}$, Eq.~(), we find \beq \hat A &=& \pm 6.50 ~10^{-2} R ~E[{\rm GeV}] \rho[{\rm gm/cm}^3] ~, \eeq with E[GeV] being the neutrino beam energy $E$ (in GeV) and $\rho$[gm/cm$^3$], is the average total density (in ${\rm gm/cm}^3$) of matter through which the neutrino beam passes on its way to the detector (the matter having average proton-nucleon ratio $R$). For our calculations we are interested in experiments close to the earth's surface, so we take \beq \rho[{\rm gm/cm}^3] &=& \rho_0 \nonumber \\ &=& 3 ~, \eeq the approximate mean density of the earth's mantle. In the SNM, $\Delta_{L}$, defined in Eq.~(), may be parametrized in the high-energy limit as \beq \Delta_L &\approx& 3.05 \times 10^{-3} \frac { L[{\rm Km}] }{E[{\rm GeV}]} ~. \eeq Here $ L[{\rm Km}] $ is the baseline and $E[{\rm GeV}]$ is the neutrino beam energy. \section{ Eigenvalue Expansions in the SNM} The fact that $\alpha$ and $\sin\theta_{13}$ are naturally small in the SNM commonly motivates approximation schemes~ based on first-order Taylor series expansions in one of these small parameters, $\xi_i'$ (where $\xi_i'$ stands for $\alpha$ or $\sin^2{\theta_{13}})$. For example, in Refs.~ the oscillation probability is expanded in $\sin^2\theta_{13}$. Reference~ makes use of an expansion in the small parameter $\alpha$. Although these expansions may be used effectively to simplify the theory, they come at a price~. This price is that neither expansion gives accurate representations for all values of the interaction strength ${\hat A}$, including values in some regions of critical importance. In our subsequent work~, we will consider simplifying the oscillation probability using the same expansions, but they will be used somewhat differently, in two stages. In the first stage, the eigenvalues are expanded, as below. In the second, the expansions will be used to simplify the oscillation probability. In this work, the eigenvalue expansion will, of course be made before introducing $R_p$. For the expansion of the oscillation probabilities, $R_p$ may be introduced before the expansion is made because of its simpler analytic structure. \subsection {Analytic Structure of Eigenvalues} The key for identifying which of the $\xi_i'$-expansions might be appropriate over specifics ranges of ${\hat A}$ is revealed by the analytic structure of the eigenvalue $ { \hat {\bar E}}_{\ell}(\xi)$. It is particularly important to identify the locations of its branch points when $\xi_i'=0$. Branch points identify where a series expansion of $ { \hat {\bar E}}_{\ell}(\xi)$, or a function of it such as $<F(H_\nu)>$, would not converge. We have seen that $ { \hat {\bar E}}_{\ell}(\xi)$ depends on $\xi$ entirely through the two functions $(U(\xi),V(\xi))$, \beq { \hat {\bar E}}_{\ell} (\xi) &=& {\hat {\bar E}}_{\ell }(U(\xi),V(\xi)) ~, \eeq where \beq U(\xi ) &\equiv& Re[d^{1/3}] \nonumber \\ V(\xi ) &\equiv& Im[d^{1/3}] ~, \eeq with $d$ defined in Eq.~(). Recalling that $4\gamma^3 - \psi^2 > 0$, we conveniently write \beq U(\xi) &=& \frac{1}{2} (\psi + i \sqrt{4\gamma^3-\psi^2 })^{1/3} \nonumber \\ &+& (\psi -i \sqrt{4\gamma^3-\psi^2 })^{1/3} ~, \eeq and \beq V(\xi) &=& \frac{1}{2i} (\psi + i \sqrt{4\gamma^3-\psi^2 })^{1/3} \nonumber \\ &-& (\psi -i \sqrt{4\gamma^3-\psi^2 })^{1/3} ~. \eeq Branch points $\xi^B$ clearly occur for parameter values satisfying \beq \psi^2(\xi^B) &=& 4 \gamma^3(\xi^B) ~. \eeq The locations of the branch points $\xi^B$ of $ {\hat {\bar E}}_{\ell}(\xi)$ are found as follows for the two small-parameter choices $\xi'_i$. When $\xi'_i=\alpha$, branch points as a function of the other parameters of the Standard Neutrino Model are found from Eq.~() with $\psi(\xi^B) $ and $\gamma(\xi^B) $ evaluated in terms of \beq a(\xi^{B} ) &=& -(1+ {\hat A} ) \nonumber \\ b(\xi^{B}) &=& \hat A \cos^2\theta_{13} \nonumber \\ c(\xi^{B}) &=& 0 ~, \eeq where $\xi^{B} = \xi|_{\alpha = 0}$. The only real solution of Eq.~() is \beq {\hat A} &=& 0 ~. \eeq When $\xi'_i= \sin^2\theta_{13}$, branch points as a function of the other parameters of the SNM are found from Eq.~() with $\psi(\xi^B) $ and $\gamma(\xi^B) $ evaluated in terms of \beq a(\xi^{B}) &=& -(1+\hat A + \alpha) \nonumber \\ b(\xi^{B}) &=& \alpha + \hat A \alpha \cos\theta_{12} + \hat A \nonumber \\ c(\xi^{B}) &=& -\hat A \alpha \cos^2\theta_{12} ~, \eeq where $\xi^{B} = \xi|_{\theta_{13} = 0}$. The only real solution of Eq.~() is \beq {\hat A} &=& {\hat A}_0 \equiv \frac{1-\alpha }{1 - \alpha \cos^2\theta_{12}} \nonumber \\ &\approx& 1 ~. \eeq \subsection{ The First-order Taylor Expansions} Defining $\hat {\bar E}^{\xi}_\ell $ to be the first two terms of the Taylor series for $\hat {\bar E}_\ell $ expanded about $\xi=0$, we find \beq \hat {\bar E}^{\xi}_\ell &\equiv& \hat {\bar E}_\ell|_{\xi=0} + \xi \frac{\partial \hat {\bar E}_\ell }{ \partial \xi} |_{\xi=0} ~, \eeq where $ \hat {\bar E}_\ell $ and $\partial \hat {\bar E}_\ell /\partial \xi $ are easily obtained from Eqs.~(,) and their derivatives. The results for $\xi = \sin^2\theta_{13} $ and $\xi = \alpha $ found in this way are given immediately below. \subsubsection{ $\xi = \sin^2\theta_{13}$ } Applying Eq.~() we find for $\xi = \sin^2\theta_{13}$ and below the corresponding branch point, \beq {\hat {\bar E^{\theta } _{1} }} &=& \frac{1}{2 } ({\hat A} +\alpha - \hat C_\theta ) \nonumber \\ &+& \frac{ \alpha R_p {\hat A} } { 2(1-y) \hat C_\theta } (2 -{\hat A}_0 - \alpha - \hat C_\theta ) \nonumber \\ &+& \frac{ y \alpha R_p } { 2 \hat C_\theta } (2 - 4{\hat A}_0 + {\hat A}_0^2 -2 \alpha + 2 {\hat A}_0 \alpha) \nonumber \\ {\hat {\bar E^{\theta } _{2} }} &=& \frac{1}{2 } ({\hat A} +\alpha + \hat C_\theta ) \nonumber \\ &-& \frac{ \alpha R_p {\hat A} } { 2(1-y) \hat C_\theta } (2 -{\hat A}_0 - \alpha + \hat C_\theta ) \nonumber \\ &-& \frac{ y \alpha R_p } { 2 \hat C_\theta } (2 - 4{\hat A}_0 + {\hat A}_0^2 -2 \alpha + 2 {\hat A}_0 \alpha) \nonumber \\ {\hat {\bar E^{\theta}_{3} }} &=& 1 + \frac{ \alpha R_p \hat A }{1-y} ~, \eeq where $ y \equiv \hat A / \hat A_0$, with ${\hat A}_0 \equiv {\hat A}^\theta_0$ the location of the branch point for the $\sin^2\theta_{13} $ expansion and \beq \hat C_\theta &=& \sqrt{ {\hat A}^2 +\alpha^2 -2 {\hat A} \alpha \cos{ 2 \theta_{12} } } ~. \eeq Note that ${\hat {\bar E^{\theta } _{2} }} = {\hat {\bar E^\theta_{1} }}|_{ {\hat C}_\theta \to -{\hat C}_\theta }$. Above the branch point, $y>1$, ${\hat {\bar E^{\theta}_{3} }}$ and ${\hat {\bar E^{\theta}_{2} }}$ exchange roles. The lack of convergence at the branch point is manifest here by inspection, ${\it i.e. }$, by the appearance of simple pole at $y=1$. Although Eq.~() would suggest a pole in all three eigenvalues, ${\hat {\bar E^{\theta}_{1} }}$ is rather accurate for all values of $y$. The absence of a singularity in ${\hat {\bar E^{\theta}_{1} }}$ can be confirmed by a simple calculation that shows the coefficient of $(1-y)^{-1}$ vanishes at $y=1$. \subsubsection{$\xi = \alpha $} Taking $R_p = \sin^2\theta_{13}/\alpha$, we may write $\cos^2\theta_{13} = 1- \alpha R_p$ and $\cos 2\theta_{13} = 1- 2 \alpha R_p$. Then, for $\xi = \alpha $ and $y>0$ (above its corresponding branch point), \beq {\hat {\bar E^\alpha_1 }} &=& \alpha \cos\theta_{12}^2 \nonumber \\ {\hat {\bar E^\alpha_{2} }} &=& \frac{1}{2} ( 1 + {\hat A} - {\hat C}_\alpha ) \nonumber \\ &+& \alpha \frac{ \sin\theta_{12}^2} {2 } ( 1+ \frac{ 1 - {\hat A} (1- 2 \alpha R_p)} { {\hat C}_\alpha } ) \nonumber \\ {\hat {\bar E^\alpha_{3} }} &=& \frac{1}{2 } ( 1 + {\hat A} + {\hat C}_\alpha ) \nonumber \\ &+& \alpha \frac{ \sin\theta_{12}^2} {2 } ( 1- \frac{ 1 - {\hat A} (1- 2 \alpha R_p) } { {\hat C}_\alpha } )~, \eeq where \beq \hat C_\alpha &=& \sqrt{ (1 - {\hat A })^2 + 4 \alpha R_p \hat A } ~. \eeq Note that ${\hat {\bar E^\alpha_{2} }} = {\hat {\bar E^\alpha_{3} }}|_{ {\hat C}_\alpha \to -{\hat C}_\alpha }$. These expressions are identical to Eqs.~(18,19) of Freund~. In contrast to $\xi = \sin^2\theta_{13}$, the lack of convergence of the $\alpha$ expansion is not obvious from a casual examination of Eq.~(). Numerical comparison to the exact result confirms that ${\hat {\bar E^{\alpha}_{1} }}$ and ${\hat {\bar E^{\alpha}_{2} }} $ are poor representations of the corresponding exact results in the vicinity of the branch point at $y = 0$. However, no evidence of the branch point is apparent in ${\hat {\bar E^{\alpha}_{3}}}$, which is rather accurate for all values of $y$. Across the branch point at $y=0$, ${\hat {\bar E^{\alpha}_{1} }}$ and ${\hat {\bar E^{\alpha}_{2} }}$ exchange roles. \subsection { Branch Points and Resonances } Resonances are heralded by the appearance of minima in the EV differences. Two well-known neutrino resonances occur in the SNM. One of these, the ${\it solar}$ resonance, is found for relatively weak interaction strengths $\hat A \approx \alpha$. The other, the {\it atmospheric } resonance, occurs for stronger interactions, ${\hat A} \approx \cos{2\theta_{13}}$. The solar resonance occurs very close to the branch point identified with the $\alpha$ expansion, and the atmospheric resonance very close to the branch point identified with the $\sin^2\theta_{13}$ expansion. Using Eq.~(), the solar resonance appears for neutrinos of energy \beq E &=& E^{sol}[{\rm GeV}] \approx 0.325 \eeq for underground experiments in the earth's mantle ($ \rho[{\rm gm/cm}^3] = \rho_0 \approx 3$). We find, similarly, that for neutrinos of energy $E$, the solar resonance occurs in matter of density \beq \rho^{sol}[{\rm gm/cm}^3] \approx E[{\rm GeV}]^{-1} ~, \eeq taking $\alpha=0.0317$ from the SNM. Likewise, the atmospheric resonance is found for neutrinos of energy \beq E &=& E^{atm}[{\rm GeV}] \approx 9.79 ~, \eeq also for underground experiments in the earth's mantle. At energy $E$, it occurs at a density of \beq \rho^{atm}[{\rm gm/cm}^3] \approx 29.4 / E[{\rm GeV}] ~, \eeq taking $ \cos{2 \theta_{13} } = 0.955$ from the SNM. Because of the close correlation between the branch points and the resonances, there is also a close correlation between resonances and viable approximation schemes. Note, however, that the branch point in the $\alpha$ expansion near the solar resonance affects both neutrino and anti-neutrino scattering since its actual location is $\alpha =0$. \subsection{Discussion} We have seen in this section that expansions of the eigenvalues in a small parameter $\xi_i'$ of the SNM (where $\xi_i'$ stands for $\alpha$ or $\sin^2{\theta_{13}})$ must be made carefully, since the EV are not analytic everywhere. Nevertheless, as noted, first-order Taylor series representations of the EV are commonly used to simplify the theory. Because these expansions do not give an accurate representation the EV for all values of the interaction strength ${\hat A}$, it is important to identify the regions where the theoretical errors of the expanded EV might be acceptable and lead to accurate representations. We address this in the next section. \section{ Approximating $\mathcal{P}(\nu_i \rightarrow \nu_f) $ with expanded EV} In this section, we begin our assessment of common procedures used to simplify calculation of the oscillation probability by expanding it in one of the small parameters of the SNM. Freund observed~ that the $\alpha$-expansion, although useful, could not be used near the solar resonance where ${\hat A}=\alpha$. However, no understanding of the limitations of the $\sin^2\theta_{13}$-expansion appears in the literature. We have shown above that the applicability of both expansions is limited by the presence of branch points in the analytical structure of the eigenvalues. The branch point responsible for the failure of the $\alpha$-expansion is located at $\hat A = 0$ when $\alpha=0$, and the branch point responsible for the failure of the $\sin^2\theta_{13}$ expansion is located at $\hat A = \hat A_0$ when $\sin^2\theta_{13}=0$, where $\hat A_0$ is defined in Eq.~(). We make our assessment numerically, comparing the oscillation probability calculated from Eqs.~(,) using the exact eigenvalues to that calculated from Eqs.~(,) using eigenvalues expanded in one of the small parameters of the SNM. \subsection { Assessing Oscillation Probabilities Expressed in terms of $\xi$-expanded Eigenvalues } Theoretical errors characterizing an approximation scheme may emerge numerically only from an examination of the dependence of $P^{e\mu}(\Delta_L,{\hat A})$ on $\Delta_L$ and $\hat A$. In this section we discuss how we will do this. It is convenient to discuss the oscillation pattern in terms of the location of the first maxima of the functions $ \Delta_L j_0 ({\hat \Delta}_\ell)$ appearing in the expressions for the partial oscillation probabilities. These peaks occur at $\Delta_L = \Delta_L^{(\ell)}$, with \beq \Delta_L^{(\ell)} &=& \frac{\pi} {2 \Delta {\hat{\bar E}} [\ell] } \eeq closely related to the period $ P_\ell $ of $ j_0 ({\hat \Delta}_\ell)$, \beq P_\ell &\equiv& 4 \Delta_L^{(\ell)} = \frac{2\pi}{ \Delta {\hat{\bar E}}[\ell] } ~. \eeq Because the exact eigenvalues never cross, the ordering of $ {\hat{\bar E}}_\ell$ is the same as it is in the vacuum, namely $ {\hat{\bar E}}_3 > {\hat{\bar E}}_2 > {\hat{\bar E}}_1 $. It can inferred from this that all $\Delta {\hat {\bar E}}[\ell]$ remain positive, and, in addition, that \beq \Delta {\hat {\bar E}}[2] &>& \Delta {\hat {\bar E}}[3] >0 \nonumber \\ \Delta {\hat {\bar E}}[2] &>& \Delta {\hat {\bar E}}[1] > 0 ~. \eeq We see, in general, that for small ${\hat A}$, $\Delta {\hat {\bar E}}[1] > \Delta {\hat {\bar E}}[3]$, and, for large ${\hat A}$, $\Delta {\hat {\bar E}}[2] > \Delta {\hat {\bar E}}[3]$. It is also clear that $\Delta {\hat {\bar E}}[2] $ is always the largest eigenvalue difference. Consequently, $P_2$ is always the smallest of the three periods, thus characterizing the most rapidly varying Bessel function. The relative sizes of $P_\ell$ are easily worked out in specific cases. In the vacuum, \beq P_1 &=& 2 \pi/(1-\alpha) \nonumber \\ P_2 &=& 2 \pi \nonumber \\ P_3 &=& 2 \pi / \alpha ~, \eeq evaluated from differences of the vacuum eigenvalues appearing in Eq.~(). In the SNM, we find that \beq P_3 &>& P_1 ~\rm{for}~ \hat A < \hat A_{2} \nonumber \\ P_1 &>& P_3 ~\rm{for}~ \hat A > \hat A_{2}~, \eeq where \beq \hat A_{2} &=& 0.538 \eeq is the value of $\hat A$ at which $P_3 = P_1$ ($\Delta {\hat {\bar E}}[1] = \Delta {\hat {\bar E}}[3]$). With $\Delta {\hat {\bar E}}[2]$ the largest eigenvalue difference, $P_2$ is always the smallest of the three periods. Thus, the value of $\Delta_L$ at the first peak of $\Delta_L j_0({\hat \Delta}_2)$ is a natural scale. \subsection {Regions of Maximum Sensitivity } Sensitivity to the approximation scheme should be most manifest within regions where all three Bessel functions of $P^{ab}( \Delta_L,{\hat A})$ are of similar size and interfere. This will happen once $\Delta_L$ becomes comparable to the first peak of its most slowly varying $ \Delta_L j_0 ({\hat \Delta}_\ell)$, which occurs at either $\Delta_L = \Delta_L^{(3)}$ or $\Delta_L = \Delta_L^{(1)}$. In general, the sensitivity to approximations increases as the distance $\Delta_L$ to the most distant peak increases. In the vacuum, the first peak of the most slowly varying $ \Delta_L j_0 ({\hat \Delta}_\ell)$ is always at $\Delta_L = \Delta_L^{(3)}$, readily established by the vacuum eigenvalue differences given in Eq.~(). The corresponding baseline is obtained from Eqs.~(,), $L^{(3)}[{\rm Km}] \approx 5170 \pi E[{\rm GeV}]$. In the SNM, the most slowly varying Bessel function, established from Eqs.~(,), is $ \Delta_L j_0 ({\hat \Delta}_3)$ when ${\hat A}< \hat A_{2}$ and $ \Delta_L j_0 ({\hat \Delta}_1)$ for ${\hat A}> \hat A_{2}$. \subsubsection {Regions of Maximum Sensitivity for Fixed $\hat A$ } Consider first the variation of $P^{(ab)}(\Delta_L, \hat A)$ with $\Delta_L$ for a given value of $\hat A$ in the SNM. According to Eq.~(), for $\hat A > \hat A_2$, $ \Delta_L j_0 ({\hat \Delta}_1)$ is the most slowly varying (having the larger period), and the region of maximum sensitivity is \beq \Delta_L &>& \Delta_L^{(1)} \equiv \frac{\pi} {2 \Delta {\hat{\bar E}} [1] } \eeq with $\Delta {\hat{\bar E}} [1]$ evaluated at $\hat A$. Similarly, for $\hat A < \hat A_2$, $ \Delta_L j_0 ({\hat \Delta}_3)$ is the most slowly varying. According to Eq.~(), its peak occurs where $ \Delta_L = \pi/(2 \Delta {\hat{\bar E}} [3] )$. Thus, the region of maximum sensitivity is \beq \Delta_L &>& \Delta_L^{(3)} \equiv \frac{\pi} {2 \Delta {\hat{\bar E}} [3] } ~. \eeq with $\Delta {\hat{\bar E}} [3]$ again evaluated at $\hat A$. \subsubsection {Regions of Maximum Sensitivity for Fixed $\Delta_L$ } Consider next the variation of $P^{(ab)}(\Delta^0_L, \hat A)$ with $\hat A$ for given $\Delta^0_L$ in the SNM. According to Eq.~(), for $\hat A > \hat A_2 $, $ \Delta_L j_0 ({\hat \Delta}_1)$ is the most slowly varying. According to Eq.~(), its peak occurs where $ \Delta^0_L = \pi/(2 \Delta {\hat{\bar E}} [1])$. Thus, the region of maximum sensitivity is \beq \Delta {\hat{\bar E}} [1] &>& \frac {\pi }{2 \Delta^0_L } \eeq with $\Delta {\hat{\bar E}} [1]$ evaluated at $\hat A$. Similarly, for $\hat A < \hat A_2 $, we find from Eq.~() that $ \Delta_L j_0 ({\hat \Delta}_3)$ is the most slowly varying, and the region of maximum sensitivity is \beq \Delta {\hat{\bar E}} [3] &>& \frac {\pi }{2 \Delta^0_L } ~. \eeq \section{ Numerical Study of $\mathcal{P}(\nu_e \rightarrow \nu_\mu)$ with expanded EV} Our main interest in the present section is to map out the regions where each of the small-parameter expansions is capable of simplifying $P^{ab}( \Delta_L,{\hat A})$. We do this by comparing three calculations using $\mathcal{P}(\nu_e \rightarrow \nu_\mu)$ taken from our exact Hamiltonian formulation. One of these is a calculation of the exact oscillation probability obtained in our Hamiltonian formulation. For this we use the expressions in Eqs.~(,) evaluated with the exact EV. The other two are calculations of our $\xi$-expanded oscillation probability for each of the small parameters of the SNM. For these we evaluate Eqs.~(,) using the $\xi$-expanded EV. For $\xi=\alpha$, we use the $\alpha$-expanded EV given in Eq.~(), and for $\xi=\sin^2\theta_{13}$ we use the $\sin^2\theta_{13}$-expanded EV given in Eq.~(). The calculation with the $\xi$-expanded EV would, of course, coincide with the exact calculation in the vacuum. Differences therefore reflect medium effects. The extent to which our oscillation probability evaluated with one of the $\xi$-expanded EV agrees with the exact result indicates regions in which it may be possible to obtain, at least in principle, a simple $\xi$-expanded expression for the oscillation probability in good agreement with the exact result. In a subsequent paper~, we make a similar comparison between the exact oscillation probability and the approximate ones given in Refs.. From the numerical results we obtain in Ref.~ and the present paper, we will be able to identify regions in which both (1) the exact oscillation probability and the results of Refs. are in poor agreement; and, (2) the exact oscillation probability and the $\xi$-expanded result are in excellent agreement. The regions where both of these conditions are satisfied indicate where it might be possible to improve the results found in Refs. using our Hamiltonian formulation. We explore this possibility in yet another paper~. Equations~(,) provide a means to extrapolate the results in any figure to a variety of baseline values, medium properties, and neutrino energies within the particular regions shown in that figure. In the earth's mantel, where the average $Z/N=1/2$, Eqs.~(,) become \beq L[{\rm Km}] &=& 1.08\times 10^4 \frac{ {\hat A} \Delta_L }{ \rho[{\rm gm/cm}^3] } \nonumber \\ E[{\rm GeV}] &=& 30.8 \frac{ {\hat A}}{ \rho[{\rm gm/cm}^3] } ~. \eeq \subsection { $\Delta_L$-Dependence of $\mathcal{P}(\nu_e \rightarrow \nu_\mu)$ } We begin our exploration of the extent to which a particular $\xi$-expansion is capable of simplifying the oscillation probability by examining $P^{(e\mu)}(\Delta_L, \hat A)$ vs $\Delta_L$ for particular values of $\hat A$. One value of $\hat A$ is chosen near the solar resonance and another near the atmospheric resonance. For each choice of $\xi$ and $\hat A$, we compare the exact result to Eq.~(). We first examine $\mathcal{P}(\nu_e \rightarrow \nu_\mu)$ below the solar resonance, at ${\hat A} = 0.0102$. For this value of $\hat A$, $\Delta {\hat{\bar E}} [3] \approx 0.0294$, Eq.~() specifies that the approximate oscillation probability becomes sensitive to approximations for $\Delta_L > \Delta_L^{(3)} \approx 53$. Our calculations for ${\hat A} = 0.0102$ are shown in Fig.~. We see from this figure that the $\alpha$-expanded oscillation probability begins to departs from the exact result at large $\Delta_L$ meaning, as expected, that the $\alpha$ expansion breaks down in the vicinity of the solar resonance. The sensitivity to medium effects shows up already for $\Delta_L \approx 20$, which is smaller than $\Delta_L^{(3)} \approx 53 $ estimated using Eq.~(). On the other hand, Eq.~() evaluated with eigenvalues expanded to first order in $\sin^2\theta_{13}$ agree well with the exact result at large $\Delta_L$ showing that the $\sin^2\theta_{13}$-expanded oscillation probability is capable of providing an excellent approximation in the vicinity of the solar resonance. For this small value of ${\hat A}$, we find that position of the first peak of the exact oscillation probability, $\Delta_L \approx 1.58$, coincides almost exactly with the location of the peak of the most rapidly varying Bessel function, $j_0({\hat \Delta}_2)$. From Eq.~(), we note that the oscillation probability at $\Delta_L=20$ in Fig.~, where the approximate calculation begins to break down in the solar resonance region with the $\alpha$-expanded EV, would correspond to a measurement of $105$ MeV neutrinos propagating in the earth's mantel at a baseline $734$ Km. We next examine $\mathcal{P}(\nu_e \rightarrow \nu_\mu)$ for ${\hat A} = 0.8$, a value of ${\hat A}$ near the atmospheric resonance. Taking the exact eigenvalue difference $\Delta {\hat{\bar E}} [1] \approx 0.328$ at ${\hat A} = 0.8$, Eq.~() specifies that the desired sensitivity of the oscillation probability to approximations should become apparent at $\Delta_L \approx 4.8$. Our calculations for ${\hat A} = 0.8$ are shown in Fig.~. The exact result shown in Fig.~ (solid curve) begins to differ from the $\sin\theta_{13}$-expanded result (long-dashed curve) at $ \Delta_L \approx 2$, which occurs a bit before $ \Delta_L \approx 4.8$, where all three Bessel functions fully contribute. Because the long-dashed curve begins to depart from the solid curve at large $\Delta_L$, these results confirm that $P^{e\mu}(\Delta_L,{\hat A})$ evaluated with $\sin^2\theta_{13}$-expanded EV breaks down near the atmospheric resonance with medium effects included. The failure of the $\sin\theta_{13}$-expansion becomes increasingly apparent as $\Delta_L$ increases to larger $\Delta_L$. The fact that the exact result (solid curve) and $\alpha$-expanded result (medium-dashed curve) seem to completely overlap demonstrates that $P^{e\mu}(\Delta_L,{\hat A})$ evaluated with the $\alpha$-expanded EV is capable of becoming a completely acceptable approximation near the atmospheric resonance. For this larger value of ${\hat A}$, the first peak of the oscillation probability, at $\Delta_L \approx 1.58 $, nearly coincides with the peak of the most rapidly varying Bessel function, $j_0({\hat \Delta}_2)$, at $\Delta_L \approx 1.5$. Applying Eq.~(), we see that the oscillation probability at the value of $\Delta_L$ where medium effects begin to become apparent in Fig.~ would correspond to neutrinos of energy $E[{\rm GeV}] \approx 4.5$ propagating in matter of density similar to the average density of the entire earth, $\rho[{\rm gm/cm}^3] = 5.52$ at a baseline of $7830$ Km (for comparison, the average earth radius is $6370$ Km). \subsection{${\hat A}$-Dependence of $\mathcal{P}(\nu_e \rightarrow \nu_\mu)$ } We next compare oscillation probabilities over various ranges of $\hat A$. For a given range of $\hat A$, whether or not the three Bessel functions maximally interfere depends on the choice of $\Delta_L$, which is determined by Eqs.~(,) depending on whether $\hat A > \hat A_2$ or $\hat A < \hat A_2$, respectively. \subsubsection{ $0< \hat A < 0.2$ } Numerical studies using Eqs.~(,) show that for $0 <{\hat A} < 0.2 $ taking $\Delta_L=60$ is sufficient to ensure that the three Bessel functions maximally interfere. In Fig.~, we compare the oscillation probabilities in this region. These results confirm that the expansion in $\sin^2\theta_{13}$ is a reasonably good approximation within the solar resonance region, whereas the expansion in $\alpha$ is evidently not. \subsubsection{ $0.2 < \hat A < \hat A_2$ } Numerical studies using Eqs.~(,) show that for $0 <{\hat A} < \hat A_2 $ taking $\Delta_L=10$ is sufficient to ensure that the three Bessel functions maximally interfere. In Fig.~, we compare the oscillation probabilities in this region. These results confirm that both the expansion in $\sin^2\theta_{13}$ and the expansion in $\alpha$ are reasonably good approximations here. Results shown here apply over the same range of neutrino energy, baselines, and medium properties as those given in Eqs.~(). The extrapolation applies, of course, only within the region $0.2 <{\hat A} < \hat A_2$. \subsubsection{ $\hat A_2 < \hat A < 0.8$ } Numerical studies using Eqs.~(,) show that for $\hat A_2 <{\hat A} < 0.8 $ taking $\Delta_L=4$ is sufficient to ensure that the three Bessel functions maximally interfere. In Fig.~, we compare the oscillation probabilities over this region. These results confirm that the expansion in $\alpha$ is a reasonably good approximation below the atmospheric resonance. The onset of the failure of the $\sin^2\theta_{13}$-expanded EV near the atmospheric resonance begins to become visible for ${\hat A} > \hat A_2$. Results shown here apply over the same range of neutrino energy, baselines, and medium properties as those given in Eqs.~(). The extrapolation applies, of course, only for $\hat A_2 <{\hat A} < 0.8$. \subsubsection{ $\hat 0.8 < \hat A < 1.2 $ } Numerical studies using Eqs.~(,) show that for $\hat 0.8 <{\hat A} < 1.2 $ taking $\Delta_L=4$ is sufficient to ensure that the three Bessel functions maximally interfere. In Fig.~, we compare the oscillation probabilities over this region. These results confirm that the expansion in $\alpha$ is a reasonably good approximation across the atmospheric resonance. The $\sin^2\theta_{13}$-expanded oscillation probability is not shown because it fails here. Results shown here apply over the same range of neutrino energy, baselines, and medium properties as those given in Eqs.~(). The extrapolation applies, of course, only for $0.8 <{\hat A} < 1.2$. \subsubsection{ $1.2 < \hat A < 2.5$ } Numerical studies using Eqs.~() show that for $1.2 < {\hat A} < 2.5$ taking $\Delta_L = 4$ is sufficient to ensure that the three Bessel functions maximally interfere. In Fig.~, we compare the oscillation probabilities over this region. These results show that $P^{e\mu}(\Delta_L,{\hat A})$ evaluated with the $\alpha$-expanded EV is relatively accurate here. The $\sin^2\theta_{13}$-expanded oscillation probability agrees with the exact result reasonably well for $\hat A > 1.6$. From Eq.~(), we see that with the value $\Delta_L=4$, and in matter of mean density similar to that of the average density of the entire earth, $\rho[{\rm gm/cm}^3] = 5.52$, the breakdown of the $\sin^2\theta_{13}$-expanded EV starts to becomes less visible for ${\hat A} > 1.2$, corresponding to neutrinos of energy $E[{\rm GeV}] \approx 6.2$ and baselines greater than $ L[{\rm Km}] = 23,500$ (considerably larger than the diameter of the earth). \subsubsection{$-0.5 < \hat A < 0.5$ } In Fig.~, we compare the oscillation probability calculated with the exact eigenvalues to the oscillation probability calculated with the expanded eigenvalues for $-0.5<{\hat A} < 0.5$, taking $\Delta_L=35$. Figure~ confirms the earlier observations that the expansion in $\sin^2\theta_{13}$ is a valid approximation within the solar resonance region, whereas the expansion in $\alpha$ is not for both neutrinos and anti-neutrinos. \subsection{ Intrinsic Limitations } It is important to note that when using expanded EV, the range of validity of the oscillation probability $P^{ab}( \Delta_L ,{\hat A})$ is limited even when it is evaluated using the convergent eigenvalue expansion. This arises because the error in the phase of the trigonometric functions $\sin(\Delta_L \Delta {\hat{\bar E}}^\xi[\ell])$ grows with $\Delta_L $. Eventually, with increasing $\Delta_L $, the error in the phase will lead to an unacceptable error in $P^{ab}( \Delta_L ,{\hat A})$. When spurious effects of this type would show up from a comparison of the exact partial oscillation probability evaluated with the exact ${\hat{\bar E}}_\ell$ to the exact partial oscillation probability evaluated with the convergent expansion of ${\hat{\bar E}}_\ell$. A divergence between the calculated results of these two calculations herald the limit of $\Delta_L$ beyond which the use of expanded eigenvalues breaks down for $P^{ab}( \Delta_L ,{\hat A})$. It is clear from Fig.~ that the region of validity for both the $\sin^2\theta_{13}$-expansion and the $\alpha$-expansion extend out as far as $\Delta_L = 35$. \subsection{Discussion} The results of this section suggest a natural division of the full range of $\hat A$ into regions. One significant region, which we call the ${\it solar }$ resonance region, covers the range $0 < \hat A < 0.2 $. The solar-resonance region contains, of course, the solar resonance at $\hat A = \alpha$. Another is the atmospheric resonance region containing the atmospheric resonance at $\hat A \approx \cos 2\theta_{13}$. This region covers the interval $\hat A_2 < {\hat A} < 2.0$. The region between the solar and atmospheric regions, $0.2 < {\hat A} < \hat A_2 $, is a ${\it transition}$ region in the sense that within it the expansion in $\alpha$ is improving rapidly and the expansion in $\sin^2 \theta_{13}$ rapidly deteriorating. Within this region, the EV expansions in $\sin^2 \theta_{13}$ and $\alpha$ are of comparable accuracy. The ${\it asymptotic}$ region covers the interval ${\hat A} > 2.0$, where the EV's are approaching their asymptotic behavior. In future work, we will make a comparison of the our oscillation probability evaluated with the exact eigenvalues to the approximate oscillation probability given in Refs.~. By comparing those results to our oscillation probability evaluated with $\xi$-expanded EV found in the present paper, we will be able to identify regions where our Hamiltonian formulation might lead to more effective approximations. Based on this information, we will subsequently present any new approximate results we believe be helpful. \section{ Summary and Conclusions} We have presented exact, closed-form expressions for the neutrino oscillation probabilities in matter using our Hamiltonian formulation within the framework of the Standard Neutrino Model assuming 3 Dirac Neutrinos. Our goal is to benchmark approximate formulations having known difficulties arising from expansions commonly used to model neutrino and anti-neutrino experiments envisioned for future neutrino facilities. We have shown explicitly that for small $\alpha$ and $\sin\theta_{13}$ there are branch points in the analytic structure of the eigenvalues that lead to singular behavior of expansions near the solar and atmospheric resonance. The numerical calculations presented indicate regions in which the small-parameter expansions are guaranteed fail and should be avoided in practice. \appendix \section{ Time-evolution operator $S(T)$ in Our Hamiltonian Formulation} Our Hamiltonian formulation was presented in Sect.~. In this appendix, and in accord with the discussion in Sect.~, we give explicit analytic formulae for the time-evolution operator $ S(L) = e^{-i {\hat {\bar H}}_\nu \Delta_L} $ in terms of ${\hat {\bar E}}_\ell $ and the parameters defining ${\hat {\bar H}}_{\nu}$. \subsection{ Expression for $S(T)$ in terms of $ F^{ab}_{\ell} $ } Introducing dimensionless variables, the time-evolution operator $ S(T)$ in Eq.~() may be written, \beq S^{ab}(L) &=& \sum_{\ell} F^{ab}_{\ell} \cos{ 2{\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& i \sum_{\ell} F^{ab}_{\ell} \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~. \eeq Here $ F^{ab}_{\ell} $, defined in Eq.~(), may be found by evaluating Eq.~() at $L=0$ and obeys the normalization condition \beq \delta_{ab} &=& \sum_{\ell} F^{ab}_{\ell} ~. \eeq Using the identity $\cos{ 2 \beta } \equiv 1- 2\sin^2{ \beta}$, we write \beq S^{ab}(L) &=& \sum_{\ell} F^{ab}_{\ell} ( 1 - 2 \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } ) \nonumber \\ &-& i \sum_{\ell} F^{ab}_{\ell} \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &=& \sum_{\ell} F^{ab}_{\ell} - 2 \sum_{\ell} F^{ab}_{\ell} \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& i \sum_{\ell} F^{ab}_{\ell} \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~. \eeq Now, using the normalization, Eq.~(), \beq S^{ab}(L) &=& \delta_{ab} - 2 \sum_{\ell} F^{ab}_{\ell} \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& i \sum_{\ell} F^{ab}_{\ell} \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~. \eeq Thus, for calculating $ P^{ab}(L) $ in Eq.~(), \beq Re[ S^{ab}(t',t) ] &=& \delta_{ab} - 2 \sum_{\ell} Re[ F^{ab}_{\ell} ] \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &+& \sum_{\ell} Im [F^{ab}_{\ell} ] \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ Im[ S^{ab}(t',t) ] &=& - 2 \sum_{\ell} Im[ F^{ab}_{\ell} ] \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& \sum_{\ell} Re[ F^{ab}_{\ell} ] \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~. \eeq \subsection{ Analytic Expressions for ${\hat {\bar W}}_{0}[\ell]$, ${\hat {\bar W}}_{sin}[\ell]$, and ${\hat {\bar W}}_{cos}[\ell]$ in terms of $H_\nu$} In this appendix we give exact, analytic expressions the $S(T)$ in terms of ${\hat {\bar E}}_\ell $ and the parameters defining ${\hat {\bar H}}_{\nu}$. We begin with Eq.~(), which expresses $ F^{ab}_{\ell}$ in terms of an operator ${\hat {\bar W}}[\ell]$ and most easily accomplish our goal by making the following rearrangement of terms in ${\hat {\bar W}}[\ell]$, \beq {\hat {\bar W}}[\ell] &=& U {\hat {\bar H}}^2_{0v} U^{\dag} - U {\hat {\bar H}}_{0v}U^{\dag}~ \Sigma[\ell] \nonumber \\ &+& U {\hat {\bar H}}_{0v} U^{\dag} {\hat V} + {\hat V} U {\hat {\bar H}}_{0v} U^{\dag} \nonumber \\ &+& {\hat V}^2 - {\hat V}~ \Sigma[\ell] + {\bf 1}~ \Pi[\ell] ~. \eeq Here, $\Sigma[\ell]$ is the sum, and $\Pi[\ell]$ the product, of two EV as defined in Eq.~(,). It is useful to recall that $\Sigma[\ell]$ and $\Pi[\ell]$, defined in this fashion, depend on $\ell$ entirely through ${\hat {\bar E}}_\ell$. In particular~, from Eq.~(), \beq \Sigma[\ell] &=& -a - {\hat {\bar E}}_\ell ~, \eeq and, from Eq.~(), \beq \Pi[\ell] &=& b + a {\hat {\bar E_\ell}} + {\hat {\bar E_\ell}}^2 ~, \eeq with $a,b$ given in Eq.~(). Explicit expressions for ${\hat {\bar W}}[\ell]$ are easily found in terms of $H_\nu$ using Eq.~(), from which it follows that its entire dependence on $\delta_{cp}$ occurs through three operators, $W_0[\ell]$, $W_{cos}[\ell]$ and $W_{sin}[\ell]$, \beq {\hat {\bar W}}[\ell] &=& {\hat {\bar W}}_0[\ell] + \cos{\delta_{cp}} {\hat {\bar W}}_{cos}[\ell] \nonumber \\ &+& i \sin{\delta_{cp}} {\hat {\bar W}}_{sin}[\ell] ~. \eeq with ${\hat {\bar W}}_0[\ell]$, ${\hat {\bar W}}_{cos}[\ell]$ and ${\hat {\bar W}}_{sin}[\ell]$ real and independent of $\delta_{cp}$. Additionally, ${\hat {\bar W}}_{sin}[\ell]$ anti-symmetric, whereas ${\hat {\bar W}}_{cos}[\ell]$ and ${\hat {\bar W}}_0[\ell] $ are symmetric, under exchange of initial and final states. It follows that $Tr{\hat {\bar W}}_{cos}[\ell]=Tr{\hat {\bar W}}_{sin}[\ell]=0$, and therefore \beq Tr {\hat {\bar W}}[\ell] &=& Tr {\hat {\bar W}}_0[\ell] ~. \eeq Then, from Eq.~() and the definition of $F^{ab}_{\ell}$ in Eq.~(), we find the trace relationship \beq Tr F_\ell &=& \frac{ Tr {\hat {\bar W}}[\ell] }{ {\hat {\bar D}}[\ell] } = 1 ~. \eeq \subsubsection{ Analytic Expressions for ${\hat {\bar W}}_{0}[\ell]$ } For flavor changing transitions, $\nu_i \to \nu_j$, the matrix elements of ${\hat {\bar W}}_0[\ell]$ are found to have the structure \beq <M(\mu)| {\hat {\bar W}}_0[\ell] |M(e)> &\equiv& {\hat {\bar W}}^{12}_0[\ell] = ~c^{(12)}[\ell] \cos{\theta_{23}} \nonumber \\ <M(\tau)|{\hat {\bar W}}_0[\ell] |M(e)> &\equiv& {\hat {\bar W}}^{13}_0[\ell] = - c^{(12)}[\ell] \sin{\theta_{23}} \nonumber \\ <M(\tau)|{\hat {\bar W}}_0[\ell] |M(\mu)> &\equiv& {\hat {\bar W}}^{23}_0[\ell] \nonumber \\ &=& ~c^{(23)}[\ell] \sin{2\theta_{23}} ~. \eeq We use the convention that a quantity $O^{(ab)}$ written {\it with} parentheses enclosing $ab$ is a number, not a matrix element. As noted, $O^{ab}$ written {\it without} parentheses surrounding $ab$ is a matrix element. This distinction may be obvious in the present context, but later on this distinction may not be so obvious. Note that ${\hat {\bar W}}^{e\mu}_0[\ell]$ and ${\hat {\bar W}}^{e\tau}_0[\ell] $ are not independent. The corresponding coefficients $ c^{(ii)}[\ell] $ of for flavor preserving transitions are defined with ${\hat {\bar E}} _{\ell}^2 $ in Eq.~() separated out, \beq <M(i)| {\hat {\bar W}}_0[\ell] |M(i)> &\equiv& {\hat {\bar W}}^{ii}_0[\ell] \nonumber \\ &=& {\hat {\bar E}}_{\ell}^2 + c^{(ii)}[\ell] ~. \eeq The matrix for ${\hat {\bar W}}_0[\ell]$ is thus \begin{widetext} \beq {\hat {\bar W}}_{0}[\ell] &=& {\bf 1}\times {\hat {\bar E}} _{\ell}^2 + \left( \begin{array}{ccc} c^{(11)} [\ell] & c^{(12)} [\ell] \cos\theta_{23} &- c^{(12)} [\ell] \sin\theta_{23} \\ c^{(12)} [\ell] \cos\theta_{23} & c^{(22)} [\ell] & c^{(23)} [\ell] \sin {2 \theta_{23} } \\ -c^{(12)} [\ell] \sin\theta_{23} & c^{(23)} [\ell] \sin {2 \theta_{23} } & c^{(33)}[\ell] \end{array} \right ) ~. \eeq Writing the dependence of $ c^{(ij)} [\ell] $ on ${\hat {\bar E}} _{\ell}$ explicitly, \beq c^{(ij)} [\ell] &=& c_0^{(ij)} + c_1^{(ij)} ~{\hat {\bar E}} _{\ell} ~, \eeq and expressing the results in terms of the combinations of mixing angles \beq C_1^{(\pm)} &\equiv& \cos^2{ \theta_{12}} \cos^2{ \theta_{23}} \pm \sin^2{ \theta_{23}} \sin^2{ \theta_{12}} \sin^2{ \theta_{13}} \nonumber \\ C_2^{(\pm)} &\equiv& \cos^2{ \theta_{12}} \pm \sin^2{ \theta_{12}} \sin^2{ \theta_{13}} ~, \eeq where $C_2^{(\pm)}$ was defined earlier in defined in Eq.~(), we find the following exact results. For the flavor-changing transitions, \beq c^{(12)}_0 &=& - c^{(12)}_1 = -\frac{\alpha}{2} \cos{\theta_{13} } \sin { 2 \theta_{12} } \nonumber \\ c^{(23)}_0 &=& - \frac{\alpha}{2} ( \sin^2{\theta_{12} } - \cos^2{\theta_{12} } \sin^2{\theta_{13}} ) - \frac{{\hat A} }{2} (\cos^2{ \theta_{13}} - \alpha C_2^{(-)} ) \nonumber \\ c^{(23)}_1 &=& \frac{1}{2} ( \cos^2{\theta_{13} } - \alpha C_2^{(-)} ) ~; \eeq and, for the flavor-preserving transitions, \beq c_0^{(11)} &=& \alpha \cos^2{\theta_{12}} \cos^2{\theta_{13}} \nonumber \\ c_1^{(11)} &=& -\cos^2{\theta_{13}} - \alpha C_2^{(+)} ~, \eeq \beq c^{(22)}_0 &=& \alpha (1 - \sin^2{\theta_{23}} \cos^2{\theta_{13}} - C_2^{(+)}) +{\hat A} ( \cos^2{\theta_{13}} \cos^2{\theta_{23}} + \alpha (C_2^{(+)}- C_1^{(+)})) \nonumber \\ c^{(22)}_1 &=& -1 + \cos^2{\theta_{13}}\sin^2{\theta_{23}} - \alpha (1- C_1^{(+)}) - {\hat A} ~. \eeq The coefficients $ c^{(33)} $ follow from $ c^{(22)}$ by the exchange $\sin{\theta_{23} } \leftrightarrow \cos{\theta_{23} }$. Consequently, there are only four independent matrix elements for ${\hat {\bar W}}_0[\ell]$, two diagonal and two off-diagonal. With the eigenvalues independent of $\theta_{23}$, it follows that \beq {\hat {\bar W}}^{\tau\tau}_0[\ell] &=& {\hat {\bar W}}^{\mu\mu}_0[\ell] |_{ \sin{\theta_{23} } \leftrightarrow \cos{\theta_{23} } } ~. \eeq \subsubsection{ Analytic Expressions for ${\hat {\bar W}}_{sin}[\ell]$ and ${\hat {\bar W}}_{cos}[\ell]$ } The matrices for ${\hat {\bar W}}_{sin}[\ell]$ and ${\hat {\bar W}}_{cos}[\ell]$ have the following simple structure \beq {\hat {\bar W}}_{sin}]\ell] &=& \left( \begin{array}{ccc} 0 & -a^{(12)}[\ell] \sin\theta_{23} & -a^{(12)}[\ell] \cos\theta_{23} \\ a^{(12)}[\ell] \sin\theta_{23} & 0 & a^{(23)}[\ell] \\ a^{(12)}[\ell]\cos\theta_{23} & - a^{(23)}[\ell] & 0 \end{array} \right ) \eeq and \beq {\hat {\bar W}}_{cos} [\ell] &=& \left( \begin{array}{ccc} 0 & a^{(12)}[\ell]\sin\theta_{23} & a^{(12)}[\ell]\cos\theta_{23} \\ a^{(12)}[\ell]\sin\theta_{23} & - a^{(23)}[\ell]\sin 2\theta_{23} & - a^{(23)}[\ell] \cos 2\theta_{23} \\ a^{(12)}[\ell] \cos\theta_{23} & - a^{(23)}[\ell] \cos 2\theta_{23} & a^{(23)}[\ell] \sin 2\theta_{23} \end{array} \right ) ~, \eeq \end{widetext} where \beq a^{(23)}[\ell] &=& a^{(23)}_0 + a^{(23)}_1 ~ {\hat {\bar E}} _{\ell} \nonumber \\ a^{(12)}[\ell] &=& a^{(12)}_0 + a^{(12)}_1 ~ {\hat {\bar E}} _{\ell} ~. \eeq We find the exact results \beq a^{(23)}_0 &=& -\frac{ {\hat A} \alpha }{2} \sin{2 \theta_{12} } \sin{ \theta_{13} } -\frac{ \alpha}{2} \sin{2 \theta_{12} } \sin{ \theta_{13} } \nonumber \\ a^{(23)}_1 &=& \frac{\alpha }{2} \sin{2 \theta_{12} } \sin{\theta_{13} } \nonumber \\ a^{(12)}_0 &=& - \frac{\alpha }{2} \cos^2{\theta_{12} } \sin{2 \theta_{13} } \nonumber \\ a^{(12)}_1 &=& \frac{1}{2} \sin{2 \theta_{13} } - \frac{\alpha}{2} \sin^2 {\theta_{12} } \sin{2 \theta_{13} } ~. \eeq \subsubsection{ Interconnections among ${\hat {\bar W}}_{0}[\ell]$, ${\hat {\bar W}}_{sin}[\ell]$, and ${\hat {\bar W}}_{cos}[\ell]$ } Significant simplifications arise from interconnections among $W_0[\ell]$, $W_{cos}[\ell]$, and $W_{sin}[\ell]$. One of these follows from Eqs.(,), which show that the matrix elements of ${\hat {\bar W}}_{cos}[\ell]$ and ${\hat {\bar W}}_{sin}[\ell]$ are proportional, namely \beq &&<M(a)| {\hat {\bar W}}_{sin}[\ell] |M(b)> \nonumber \\ &=& f_{ab} <M(a)| {\hat {\bar W}}_{cos}[\ell] |M(b)> ~, \eeq with the constants of proportionality $f_{ab}$ independent of $\ell$, \beq f_{ab} &=& \left( \begin{array}{ccc} 0 & -1 & -1 \\ 1 & 0 & -\cos^{-1}{2 \theta_{23}} \\ 1 & \cos^{-1}{2 \theta_{23}} & 0 \end{array} \right ) ~. \eeq Another follows from Eqs.~(,), from which it follows that \beq a^{(12)}_1 c^{(12)}_0 - a^{(12)}_0 c^{(12)}_1 &=& 2 K \nonumber \\ a^{(23)}_1 c^{(23)}_0 - a^{(23)}_0 c^{(23)}_1 &=& - K ~. \eeq The quantity $K$ is \beq K &=& -\frac{\alpha(1-\alpha) }{8} \nonumber \\ &\times& \cos{\theta_{13}} \sin{2\theta_{12}} \sin{2\theta_{13}} ~. \eeq \subsection{ Expressions for $ S(T)$ in terms of $\delta_{cp}$ } The dependence of $ S(T)$ on the CP violating phase $\delta_{cp}$ is very simple and follows from Eqs.~(,), noting that $ F^{ab}_{\ell} = {\hat {\bar W}}^{ab}[\ell] / {\hat {\bar D}}[\ell] $, Eq.~(). We thus find, \beq Re[ S^{ab}(t',t) ] &=& \delta_{ab} - 2 \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_0[\ell]} { {\hat {\bar D}}[\ell] }\sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& 2 \cos{\delta_{cp}} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{cos}[\ell] } { {\hat {\bar D}}[\ell] }\sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &+& \sin\delta_{cp} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{sin}[\ell] } { {\hat {\bar D}}[\ell] }\sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~, \eeq and \beq Im[ S^{ab}(t',t) ] &=& - 2 \sin\delta_{cp} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{sin}[\ell] } { {\hat {\bar D}}[\ell] } \sin^2{ {\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_0[\ell] } { {\hat {\bar D}}[\ell] } \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } \nonumber \\ &-& \cos{\delta_{cp}} \sum_{\ell} \frac{ {\hat {\bar W}}^{ab}_{cos}[\ell] } { {\hat {\bar D}}[\ell] } \sin{ 2{\hat{\bar E}}_{\ell} \Delta_L } ~. \eeq The dependence of $S(L)$ on ${\hat {\bar E}} _{\ell}$ and the remaining parameters of $H_\nu$ is given analytically through Eqs.~(,,). \section {Oscillation Probability $\mathcal{P}(\nu_a \rightarrow \nu_b)$ in our Hamiltonian formulation } The neutrino oscillation probabilities are obtained directly from our expression for the time-evolution operator, Eq.~(). In the high-energy limit, the oscillation probability in Eq.~(), expressed in terms of $Re[F^{ab}_{\ell\ell'}]$ and $Im[F^{ab}_{\ell\ell'}]$, is \beq &&\mathcal{P}(\nu_a \rightarrow \nu_b) = 2\sum_{\ell > \ell'} Im[F^{ab}_{\ell\ell'}] \sin{ (2 \Delta {\hat{\bar E}}_{\ell\ell'} \Delta_L )} \nonumber \\ &+& 2\sum_{\ell>\ell'} Re [F^{ab}_{\ell \ell'}] \cos{( 2 \Delta {\hat{\bar E}}_{\ell\ell'} \Delta_L ) } ~, \eeq where $1 \leq \ell \leq 3$, where $\Delta_L$ was defined in Eq.~(), and $\Delta {\hat{\bar E}}_{\ell\ell'} $ in Eq.~(). Note that $Re [F^{ab}_{\ell \ell'}]$ satisfies a normalization condition \beq \sum_{\ell\ell'} Re F^{ab}_{\ell\ell'} &=& \delta_{ab} ~, \eeq found by evaluating Eq.~() at $t = t'$ and recognizing that $\mathcal{P}(\nu_a \rightarrow \nu_b)$ is a real number. Taking Eq.~() into account and using the identity $1 - \cos{ 2 \beta } \equiv 2\sin^2{ \beta}$, we find an equivalent expression \beq &&\mathcal{P}(\nu_a \rightarrow \nu_b) = \delta_{ab} + 2\sum_{\ell >\ell'} Im[ F^{ab}_{\ell\ell'}] \nonumber \\ &\times& \sin{ ( 2 \Delta {\hat{\bar E}}_{\ell\ell'} \Delta_L )} \nonumber \\ &-& 4\sum_{\ell >\ell'} Re F^{ab}_{\ell \ell'} \sin^2{ (\Delta {\hat{\bar E}}_{\ell\ell'} \Delta_L ) } ~, \eeq that bears a striking similarity to the familiar vacuum expression with $F^{ab}_{\ell\ell'}$ playing a role analogous to $J^{ab}_{\ell\ell'}$ (as, for example, in Eq.~(1) of Ref.~). Equations~(,) make use of the fact that a probability is purely real. \subsection {Properties of $F^{ab}_{\ell\ell'}$ } It follows from Eq.~() and the observation that the energy of a neutrino or antineutrino in matter is independent of $\delta_{cp}$ that $F^{ab}_{\ell\ell'}$ is symmetric under the simultaneous exchange of $a,b$ and $\ell,\ell'$, \beq F^{ab}_{\ell\ell'} &=& F^{ba}_{\ell' \ell} ~. \eeq Using in addition the Hermiticity of ${\hat {\bar W}}[\ell]$, Eq.~() we find the reflection symmetry \beq F^{ba}_{\ell\ell'} &=& F^{ab*}_{\ell\ell'} \nonumber \\ F^{ab}_{\ell'\ell} &=& F^{ab*}_{\ell\ell} ~. \eeq From this, along with Eq.~(), it follows that $Im F^{ab}_{\ell\ell'}$ is antisymmetric under the exchange of either $(a,b)$ or $(\ell,\ell')$, \beq Im F^{ab}_{\ell \ell'} &=& -Im F^{ab}_{\ell' \ell} = -Im F^{ba}_{\ell\ell' } ~, \eeq whereas $ReF^{ab}_{\ell\ell'}$ is symmetric, \beq Re F^{ab}_{\ell \ell'} &=& Re F^{ab}_{\ell'\ell} = Re F^{ba}_{\ell\ell' } ~. \eeq There are no symmetries connecting the $F^{ab}_{\ell\ell'}$ of neutrinos to those of antineutrinos because these energies are, in general, different. This is not the case, however, in the vacuum for theories invariant under CPT. \subsection{ General Expressions for $ F^{ab}_{\ell\ell'}$ } The quantity $ F^{ab}_{\ell \ell'} $ is most easily obtained from $ w^{ab}_{\ell \ell'} $, \beq w^{ab}_{\ell \ell'} &\equiv& <M(b)| {\hat {\bar W}}[\ell] |M(a)> \nonumber \\ &\times& <M(b)|{\hat {\bar W}}[\ell'] |M(a)>^* ~. \eeq Analytic formulae for $ w^{ab}_{\ell,\ell'} $ are easily obtained in terms of the parameters of $H_\nu$ using ${\hat {\bar W}} [\ell] $ given in Eq.~(). Equation~() then gives $ F^{ab}_{\ell \ell'} $ as \beq F^{ab}_{\ell \ell'} &=& \frac{ w^{ab}_{\ell \ell'} } {{\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } ~. \eeq Because ${\hat {\bar W}} [\ell] $ consists of three terms, one proportional to $\sin\delta_{cp}$, one proportional to $\cos\delta_{cp}$, and one independent of $\delta_{cp}$, the dependence of $ w^{ab}[\ell,\ell'] $ on $\delta_{cp}$ can be expressed {\it a priori} through the five operators, \beq w^{ab}_{\ell\ell'} &=& w_{0\ell\ell'} ^{ab} + \cos{\delta_{cp}} w_{cos\ell\ell'}^{ab} \nonumber \\ &+& \cos^2{\delta_{cp}} w_{cos^2\ell\ell'}^{ab} + i( \sin{\delta_{cp}} w_{sin\ell\ell'}^{ab} \nonumber \\ &+& \sin{\delta_{cp}}\cos{\delta_{cp}}w^{ab}_{sin\times cos\ell\ell'} ~, \eeq each of which is uniquely determined by Eqs.~(,). Clearly, just as for the ${\hat {\bar W}}_i[\ell]$, the matrix elements of $w_i[\ell,\ell']$ are real and independent of $\delta_{cp}$. Using the hermiticity of ${\hat {\bar W}}_{0}[\ell] $ and ${\hat {\bar W}}_{cos}[\ell] $, and the anti-hermiticity of ${\hat {\bar W}}_{sin}[\ell] $, we then find \beq w^{ab}_{0\ell\ell'} &\equiv& {\hat {\bar W}}^{ab}_0[\ell] {\hat {\bar W}}^{ab}_0[\ell'] + {\hat {\bar W}}^{ab}_{sin}[\ell] {\hat {\bar W}}^{ab}_{sin}[\ell'] \nonumber \\ w_{sin\ell\ell'}^{ab} &\equiv& {\hat {\bar W}}^{ab}_{sin}[\ell] {\hat {\bar W}}^{ab}_0[\ell'] - {\hat {\bar W}}^{ab}_0[\ell] {\hat {\bar W}}^{ab}_{sin}[\ell'] \nonumber \\ w^{ab}_{cos\ell \ell'} &\equiv& {\hat {\bar W}}^{ab}_{cos}[\ell] {\hat {\bar W}}^{ab}_0[\ell'] + {\hat {\bar W}}^{ab}_0[\ell] {\hat {\bar W}}^{ab}_{cos}[\ell'] \nonumber \\ w_{cos^2\ell\ell'}^{ab} &\equiv& {\hat {\bar W}}^{ab}_{cos}[\ell] {\hat {\bar W}}^{ab}_{cos}[\ell'] - {\hat {\bar W}}^{ab}_{sin}[\ell] {\hat {\bar W}}^{ab}_{sin}[\ell'] \nonumber \\ w^{ab}_{sin\times cos\ell\ell'} &\equiv& {\hat {\bar W}}^{ab}_{sin}[\ell] {\hat {\bar W}}^{ab}_{cos}[\ell'] \nonumber \\ &~&~~~~~~~~~ - {\hat {\bar W}}^{ab}_{cos}[\ell] {\hat {\bar W}}^{ab}_{sin}[\ell']~. \eeq Note that Eqs.~(,) require that $w_{sin\times cos}$ vanish identically, \beq w^{ab}_{sin\times cos} &=& 0 ~, \eeq so this term need not be considered further. Note also that the dependence of $ w^{ab}_{i\ell\ell'} $ on $\ell$ and $\ell'$ arises entirely from the eigenvalues ${\hat {\bar E}} _{\ell}$ and ${\hat {\bar E}} _{\ell'}$, as ${\hat {\bar W}}^{ab}[\ell]$ depends on $\ell$ entirely through ${\hat {\bar E}} _{\ell}$. Finally, we will find it useful to define $ w_{i}^{ab} [\ell] $ in analogy to ${\hat {\bar W}}[\ell] $ in Eq.~(), \beq w_{i}^{ab} [1] &\equiv& w_{i}^{ab} [3,2] \nonumber \\ w_{i}^{ab} [2] &\equiv& w_{i}^{ab} [3,1] \nonumber \\ w_{i}^{ab} [3] &\equiv& w_{i}^{ab} [2,1] ~. \eeq These are the only three $ w_{i}^{ab} [\ell,\ell']$ needed because of the restriction $\ell >\ell'$ in Eq.~(). \subsection{ Analytic expressions for $ w_{sin}^{ab} [\ell,\ell'] $ } From the symmetries of ${\hat {\bar W}} ^{ab}_{0}[\ell]$, ${\hat {\bar W}}^{ab}_{cos}[\ell]$, and ${\hat {\bar W}}^{ab}_{sin}[\ell]$, we see from Eq.~() that $w^{ab}_{sin\ell\ell'}$ is odd under the exchange of either $a,b$ or $\ell,\ell'$. The term $w^{ab}_{sin\ell,\ell'}$ therefore vanishes for $a=b$ [and for $\ell = \ell'$], but again recall that restrictions on the sums is such that $w^{ab}_{\ell\ell'}$ contributes only for $\ell > \ell'$. Using the general results in Eqs.~(,,), we find $ w_{sin\ell\ell'}^{ab} $ given in Eq.~() is, as a matrix, \beq w_{sin\ell\ell'} &=& \frac{1}{2} \sin{2\theta_{23}} \left( \begin{array}{ccc} 0 & A & -A \\ -A & 0 & B \\ A & -B & 0 \end{array} \right ) ~, \eeq where \beq A &=& -( a^{(12)}[\ell] c^{(12)}[\ell'] - c^{(12)}[\ell] a^{(12)}[\ell']) \nonumber \\ &=& -(a^{(12)}_1 c^{(12)}_0 - a^{(12)}_0 c^{(12)}_1) \Delta {\hat{\bar E}}_{\ell\ell'} \eeq and \beq B &=& 2 (a^{(23)}[\ell] c^{(23)}[\ell'] - c^{(23)}[\ell] a^{(23)}[\ell']) \nonumber \\ &=& 2(a^{(23)}_1 c^{(23)}_0 - a^{(23)}_0 c^{(23)}_1) \Delta {\hat{\bar E}}_{\ell\ell'} ~. \eeq We see from Eq.~() that that $A=B$, so from Eq.~() we find the following simple expression, \beq w^{ab} _{sin\ell\ell'} &=& K \sin{2\theta_{23}} \Delta {\hat{\bar E}}_{\ell\ell'} \epsilon^{ab}_{\sin} ~, \eeq and correspondingly \beq w^{ab} _{sin}[\ell] &=& K \sin{2\theta_{23}} \Delta {\hat{\bar E}}[\ell] \epsilon^{ab}_{\sin} ~, \eeq where $\epsilon_{\sin}$ is the anti-symmetric matrix \beq \epsilon_{sin} &\equiv& \left( \begin{array}{ccc} 0 & 1 & -1 \\ -1 & 0 & 1 \\ 1 & -1 & 0 \end{array} \right) ~. \eeq \subsection{ Analytic expressions for $ w^{ab}_{cos^2} $, $ w^{ab}_{cos} $, and $ w^{ab}_0 $ } We also see from the symmetries of ${\hat {\bar W}} ^{ab}_{0}[\ell]$, ${\hat {\bar W}}^{ab}_{cos}[\ell]$, and ${\hat {\bar W}}^{ab}_{sin}[\ell]$, that $w^{ab}_{0\ell\ell'}$, $w^{ab}_{cos\ell\ell'}$, and $w^{ab}_{\cos^2\ell\ell'}$ are even under the exchange of $a,b$ and/or $\ell,\ell'$. Since the dependence of all $ w^{ab}_{i\ell\ell'} $ on $\ell$ and $\ell'$ arises entirely from the eigenvalues ${\hat {\bar E}} _{\ell}$ and ${\hat {\bar E}} _{\ell'}$, it follows that the terms symmetric in $\ell$ and $\ell'$ must be functions of symmetric combinations of ${\hat {\bar E}} _{\ell}$ and ${\hat {\bar E}} _{\ell'}$. There are only three irreducible symmetric functions of ${\hat {\bar E}} _{\ell}$ and ${\hat {\bar E}} _{\ell'}$: a constant, the sum ${\hat {\bar E}} _{\ell}+{\hat {\bar E}} _{\ell'}$, and the product ${\hat {\bar E}} _{\ell}{\hat {\bar E}} _{\ell'}$, just as for ${\hat {\bar W}}[\ell] $ in Eq.~(). Thus, we anticipate that the entire dependence of $w^{ab}_0[\ell]$, $w^{ab}_{cos}[\ell]$, and $w^{ab}_{\cos^2}[\ell]$ on $\ell$ will occur through $\Sigma[\ell]$ and $\Pi[\ell]$ in Eqs.~(,), and hence through ${\hat {\bar E}} _{\ell}$, as for ${\hat {\bar W}}[\ell] $. In the following discussion, it is important to recall our convention that a quantity, such as $w^{(ab)}[\ell]$, written {\it with} parentheses enclosing $ab$ is a number, whereas $w^{ab}[\ell]$ written {\it without} parentheses surrounding $ab$ is the element of a matrix $w[\ell]$, \beq w^{ab}[\ell] &\equiv& <M(b)| w[\ell] |M(a)> ~. \eeq \subsubsection{ Matrix Elements of $ w_{cos^2} $ } Using Eq.~() with Eqs.~(,), the matrix for $w_{cos^2} $ is found to be \beq w_{cos^2} &=& \sin^2{2\theta_{23}} \nonumber \\ &\times& \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & w^{(22)}_{cos^2} & - w^{(22)}_{cos^2} \\ 0 & - w^{(22)}_{cos^2} & w^{(22)}_{cos^2} \end{array} \right ) ~, \eeq where \beq w^{(22)}_{cos^2\ell\ell'} &=& a^{(23)}[\ell] a^{(23)}[\ell'] \nonumber \\ &=& a^{(23)2}_0 + a^{(23)}_0 a^{(23)}_1 ({\hat {\bar E}} _{\ell} + {\hat {\bar E}} _{\ell'}) \nonumber \\ &+& a^{(23)2}_1 {\hat {\bar E}} _{\ell}{\hat {\bar E}} _{\ell'} ~. \eeq Correspondingly, \beq w^{(22)}_{cos^2} [\ell] &=& a^{(23)2}_0 + a^{(23)}_0 a^{(23)}_1~ \Sigma[\ell] \nonumber \\ &+& a^{(23)2}_1~ \Pi[\ell] ~. \eeq Note the somewhat subtle notational distinction between the ${\it operator}$ $ w_{cos^2} $ and the ${\it coefficients }$ $ w^{(ij)}_{cos^2} $ in terms of which it is defined. The latter is indicated by parentheses that surrounding the superscripts. By contrast, superscripts without parentheses, as in $ w^{ab}_i $, indicate the transition $\nu_a \rightarrow \nu_b$, ${i.e.}$, $w^{ab}_i \equiv <M(b) | w_i | M(a)>$. \subsubsection{ Matrix Elements of $ w_{cos} $ } We find only three independent, non-vanishing matrix elements for $w_{cos}$, one diagonal element and two off-diagonal elements. As a matrix, \beq && w_{cos} = \sin 2\theta_{23} \nonumber \\ &\times& \left( \begin{array}{ccc} 0 & w^{(12)}_{cos} & - w^{(12)}_{cos} \\ w^{(12)}_{cos} & w^{(22)}_{cos} & w^{(23)}_{cos} \\ - w^{(12)}_{cos} & w^{(23)}_{cos} & w^{(33)}_{cos} \end{array} \right )~. \eeq The two independent off-diagonal elements, \beq w^{(12)}_{cos\ell\ell'} &=& \frac{1}{2} (a^{(12)}[\ell] c^{(12)} [\ell'] + c^{(12)} [\ell] a^{(12)}[\ell'] ) \nonumber \\ w^{(23)}_{cos\ell\ell'} &=& -( a^{(23)}[\ell] c^{(23)} [\ell'] + c^{(23)} [\ell] a^{(23)}[\ell'] ) \nonumber \\ &\times& \cos{2\theta_{23}} ~, \eeq follow immediately from the structure of ${\hat {\bar W}}^{ab}_0[\ell'] $ and ${\hat {\bar W}}^{ab}_{cos}[\ell]$ given in Eqs.~(,), respectively. Taking $ w^{(22)}_{cos} $ from Eqs.~(,,), we obtain \beq w^{(22)}_{cos\ell\ell'} &=& - a^{(23)}[\ell] ({\hat {\bar E}} _{\ell'}^2 + c^{(22)} [\ell'] ) \nonumber \\ &-& ({\hat {\bar E}} _{\ell}^2 + c^{(22)} [\ell] ) a^{(23)}[\ell'] ~. \eeq Similarly, for the diagonal matrix element $D$, \beq w^{(33)}_{cos\ell\ell'} &=& a^{(23)}[\ell] ({\hat {\bar E}} _{\ell'}^2 + c^{(33)} [\ell'] ) \nonumber \\ &+& ({\hat {\bar E}} _{\ell}^2 + c^{(33)} [\ell] ) a^{(23)}[\ell'] ~. \eeq The element $ w^{(11)}_{cos\ell\ell'}$ vanishes as a consequence of ${\hat {\bar W}}^{(11)}_{cos}[\ell]=0$. Correspondingly, for the off-diagonal matrix elements we find, \beq w^{(12)}_{cos} [\ell] &=& a^{(12)}_0 c^{(12)}_0 + \frac{1}{2} ( a^{(12)}_1 c^{(12)}_0 + a^{(12)}_0 c^{(12)}_1 ) \nonumber \\ &\times& \Sigma[\ell] + a^{(12)}_1 c^{(12)}_1 \Pi[\ell] \nonumber \\ w^{(23)}_{cos} [\ell] &=& -2 \cos{2\theta_{23}} a^{(23)}_0 c^{(23)}_0 \nonumber \\ &-& \cos{2\theta_{23}} ( a^{(23)}_1 c^{(23)}_0 + a^{(23)}_0 c^{(23)}_1 ) \nonumber \\ &\times& \Sigma[\ell] - 2 \cos{2\theta_{23}} a^{(23)}_1 c^{(23)}_1 ~\Pi[\ell] ~, \eeq which simplify somewhat by using Eqs.~(,). For the diagonal matrix elements, \beq w^{(22)}_{cos} [\ell] &=& -2 a^{(23)}_0 c^{(22)}_0 - ( a^{(23)}_1 c^{(22)}_0 + a^{(23)}_0 c^{(22)}_1) \nonumber \\ &\times& \Sigma[\ell] + 2( a^{(23)}_0 - a^{(23)}_1 c^{(22)}_1)~ \Pi[\ell] \nonumber \\ &-& a^{(23)}_1 \Sigma[\ell] ~\Pi[\ell] - a^{(23)}_0 \Sigma[\ell]^2 \eeq and \beq w^{(33)}_{cos} [\ell] &=& 2 a^{(23)}_0 c^{(33)}_0 + ( a^{(23)}_1 c^{(33)}_0 + a^{(23)}_0 c^{(33)}_1) \nonumber \\ &\times& \Sigma[\ell] - 2( a^{(23)}_0 - a^{(23)}_1 c^{(33)}_1)~ \Pi[\ell] \nonumber \\ &+& a^{(23)}_1 \Sigma[\ell] ~\Pi[\ell] + a^{(23)}_0 \Sigma[\ell]^2 ~. \eeq The matrix elements $ w^{(22)}_{cos}[\ell] $ and $ w^{(33)}_{cos}[\ell] $ are not independent. The reason is that $ w_{cos}^{(33)} [\ell]$ may be found from $ w_{cos}^{(22)} [\ell]$ by making the replacement $\sin{\theta_{23}} \leftrightarrow \cos{\theta_{23}} $ [Eq.~()] and by flipping the overall sign. Recall that under this replacement, $ c^{(22)} \leftrightarrow c^{(33)} $ and that $ a^{(23)} $ is independent of $\sin{\theta_{23} }$. \subsubsection{ Matrix Elements of $ w_{0} $ } We find four independent, non-vanishing matrix elements for $w_{0}$, two diagonal and two off-diagonal elements. As a matrix, \beq && w_{0} = \left( \begin{array}{ccc} w_{0}^{(11)} & w_{0}^{(12)} & w_{0}^{(13)} \\ w_{0}^{(12)} & w_{0}^{(22)} & w_{0}^{(23)} \\ w_{0}^{(13)} & w_{0}^{(23)} & w_{0}^{(33)} \end{array} \right )~. \eeq The off-diagonal elements, \beq w_{0\ell\ell'}^{(12)} &=& c^{(12)} [\ell] c^{(12)} [\ell'] \cos^2\theta_{23} \nonumber \\ &+& a^{(12)}[\ell] a^{(12)}[\ell'] \sin^2{\theta_{23} } \nonumber \\ w_{0\ell\ell'}^{(13)} &=& c^{(12)} [\ell] c^{(12)} [\ell'] \sin^2\theta_{23} \nonumber \\ &+& a^{(12)}[\ell] a^{(12)}[\ell'] \cos^2{\theta_{23} } \nonumber \\ w_{0\ell\ell'}^{(23)} &=& c^{(23)}[\ell] c^{(23)}[\ell'] \sin^2 2 \theta_{23} \nonumber \\ &+& a^{(23)}[\ell] a^{(23)}[\ell'] ~, \eeq follow immediately from the structure of ${\hat {\bar W}}^{ab}_0[\ell'] $ and ${\hat {\bar W}}^{ab}_{sin}[\ell]$ given in Eqs.~(,), respectively. The diagonal elements $ w^{(nn)}_{0} $ are found from Eqs.~(,), \beq w_{0\ell\ell'}^{(nn)} &=& ( {\hat {\bar E}} _{\ell}^2 + c^{(nn)} [\ell] ) \nonumber \\ &\times& ( {\hat {\bar E}} _{\ell'}^2 + c^{(nn)} [\ell'] ) ~. \eeq Correspondingly, for the off-diagonal matrix elements, \beq w_{0}^{(12)} [\ell] &=& c^{(12)2}_0 \cos^2{\theta_{23}} + a^{(12)2}_0 \sin^2{\theta_{23}} \nonumber \\ &+& (c^{(12)}_0 c^{(12)}_1 \cos^2{\theta_{23}} + a^{(12)}_0 a^{(12)}_1 \sin^2{\theta_{23}} ) \Sigma[\ell] \nonumber \\ &+& ( c^{(12)2}_1 \cos^2{\theta_{23}} + a^{(12)2}_1 \sin^2{\theta_{23}})~ \Pi[\ell] \nonumber \\ w_{0}^{(13)} [\ell] &=& c^{(12)2}_0 \sin^2{\theta_{23}} + a^{(12)2}_0 \cos^2{\theta_{23}} \nonumber \\ &+& (c^{(12)}_0 c^{(12)}_1 \sin^2{\theta_{23}} + a^{(12)}_0 a^{(12)}_1 \cos^2{\theta_{23}} ) \Sigma[\ell] \nonumber \\ &+& ( c^{(12)2}_1 \sin^2{\theta_{23}} + a^{(12)2}_1 \cos^2{\theta_{23}})~ \Pi[\ell] \nonumber \\ w_{0}^{(23)} [\ell] &=& c^{(23)2}_0 \sin ^2{2 \theta_{23}} + a^{(23)2}_0 \nonumber \\ &+& ( a^{(23)}_0 a^{(23)}_1 + c^{(23)}_0 c^{(23)}_1 \sin ^2{2 \theta_{23}} ) ~\Sigma[\ell] \nonumber \\ &+& ( c^{(23)2}_1 \sin^2{\theta_{23}} + a^{(23)2}_1 )~ \Pi[\ell] ~. \eeq For the diagonal matrix elements, \beq w_{0}^{(nn)} [\ell] &=& c^{(nn)2}_0 - (2 c^{(nn)}_0 - c^{(nn)2}_1 )~\Pi[\ell] \nonumber \\ &+& c^{(nn)}_0 c^{(nn)}_1 ~ \Sigma[\ell] + c^{(nn)}_1 \Sigma[\ell] ~ \Pi[\ell] \nonumber \\ &+& \Pi[\ell]^2 ~. \eeq As for $ w_{cos}$, the matrix elements $ w^{(22)}_{0}[\ell]$ and $ w^{(33)}_{0}[\ell]$ are not independent. The same is true for $ w_{0}^{(13)} [\ell]$ and $ w_{0}^{(12)} [\ell]$. In both cases the matrix elements may be found from one another by making the replacement $\sin{\theta_{23}} \leftrightarrow \cos{\theta_{23}} $ [Eq.~()]. Recall that under such a replacement $ c^{(22)} \leftrightarrow c^{(33)} $ and that $ c^{(12)} $ and $ a^{(23)} $ are independent of $\sin{\theta_{23} }$. Writing the dependence of the coefficients $ w_i^{(mn)} [\ell] $ on ${\hat {\bar E}} _{\ell}$ explicitly, \beq w_i^{(mn)} [\ell] &=& w_{i;0}^{(mn)} + w_{i;1}^{(mn)} ~{\hat {\bar E}} _{\ell} \nonumber \\ &+& w_{i;2}^{(mn)} ~{\hat {\bar E}} _{\ell}^2 ~. \eeq \subsection{ Analytic expression for $ Im[F^{ab}_{\ell\ell'}] $ and $ Re[ F^{ab}_{\ell\ell'}] $ } A general form for the real and imaginary parts of $F^{ab}_{\ell\ell'}$ in terms of the four operators, ($w_0$, $w_{sin}$, $w_{cos}$, $w_{cos^2}$) is found from Eq.~() and Eq.~(). For $Im[F^{ab}_{\ell\ell'}]$, using Eq.~(), we find \beq Im[F^{ab}_{\ell \ell'}] &=& \sin{\delta_{cp}} \frac{ w_{sin\ell\ell'}^{ab} }{ {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } \nonumber \\ &=& \frac{1}{2} \sin{\delta_{cp} } \sin{2\theta_{23}} K \epsilon^{ab}_{\sin} \nonumber \\ &\times& \frac{ \Delta {\hat{\bar E}}_{\ell\ell'} }{ {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } ~. \eeq Then, using the easily-verified result \beq &&\frac{ \Delta {\hat{\bar E}}_{\ell\ell'} }{ {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } = \frac{ \epsilon^{\ell\ell'}_{\sin} } { {\hat {\bar D}} } ~, \eeq we find \beq Im F^{ab}_{\ell\ell'} &=& -\sin{\delta_{cp}} \epsilon_{sin}^{ab} \epsilon_{sin}^{\ell\ell' } \frac{\alpha(1-\alpha) }{ 8{\hat {\bar D}} } \cos{\theta_{13}} \nonumber \\ &\times& \sin{2\theta_{12}} \sin{2\theta_{13}} \sin{2\theta_{23}} ~. \eeq For $Re[F^{ab}_{\ell\ell'}]$ we find \beq Re[F^{ab}_{\ell \ell'}] &=& \frac{ w_{0\ell\ell'}^{ab} }{ {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } + \cos{\delta_{cp}} \frac{ w_{cos\ell\ell'}^{ab} }{ {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } \nonumber \\ &+& \cos^2{\delta_{cp}} \frac{ w_{cos^2\ell\ell'}^{ab} }{ {\hat {\bar D}}[\ell] {\hat {\bar D}}[\ell'] } ~. \eeq Equations~(,) show that $F^{ab}_{\ell\ell'}$ plays a role for $\mathcal{P}(\nu_a \rightarrow \nu_b)$ similar to the one that $F^{ab}_{\ell}$ plays for $S(T)$. Using Eq.~(), it follows from Eqs.~(,) that $Im F^{ab}_{\ell\ell'}$ is odd, and $Re F^{ab}_{\ell\ell'}$ even, under $\delta_{cp}\rightarrow -\delta_{cp}$. \newpage \begin{thebibliography}{99} \bibitem{wolf} L. Wolfenstein, , Phys. Rev. {\bf D17}, 2369 (1978). \bibitem{ISS} The ISS Working group, arXiv:0710.4947/hep-ph. \bibitem{barg} V. Barger, K. Whisnant, S. Pakvasa, and R. J. N. Phillips, Phys. Rev. {\bf D22}, 2718 (1980). \bibitem{f} M. Freund, Phys. Rev. {\bf{D64}}, 053003 (2001). \bibitem{ahlo} E. Akhmedov, P. Huber, M. Lindner, and T. Ohlsson, Nucl. Phys. {\bf{B608}}, 394 (2001). \bibitem{JO} M. Jacobson and T. Ohlsson, Phys. Rev. D {\bf 69}, 013003 (2004). \bibitem{hjk1} E. M. Henley, M. B. Johnson, and L. S. Kisslinger, Int J. Mod Phys. E {\bf 20}, 2463 (2011); arXiv:1102.5106/hep-ph. \bibitem{khj2} L. S. Kisslinger, E. M. Henley, and M. B. Johnson, arXiv:1105.2741/hep-ph (2011). \bibitem{hjk1a} L. S. Kisslinger, E. M. Henley, and M. B. Johnson, Int J. Mod Phys. E {\bf 21}, 12920 (2012) (erratum); arXiv:1205.6430/hep-ph. \bibitem{khj1E} L. S. Kisslinger, E. M. Henley, and M. B. Johnson, Int J. Mod Phys. E {\bf 21}, 1250065 (2012); arXiv:1203.6613/hep-ph. \bibitem{kiss1} L. S. Kisslinger, Mod. Phys. Lett. A, {\bf 30}, 1550014 (2015); arXiv:1408.0310/hep-ph. \bibitem{KS} C. W. Kim and W. K. Sze, Phys. Rev D {\bf 35}, 1404 (1987). \bibitem{Dav} H. Davoudiasl, H. S. Lee, and W. J. Marciano, Phys. Rev.~{\bf D84}, 013009, (2011). \bibitem{Gon} M. C. Gonzalez-Garcia, M. Maltoni, and J. Salvado, JHEP 1105: 075 (2011); arXiv:1103.4365/hep-ph. \bibitem{DB} F.P. An et.al., Daya Bay Collaboration, Chin. Phys. C37, 01101 (2013); arXiv:1210.6327/hep-ex. \bibitem{Cer} A. Cervera, {\it et al}, Nucl. Phys. {\bf B579}, 17 (2000). \bibitem{jhk1} M. B. Johnson, E. M. Henley, and L. S. Kisslinger, manuscript in preparation. \bibitem{jhk2} M. B. Johnson, E. M. Henley, and L. S. Kisslinger, manuscript in preparation. \bibitem{OS} T. Ohlsson and H. Snellman, Phys. Lett. {\bf B 474}, 153 (2000). \end{thebibliography} |
1501.04099 | Title: Topological phase transitions in superradiance lattices
Abstract: Topological phases of matters are of fundamental interest and have promising
applications. Fascinating topological properties of light have been unveiled in
classical optical materials. However, the manifestation of topological physics
in quantum optics has not been discovered. Here we study the topological phases
in a two-dimensional momentum-space superradiance lattice composed of timed
Dicke states (TDS) in electromagnetically induced transparency (EIT). By
periodically modulating the three EIT coupling fields, we can create a Haldane
model with in-situ tunable topological properties, which manifest themselves in
the contrast between diffraction signals emitted by superradiant TDS. The
topological superradiance lattices provide a controllable platform for
simulating exotic phenomena in condensed matter physics and offer a basis of
topological quantum optics and novel photonic devices.
Body: \title{Topological phase transitions in superradiance lattices} \author{Da-Wei Wang} \email{cuhkwdw@gmail.com} \affiliation{Department of Physics and Centre for Quantum Coherence, The Chinese University of Hong Kong, Hong Kong, China} \affiliation{Institute for Quantum Science and Engineering and Department of Physics and Astronomy, Texas A$\&$M University, College Station, TX 77843, USA} \author{Han Cai} \affiliation{Institute for Quantum Science and Engineering and Department of Physics and Astronomy, Texas A$\&$M University, College Station, TX 77843, USA} \author{Luqi Yuan} \affiliation{Institute for Quantum Science and Engineering and Department of Physics and Astronomy, Texas A$\&$M University, College Station, TX 77843, USA} \affiliation{Department of Electrical Engineering, and Ginzton Laboratory, Stanford University, Stanford, CA 94305, USA} \author{Shi-Yao Zhu} \affiliation{Beijing Computational Science Research Centre, Beijing 100084, China} \affiliation{Synergetic Innovation Center of Quantum Information and Quantum Physics, University of Science and Technology of China, Hefei, Anhui 230026, China} \author{Ren-Bao Liu} \email{rbliu@phy.cuhk.edu.hk} \affiliation{Department of Physics and Centre for Quantum Coherence, The Chinese University of Hong Kong, Hong Kong, China} \date{\today } \begin{abstract} Topological phases of matters are of fundamental interest and have promising applications. Fascinating topological properties of light have been unveiled in classical optical materials. However, the manifestation of topological physics in quantum optics has not been discovered. Here we study the topological phases in a two-dimensional momentum-space superradiance lattice composed of timed Dicke states (TDS) in electromagnetically induced transparency (EIT). By periodically modulating the three EIT coupling fields, we can create a Haldane model with in-situ tunable topological properties, which manifest themselves in the contrast between diffraction signals emitted by superradiant TDS. The topological superradiance lattices provide a controllable platform for simulating exotic phenomena in condensed matter physics and offer a basis of topological quantum optics and novel photonic devices. \end{abstract} \maketitle \noindent The quantum Hall effect (QHE) reveals a new class of matter phases, topological insulators (TIs), which have been extensively studied in solid-state materials and recently in photonic structures , time-periodic systems and optical lattices of cold atoms . The first TI is the Haldane model proposed in 1988, which shows that the QHE is an intrinsic topological property of the energy bands. It inspired the discovery of the quantum spin Hall effect and topological superconductors . The Haldane model consists of a honeycomb tight-binding lattice with complex next-nearest-neighbour (NNN) hopping, which breaks the time-reversal symmetry and induces an energy gap between two bands that have opposite Chern numbers. Inversion symmetry breaking by an on-site potential can also open band gaps. The interplay of these two types of symmetry breaking leads to transitions between phases with Chern numbers 0 and $\pm 1$. Notwithstanding its foundational role in topological condensed matter physics, the Haldane model has never been realized in solid-state systems. Floquet modulation of a circular polarised light in graphene was proposed to induce the complex NNN tunnelling . However, the required light is soft X-ray which would directly excite electrons and be absorbed. The Haldane model of cold atoms in optical lattices were recently realised in experiments. Our recent study shows that Scully's timed Dicke states (TDS) can form a superradiance lattice (SL) in momentum space . Here we propose a quantum optics realisation of an in-situ tunable Haldane model using two-dimensional SLs in a simple electromagnetically induced transparency (EIT) configuration . Since Dicke's seminal paper in 1954 , superradiance has been an important topic in quantum optics. A single photon with wave vector $\mathbf{k}_p$ can excite a spatially extended $N$-atom ensemble from the ground state $|G\rangle=|g_1,g_2,\ldots, g_N\rangle$ to the TDS \begin{equation} |e_{\mathbf{k}_p}\ra=\frac{1}{\sqrt{N}}\sum\limits_{j=1}^{N}e^{i\mathbf{k}_p\cdot \mathbf{r}_j}|g_1, g_2,...,e_j,...,g_N\ra, \end{equation} where $e_i$ and $g_i$ are the excited and ground states of the $i$th atom at position $\mathbf{r}_i$, respectively. The TDS stores a light momentum $\hbar\mathbf{k}_p$ via phase correlations between atoms excited at different positions. This momentum can be transferred back to a single photon via directional emission . By coupling $|e\ra$ to another state $|m\ra$ with three coherent plane wave fields, we construct a honeycomb SL of TDS in momentum space , as shown in Fig.. The SL Hamiltonian with rotating-wave approximation is (see Supplementary Information) \begin{equation} \begin{aligned} H=&\frac{\hbar\Delta_c}{2}\sum\limits_{\mathbf{k}}(|e_\mathbf{k}\ra\la e_\mathbf{k}|-|m_{\mathbf{k}-\mathbf{k}_1}\ra\la m_{\mathbf{k}-\mathbf{k}_1}|)\\ &-\sum\limits_{\mathbf{k}}\sum\limits_{l=1}^{3}\hbar\Omega_l [|e_{\mathbf{k}}\ra\la m_{\mathbf{k}-\mathbf{k}_l}|+h.c.], \end{aligned} \end{equation} where $\mathbf{k}=\mathbf{k}_p+r(\mathbf{k}_2-\mathbf{k}_1)+s(\mathbf{k}_3-\mathbf{k}_2)$ with integers $r$ and $s$, $\Omega_l$ is the Rabi frequency of the coupling field along $\mathbf{k}_l$, $|m_\mathbf{k}\ra$ are defined the same as in Eq.() but for $|m\ra$ states, and $\Delta_c=\omega_{em}-\nu_c$ is the detuning of the transition frequency $\omega_{em}$ between $|e\ra$ and $|m\ra$ from the angular frequency $\nu_c$ of the EIT coupling fields. The states $|e_{\mathbf{k}}\rangle$ and $|m_{\mathbf{k}-\mathbf{k}_1}\rangle$ form two sublattices of a honeycomb structure in momentum space [see Fig. (c)]. $\Omega_l$ causes the nearest-neighbour hopping. As we will show below, periodically modulating $\Omega_l$ can introduce NNN hopping with controllable phases. On the other hand, the detuning $\Delta_c$ causes an energy offset between the two sublattices, breaking the inversion symmetry. Thus three Haldane phases with Chern numbers $\pm 1$ and $0$ can be in-situ realised. We first show how to induce complex NNN hopping. We consider periodically modulated Rabi frequencies $\Omega_l$ $(l=1,2,3)$ of coupling fields, \begin{equation} \Omega_l=\Omega_s+2\Omega_d\cos(\nu_d t+\phi_l), \end{equation} where $\Omega_{s}$ and $\Omega_{d}$ are the static and dynamic components of the Rabi frequency, $\nu_d$ is the modulation frequency and $\phi_l$ is the modulation phase. Here we have assumed the size of the atomic ensemble be much smaller than $c/\nu_d$ (where $c$ is the speed of light) and hence neglected the position dependence of the modulation. On the other hand, the ensemble size is much larger than $c/\nu_c$ and the number of atoms $N\gg 1$ such that the TDS in the SL are approximately orthogonal to each other, i.e., $\la e_{\mathbf{k}^\prime}|e_{\mathbf{k}}\ra\approx \delta_{\mathbf{k}\mathbf{k}^\prime}$. The periodic modulation induces Floquet sidebands with energy separation $\hbar\nu_d$. While $\Omega_s$ is the intra-sideband nearest-neighbour hopping, $\Omega_d$ induces the inter-sideband hopping. We choose $\nu_d\gg \Omega_{s/d},\Delta_c$ so that the Floquet sidebands are well separated in energy. Thus by second-order perturbation, the effective Hamiltonian of the NNN intra-sideband transition mediated by $\Omega_d$ is (see Supplementary Information) \begin{equation} \begin{aligned} H^\prime=&\sum\limits_{\mathbf{k}}\sum\limits_{l\ne j=1}^{3}\hbar\Omega_{lj} (|e_{\mathbf{k}+\mathbf{k}_l-\mathbf{k}_j}\ra \la e_{\mathbf{k}}|\\ &+|m_{\mathbf{k}-\mathbf{k}_1}\ra\la m_{\mathbf{k}-\mathbf{k}_1+\mathbf{k}_l-\mathbf{k}_j}|), \end{aligned} \end{equation} where the NNN hopping coefficient $\Omega_{lj}=2i\Omega^\prime\sin(\phi_l-\phi_j)$ with $\Omega^\prime=\Omega^2_d/\nu_d$. The fact that the NNN hopping coefficient is purely imaginary is crucial to the topological phases. The loop transitions via NNN hopping accumulate nonzero phases, as shown in Fig. (c). $H^\prime$ opens a band gap $\epsilon_{tr}=4\sqrt{3}\hbar|\alpha|\Omega^\prime$ where $\alpha$ is a dimensionless quantity determined by the summation of $\Omega_{(l+1)l}$, \begin{equation} \alpha=-\sum\limits_{l=1}^3\sin(\phi_{l+1}-\phi_l)=4\prod\limits_{l=1}^3\sin\frac{\phi_{l+1}-\phi_l}{2}. \end{equation} The Chern numbers of the upper and lower bands $C$ and $C^\prime$ are $C=-C^\prime=\text{sign}(\alpha)$ (see Supplementary Information). In Fig. (a), we plot $\alpha$ with $\phi_1=0$ and $0\le\phi_{2,3}<2\pi$. The topological property of this SL Haldane model can be represented by the distribution of $\phi_l$ on a unit circle. There are two distinct topological configurations, counter-clockwise $\phi_1$, $\phi_2$ and $\phi_3$ for $C=-1$, and clockwise $\phi_1$, $\phi_2$ and $\phi_3$ for $C=1$, as shown in Fig. (b). The time reversal $t\rightarrow-t$ in Eq.() is equivalent to $\phi_l\rightarrow-\phi_l$, which leads to $C\rightarrow-C$. Unlike TIs , topological superradiance lattices (TSLs) have no outer edges in the semiclassical limit of the coupling fields (see Supplementary Information). Neither do TSLs have Fermi surfaces. Nonetheless, the TSL has its unique topological properties that are observable. The TDS have directional superradiance emission. Of all the TDS in the SL, only those $|e_\mathbf{k}\ra$ with $c|\mathbf{k}|-\omega_{eg}$ within the energy bands ($\omega_{eg}$ is the transition frequency between $|e\ra$ and $|g\ra$) can satisfy both energy and momentum conservation, and have directional emission in $\mathbf{k}$ . We call these states superradiant TDS and the other ones subradiant TDS. We can regard these superradiant TDS as an inner edge of a honeycomb lattice of subradiant TDS . The topological orders lead to different light emissions from different superradiant TDS. Alternatively, we can also tune the probe field frequency to test the topological band properties at certain energy, which is analogue to tuning the Fermi surface in a fermionic system. For the sake of simplicity, we set the wavevectors of the three EIT coupling fields to be $\mathbf{k}_1=-k_c\hat{x}$, $\mathbf{k}_2=k_c(\hat{x}-\sqrt{3}\hat{y})/2$ and $\mathbf{k}_3=k_c(\hat{x}+\sqrt{3}\hat{y})/2$, and the probe field wavevector $\mathbf{k}_p=-\mathbf{k}_1$. In this case we have only three superradiant TDS with wavevectors $\mathbf{k}_p$, $\mathbf{k}_+=-\mathbf{k}_3$ and $\mathbf{k}_-=-\mathbf{k}_2$, and the diffraction fields are along $\mathbf{k}_\pm$, as shown in Fig. (a). We set all fields on resonance, i.e., $\Delta_c=0$ and $\Delta_p\equiv\omega_{eg}-\nu_p=0$. The excitation $|e_{\mathbf{k}_p}\ra$ flows to $|e_{\mathbf{k}_\pm}\ra$ and emits photons along $\mathbf{k}_\pm$. We denote the steady state probability amplitudes of states $|e_{\mathbf{k}_{\pm}}\ra$ as $c_{\mathbf{k}_\pm}$ and define the superradiance contrast \begin{equation} \eta=\frac{|c_{\mathbf{k}_+}|^2-|c_{\mathbf{k}_-}|^2}{|c_{\mathbf{k}_+}|^2+|c_{\mathbf{k}_-}|^2}. \end{equation} For $C=1$, the excitation current flows along $|e_{\mathbf{k}_p}\ra \rightarrow|e_{\mathbf{k}_+}\ra \rightarrow |e_{\mathbf{k}_-}\ra$ , as shown in Fig. (c). Since each TDS $|e_\mathbf{k}\ra$ or $|m_\mathbf{k}\ra$ has decoherence rate $\gamma_e$ or $\gamma_m$, respectively, the excitation decays while flowing and it is more probable in state $|e_{\mathbf{k}_+}\ra$ than in state $|e_{\mathbf{k}_-}\ra$. We therefore have $\eta>0$. Similarly, for $C=-1$, $\eta<0$. Thus the sign change of the superradiance contrast signatures the topological phase transition, as seen in Fig. (a). The superradiance contrast in Fig. (a) is consistent with Fig. (b) except for the two diagonal corners where the topological currents are weak and the local effect inside a unit cell dominates (see Supplementary Information). The topological phase with $C=0$ for non-zero $\Omega^\prime$ can be reached by breaking the inversion symmetry. Substantially easier than in graphene electrons, a sublattice offset in SLs can be introduced by choosing $\Delta_c\ne 0$. We set $\phi_l=4\pi(l-1)/3$ and thus $C=1$ when $\Delta_c=0$. The energy gap in absence of sublattice offset is $\epsilon_{tr}=18\hbar\Omega^\prime$. Topological phase transitions occur at $\hbar|\Delta_c|=\epsilon_{tr}$ (see Supplementary Information). In Fig. (b), we plot $\eta$ as a function of $\hbar\Delta_c/\epsilon_{tr}$. For $\Omega^\prime=0.01$, the phase transition is smeared out by the relatively large $\gamma_e$. The phase transition gets more apparent for larger $\Omega^\prime$. Depending on $\Omega^\prime$, $\eta$ has peaks or kinks near the phase transition point because the two bands touch at the middle of the band gap, where the probe field probes, as shown in Fig. (c). The TSLs have unique features in transient light propagation under pulse probe. In Fig., we compare the pulse propagation in a trivial SL with zero $\Omega^\prime$ and in a TSL. For a weak probe pulse, the linear susceptibility is $\chi^{(1)}\propto c_{\mathbf{k}_p}$ and the linear absorption is $\text{Im}\chi^{(1)}$. The two multi-wave-mixing signals along $\mathbf{k}_\pm$ correspond to the nonlinear susceptibilities $\chi_\pm\propto c_{\mathbf{k}_\pm}$ and can be understood as a result of optical grating . We simulate the pulse propagation for the three modes along $\mathbf{k}_{p,\pm}$ using coupled wave equations (see Supplementary Information). For a trivial SL without modulation, the light propagating along $\mathbf{k}_\pm$ is symmetric, while for a TSL with $C=1$, the topological currents drive the probe pulse to $\mathbf{k}_+$, even if the NNN hopping is two orders of magnitude smaller than the nearest-neighbour hopping. TSLs can be readily realised in experiments for cold alkali atoms. Taking $^{85}$Rb D1 line for example, we can have $|g\ra=|5^{2}S_{1/2}, F=2\ra$, $|e\ra=|5^{2}P_{1/2}, F=2\ra$ and $|m\ra=|5^{2}S_{1/2}, F=3\ra$. $\gamma_e=2.9$MHz and $\gamma_m$ is controllable via inhomogeneous magnetic field . The Rabi frequency $\Omega_s=3\gamma_e=8.6$MHz (intensity 25mW/cm$^{2}$). The modulation frequency can be $\nu_d=10\gamma_e=28.8$MHz which is large enough to separate the Floquet bands. One can trap $10^6$ atoms in 1 mm$^3$ such that $N\gg 1$ and the size $L$ of the ensemble $c/\nu_c\ll L\ll c/\nu_d$. In the $\mu$K regime, the thermal random motions have negligible Doppler shifts ($\sim$kHz). Another possible type of physical systems are rare earth atoms doped in solids . One should first optically pump nearly all population to $|g\ra$, then turn on three optical fields coupling $|e\ra$ to $|m\ra$, send in a weak field probing the $|g\ra$ to $|e\ra$ transition, and detect the diffraction signals. Similar to the electronic TIs and the optical lattice simulations of TIs , the topological properties of TSLs are determined by Schr\"odinger equations, different from the photonic TIs , which are governed by Maxwell equations. The generation and detection of the topological properties of TSLs, however, can be easily controlled by light. The unique feature of the TSL is that its lattice sites have discrete momenta rather than positions. It has the advantage to be extended to dimensions higher than three where no real space lattices exist and offers a platform for high-dimensional topological physics . \section*{Funding Information} \noindent Science Foundation Grants No. PHY-1241032 (INSPIRE CREATIV) and PHY-1068554; Robert A. Welch Foundation (Grant No. A-1261); Herman F. Heep and Minnie Belle Heep Texas A\&M University Endowed Fund; Hong Kong RGC/GRF; CUHK VC's One-Off Discretionary Fund; Natural Science Foundation of China U1330203.\\ The authors thank Marlan O. Scully for helpful discussions. \bibliographystyle{apsrev4-1} \begin{thebibliography}{10} \expandafter\ifx\csname url\endcsname\relax \def\url#1{\texttt{#1}}\fi \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi \providecommand{\bibinfo}[2]{#2} \providecommand{\eprint}[2][]{} \bibitem{Klitzing1980} \bibinfo{author}{K. von Klitzing, G. Dorda and M. Pepper}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{45}}, \bibinfo{pages}{494--497} (\bibinfo{year}{1980}). \bibitem{Haldane1988} \bibinfo{author}{F. D.~M. Haldane}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{61}}, \bibinfo{pages}{2015--2018} (\bibinfo{year}{1988}). \bibitem{Kane2005} \bibinfo{author}{C. L. Kane} {and} \bibinfo{author}{E. J. Mele}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{95}}, \bibinfo{pages}{226801} (\bibinfo{year}{2005}). \bibitem{Bernevig2006} \bibinfo{author}{B.~A. Bernevig} {and} \bibinfo{author}{S. C. Zhang}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{96}}, \bibinfo{pages}{106802} (\bibinfo{year}{2006}). \bibitem{Konig2007} \bibinfo{author}{M. K\"onig, S. Wiedmann, C. Brüne, A. Roth, H. Buhmann, L. W. Molenkamp, X.-L. Qi, and S.-C. Zhang}, \newblock \emph{\bibinfo{journal}{Science}} \textbf{\bibinfo{volume}{318}}, \bibinfo{pages}{766--770} (\bibinfo{year}{2007}). \bibitem{Chang2013} \bibinfo{author}{C.-Z. Chang, J. Zhang, X. Feng, J. Shen, Z. Zhang, M. Guo, K. Li, Y. Ou, P. Wei, L.-L. Wang, J.-Q. Ji, Y. Feng, S. Ji, X. Chen, J. Jia, X. Dai, Z. Fang, S.-C. Zhang, K. He, Y. Wang, L. Lu, X.-C. Ma, Q.-K. Xue}, \newblock \emph{\bibinfo{journal}{Science}} \textbf{\bibinfo{volume}{340}}, \bibinfo{pages}{167--170} (\bibinfo{year}{2013}). \bibitem{Qi2009} \bibinfo{author}{X.-L. Qi, T. L. Hughes, S. Raghu, S.-C. Zhang}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{102}}, \bibinfo{pages}{187001} (\bibinfo{year}{2009}). \bibitem{Haldane2008} \bibinfo{author}{F. D.~M. Haldane} {and} \bibinfo{author}{S. Raghu}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{100}}, \bibinfo{pages}{013904} (\bibinfo{year}{2008}). \bibitem{HafeziM2013} \bibinfo{author}{M. Hafezi, S. Mittal, J. Fan, A. Migdall and J. M. Taylor}, \newblock \emph{\bibinfo{journal}{Nature Photonics}} \textbf{\bibinfo{volume}{7}}, \bibinfo{pages}{1001--1005} (\bibinfo{year}{2013}). \bibitem{Rechtsman2013} \bibinfo{author}{M.~C. Rechtsman, J. M. Zeuner, Y. Plotnik, Y. Lumer, D. Podolsky, F. Dreisow, S. Nolte, M. Segev and A. Szameit}, \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{496}}, \bibinfo{pages}{196--200} (\bibinfo{year}{2013}). \bibitem{Khanikaev2013} \bibinfo{author}{A.~B. Khanikaev, M. S. Hossein, W.-K. Tse, M. Kargarian, A. H. MacDonald and G. Shvets}, \newblock \emph{\bibinfo{journal}{Nature Materials}} \textbf{\bibinfo{volume}{12}}, \bibinfo{pages}{233--239} (\bibinfo{year}{2013}). \bibitem{Nalitov2015} \bibinfo{author}{A.~V. Nalitov, G. Malpuech, H. Terças and D. D. Solnyshkov}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{114}}, \bibinfo{pages}{026803} (\bibinfo{year}{2015}). \bibitem{Oka2009} \bibinfo{author}{T. Oka} {and} \bibinfo{author}{H. Aoki}, \newblock \emph{\bibinfo{journal}{Physical Review B}} \textbf{\bibinfo{volume}{79}}, \bibinfo{pages}{081406} (\bibinfo{year}{2009}). \bibitem{Inoue2010} \bibinfo{author}{J.-I. Inoue} {and} \bibinfo{author}{A. Tanaka}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{105}}, \bibinfo{pages}{017401} (\bibinfo{year}{2010}). \bibitem{Kitagawa2010} \bibinfo{author}{T. Kitagawa, E. Berg, M. Rudner and E. Demler}, \newblock \emph{\bibinfo{journal}{Physical Review B}} \textbf{\bibinfo{volume}{82}}, \bibinfo{pages}{235114} (\bibinfo{year}{2010}). \bibitem{Lindner2011} \bibinfo{author}{N.~H. Lindner, G. Refael, and V. Galitski}, \newblock \emph{\bibinfo{journal}{Nature Physics}} \textbf{\bibinfo{volume}{7}}, \bibinfo{pages}{490--495} (\bibinfo{year}{2011}). \bibitem{Jotzu2014} \bibinfo{author}{G. Jotzu, M. Messer, R. Desbuquois, M. Lebrat, T. Uehlinger, D. Greif and T. Esslinger}, \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{515}}, \bibinfo{pages}{237--240} (\bibinfo{year}{2014}). \bibitem{Scully2006} \bibinfo{author}{M.~O. Scully, E. S. Fry, C. H. R. Ooi and K. Wódkiewicz}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{96}}, \bibinfo{pages}{010501} (\bibinfo{year}{2006}). \bibitem{Wang2015} \bibinfo{author}{D.-W. Wang, R.-B. Liu, S.-Y. Zhu and M. O. Scully}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{114}}, \bibinfo{pages}{043602} (\bibinfo{year}{2015}). \bibitem{Boller1991} \bibinfo{author}{K.-J. Boller, A. Imamo{\v g}lu and S. E. Harris}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{66}}, \bibinfo{pages}{2593--2596} (\bibinfo{year}{1991}). \bibitem{Dicke1954} \bibinfo{author}{R.~H. Dicke}, \newblock \emph{\bibinfo{journal}{Physical Review}} \textbf{\bibinfo{volume}{93}}, \bibinfo{pages}{99--110} (\bibinfo{year}{1954}). \bibitem{Shirley1965} \bibinfo{author}{J.~H. Shirley}, \newblock \emph{\bibinfo{journal}{Physical Review}} \textbf{\bibinfo{volume}{138}}, \bibinfo{pages}{B979--B987} (\bibinfo{year}{1965}). \bibitem{Lumer2013} \bibinfo{author}{Y. Lumer, Y. Plotnik, M. C. Rechtsman, and M. Segev}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{111}}, \bibinfo{pages}{243905} (\bibinfo{year}{2013}). \bibitem{Wang2013} \bibinfo{author}{D.-W. Wang, H.-T. Zhou, M.-J. Guo, J.-X. Zhang, J. Evers and S.-Y. Zhu}, \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{110}}, \bibinfo{pages}{093901} (\bibinfo{year}{2013}). \bibitem{Tiwari2010} \bibinfo{author}{V.~B. Tiwari, S. Singh, H. S. Rawat, M. P. Singh and S. C. Mehendale}, \newblock \emph{\bibinfo{journal}{Journal of Physics B}} \textbf{\bibinfo{volume}{43}}, \bibinfo{pages}{095503} (\bibinfo{year}{2010}). \bibitem{Thiel2011} \bibinfo{author}{C.~W. Thiel, T. B\"ottger, and R. L. Cone}, \newblock \emph{\bibinfo{journal}{Journal of Luminescence}} \textbf{\bibinfo{volume}{131}}, \bibinfo{pages}{353--361} (\bibinfo{year}{2011}). \bibitem{Zhang2001} \bibinfo{author}{S.-C. Zhang} {and} \bibinfo{author}{J. Hu}, \newblock \emph{\bibinfo{journal}{Science}} \textbf{\bibinfo{volume}{294}}, \bibinfo{pages}{823--828} (\bibinfo{year}{2001}). \end{thebibliography} \begin{center} \textbf{\Large{Supplementary I}} \end{center} \setcounter{equation}{0} \renewcommand{\theequation}{S.\arabic{equation}} \setcounter{figure}{0} \renewcommand{\thefigure}{S.\arabic{figure}} \section{Effective Hamiltonian} The three-level atoms we use to construct the two-dimensional (2D) superradiance lattice (SL) have a ground state $|g\ra$, an excited state $|e\ra$ and a third state $|m\ra$. The optical fields that couple $|e\ra$ and $|m\ra$ have three modes with wave vectors $\mathbf{k}_1=-k_c\hat{x}$, $\mathbf{k}_2=k_c(\frac{1}{2}\hat{x}-\frac{\sqrt{3}}{2}\hat{y})$ and $\mathbf{k}_3=k_c(\frac{1}{2}\hat{x}+\frac{\sqrt{3}}{2}\hat{y})$. The interaction Hamiltonian with rotating-wave approximation is \begin{equation} H_c=-\hbar\sum\limits_{l=1}^{3}\sum\limits_{j=1}^{N}\kappa_l a_l e^{i\mathbf{k}_l\cdot \mathbf{r}_j}|e_j\ra\la m_j|+H.c., \end{equation} where $\kappa_l$ and $a_l$ are the vacuum coupling strength and annihilation operator of the $l$th mode, respectively. $N$ is the number of atoms and $\mathbf{r}_j$ is the position of the $j$th atom. Initially the atomic ensemble is in the ground state $|G\ra=|g_1, g_2,...,g_N\ra$. A weak probe field with wave vector $\mathbf{k}_p$ can prepare the atomic ensemble in the timed Dicke state (TDS) \begin{equation} |e_{\mathbf{k}_p}\ra=\frac{1}{\sqrt{N}}\sum\limits_{j=1}^{N}e^{i\mathbf{k}_p\cdot \mathbf{r}_j}|g_1, g_2,...,e_j,...,g_N\ra \end{equation} by the collective absorption of a single photon. The combinational quantum states of the atoms and three coupling modes can be written as $|e_{\mathbf{k}_p},n_1,n_2,n_3\ra$, where $n_l$ $(l=1,2,3)$ is the photon number in mode $\mathbf{k}_l$. This state is coupled to $|m_{\mathbf{k}_p-\mathbf{k}_2},n_1,n_2+1,n_3\ra$ with coupling strength $-\hbar\kappa_2\sqrt{n_2+1}$. The state $|m_{{\mathbf k}_p-{\mathbf k}_2}\ra$ is defined the same as in equation () with $e_j$ replaced by $m_j$ and ${\mathbf k}_p$ replaced by ${\mathbf k}_p-{\mathbf k}_2$. Similar coupling also exists for the other two modes. $|m_{\mathbf{k}_p-\mathbf{k}_2},n_1,n_2+1,n_3\ra$ is in turn coupled to $|e_{\mathbf{k}_p+\mathbf{k}_1-\mathbf{k}_2},n_1-1,n_2+1,n_3\ra$ or $|e_{\mathbf{k}_p+\mathbf{k}_3-\mathbf{k}_2},n_1,n_2+1,n_3-1\ra$ via excitation by modes $\mathbf{k}_1$ or $\mathbf{k}_3$. These states form a honeycomb lattice with discrete momentum coordinates, as shown in Fig. , called the superradiance lattice . The two sublattices of the SL correspond to TDS for $|e\ra$ and $|m\ra$. We denote the coupling field frequency as $\nu_c$ and the transition frequency between $|e\ra$ and $|m\ra$ as $\omega_{em}$. The energy difference between the two sublattice is $\Delta_c=\omega_{em}-\nu_c$. We set the zero energy at the middle of the energies of the two sublattices. Then the energies of the $|e\ra$ and $|m\ra$ sublattices are $\Delta_c/2$ and $-\Delta_c/2$, respectively. Although the coupling strengths in the SL are site-dependent, if we use coherent coupling fields with large average photon numbers $\left\langle n_l\right\rangle\gg 1$, the Rabi frequency of the $l$th mode can be approximated as the classical Rabi frequency $\Omega_l=\kappa_l\sqrt{\left\langle n_l\right\rangle}$. The SL with quantized photon numbers has edges when either of the three coupling field photon numbers reduces to zero. The total lattice has a triangular boundary. However, since the TDS have finite life time, which we suppose to be $\tau$ on average, once we create an excitation $|e_{\mathbf{k}_p}\ra$ in the SL, the average distance this excitation can travel is in the order of $\Omega_l\tau$. In this paper we have $\Omega_l\tau\ll \la n_l\ra$ and the edges can never be reached. Interesting edge effects might exist in the few-photon limit and it will be discussed elsewhere. Since the photon numbers are correlated to the momenta $\mathbf{k}$'s in the TDS, we can simplify the notations by dropping the photon numbers. The Hamiltonian of these TDS is \begin{equation} \begin{aligned} H=&\frac{\hbar\Delta_c}{2}\sum\limits_{\mathbf{k}}(|e_\mathbf{k}\ra\la e_\mathbf{k}|-|m_{\mathbf{k}-\mathbf{k}_1}\ra\la m_{\mathbf{k}-\mathbf{k}_1}|)\\ &-\sum\limits_{\mathbf{k}}\sum\limits_{l=1}^{3}\hbar\Omega_l [|e_{\mathbf{k}}\ra\la m_{\mathbf{k}-\mathbf{k}_l}|+h.c.], \end{aligned} \end{equation} where $\mathbf{k}=\mathbf{k}_p+r(\mathbf{k}_2-\mathbf{k}_1)+s(\mathbf{k}_3-\mathbf{k}_2)$ with integers $r$ and $s$. The next-nearest-neighbour (NNN) hopping terms are induced by the periodic modulation of the three coupling field Rabi frequencies, \begin{equation} \Omega_l=\Omega_s+2\Omega_d\cos(\nu_d t-\mathbf{k}_d^l\cdot\mathbf{r}+\phi_l), \end{equation} where $\Omega_{s}$ and $\Omega_{d}$ are the static and dynamic components of the Rabi frequencies. $\nu_d$ is the modulation frequency. $\phi_l$ is the modulation phase of the $l$th field. $\mathbf{k}_d^l$ is the modulation wavevector. We choose the atomic ensemble much smaller than $1/|\mathbf{k}_d^l|$ such that the position dependence of the phase $-\mathbf{k}_d^l\cdot\mathbf{r}$ can be neglected. We set $-\mathbf{k}_d^l\cdot\mathbf{r}=0$ for simplicity. On the other hand, the size of the atomic ensemble shall be much larger than $1/\nu_c$ and the total number of atoms $N\gg 1$ such that \begin{equation} \la e_{\mathbf{k}^\prime}|e_\mathbf{k}\ra=\frac{1}{N}\sum\limits_{j=1}^{N}e^{i(\mathbf{k}-\mathbf{k}^\prime)\cdot\mathbf{r}_j}\approx \delta_{\mathbf{k}\mathbf{k}^\prime}. \end{equation} We assume the atoms be randomly distributed and their number is large enough to cover all possible points in the real space Brillouin Zone. The SL can be regarded as infinite. We expand the Hamiltonian into static, positive- and negative-frequency components, \begin{equation} H=H_0+H_{+1}e^{i\nu_dt}+H_{-1}e^{-i\nu_dt}, \end{equation} where \begin{equation} \begin{aligned} H_0=&\frac{\hbar\Delta_c}{2}\sum\limits_{\mathbf{k}}(|e_\mathbf{k}\ra\la e_\mathbf{k}|-|m_{\mathbf{k}-\mathbf{k}_1}\ra\la m_{\mathbf{k}-\mathbf{k}_1}|)\\ &-\hbar\Omega_s\sum\limits_{\mathbf{k}}\sum\limits_{l=1}^{3} [|e_{\mathbf{k}}\ra\la m_{\mathbf{k}-\mathbf{k}_l}|+h.c.], \end{aligned} \end{equation} \begin{equation} H_{+1}=-\hbar\Omega_d\sum\limits_{\mathbf{k}}\sum\limits_{l=1}^{3}e^{i\phi_l} [|e_{\mathbf{k}}\ra\la m_{\mathbf{k}-\mathbf{k}_l}|+h.c.], \end{equation} \begin{equation} H_{-1}=-\hbar\Omega_d\sum\limits_{\mathbf{k}}\sum\limits_{l=1}^{3}e^{-i\phi_l} [|e_{\mathbf{k}}\ra\la m_{\mathbf{k}-\mathbf{k}_l}|+h.c.]. \end{equation} Note that $H_{\pm 1}$ are not Hermitian themselves, but $H_{+1}$ is the Hermitian conjugate of $H_{-1}$. The phase factors $e^{\pm i\phi_l}$ in $H_{\pm 1}$ play the crucial role for the complex NNN hopping terms in the Haldane model. The dynamics of the system is a Floquet problem . Based on the Floquet theorem, the wave function can be written as \begin{equation} |\Psi\ra=e^{-i\epsilon t/\hbar}|\psi(t)\ra, \end{equation} where $\epsilon$ is the quasi-eigenenergy and \begin{equation} |\psi(t)\ra=\sum_n e^{in\nu_dt}|\psi_n\ra, \end{equation} is a periodic wave function with period $2\pi/\nu_d$. Substituting equations ()-() to the Schr\"odinger equation, $i\hbar\partial|\Psi\ra/\partial t=H|\Psi\ra$, we obtain \begin{equation} \begin{aligned} &i\hbar\frac{\partial}{\partial t}e^{-i\epsilon t/\hbar}\sum_n e^{in\nu_dt}|\psi_n\ra\\ &=(H_0+H_{+1}e^{i\nu_dt}+H_{-1}e^{-i\nu_dt})e^{-i\epsilon t/\hbar}\sum_n e^{in\nu_dt}|\psi_n\ra. \end{aligned} \end{equation} The terms with the same time evolution phase factors should be equal on both sides. We therefore have \begin{equation} (\epsilon-\hbar n\nu_d)|\psi_n\ra=H_0|\psi_n\ra+H_{+1}|\psi_{n-1}\ra+H_{-1}|\psi_{n+1}\ra. \end{equation} The quasi-eigenenergy can be obtained by diagonalizing the above Hamiltonian. For the sake of simplicity, we assume the separation between the Floquet sidebands is much larger than the bandwidth, $\nu_d\gg\Omega_{s,d},\Delta_c$, where the perturbation theory can be applied . When the probe field is near resonance, $\Delta_p=\omega_{eg}-\nu_p\ll \Omega_s$ where $\omega_{eg}$ is the transition frequency between $|e\ra$ and $|g\ra$ and $\nu_p$ is the probe field frequency, only the Floquet band with $n=0$ in Eq.() is relevant. For states with eigenfrequencies near $\epsilon=0$, the effective Hamiltonian can be obtained by standard second-order perturbation as \begin{equation} \epsilon|\psi_0\ra=H_\text{eff}|\psi_0\ra, \end{equation} where \begin{equation} H_{\text{eff}}=H_0+H^\prime, \end{equation} with NNN Hamiltonian \begin{equation} H^\prime=\frac{1}{\hbar\nu_d}(H_{+1}H_{-1}-H_{-1}H_{+1}). \end{equation} Explicitly, \begin{equation} \begin{aligned} H^\prime=&\sum\limits_{\mathbf{k}}\sum\limits_{l\ne j=1}^{3}\hbar\Omega_{lj}\\ &\left(|e_{\mathbf{k}+\mathbf{k}_l-\mathbf{k}_j}\ra \la e_{\mathbf{k}}|+|m_{\mathbf{k}-\mathbf{k}_1}\ra\la m_{\mathbf{k}-\mathbf{k}_1+\mathbf{k}_l-\mathbf{k}_j}|\right), \end{aligned} \end{equation} where $\Omega_{lj}=\Omega^\prime[e^{i(\phi_l-\phi_j)}-e^{i(\phi_j-\phi_l)}]=2i\Omega^\prime\sin(\phi_l-\phi_j)$ with $\Omega^\prime=\Omega^2_d/\nu_d$. The crucial factor $i =\sqrt{-1}$ comes from the quantum interference between the two pathways shown in Fig. (a). The loop transitions via NNN hopping accumulate nonzero phases due to these complex factors, as shown in Fig. (b). The effective Hamiltonian is greatly simplified in the real-space representation. We denote the real-space basis states as \begin{equation} \begin{aligned} |e_{\mathbf{r}_j}\rangle&=|g_1,g_2,\ldots, e_j,\ldots, g_N\rangle,\\ |m_{\mathbf{r}_j}\rangle&=|g_1,g_2,\ldots, m_j,\ldots, g_N\rangle, \end{aligned} \end{equation} The effective Hamiltonian can be written as \begin{equation} H_{\text{eff}}=\sum_{j}{\mathbf h}({\mathbf r}_j)\cdot{\boldsymbol\sigma}_j \end{equation} where the effective magnetic field ${\mathbf h}({\mathbf r}_j)=(h_x,h_y,h_z)$ with \begin{equation} h_x=-\hbar\Omega_s\sum\limits_{l=1}^3\cos(\mathbf{r}_j\cdot\mathbf{k}_l), \end{equation} \begin{equation} h_y=\hbar\Omega_s\sum\limits_{l=1}^3\sin(\mathbf{r}_j\cdot\mathbf{k}_l), \end{equation} \begin{equation} h_z=\frac{\hbar\Delta_c}{2}+2i\hbar\sum\limits_{l=1}^3\Omega_{(l+1)l}\sin[\mathbf{r}_j\cdot(\mathbf{k}_{l+1}-\mathbf{k}_{l})], \end{equation} and the pseudo spin ${\boldsymbol\sigma}_j=(\sigma_j^x,\sigma_j^y,\sigma_j^z)$ with the Pauli matrices for the $j$th atom defined as $\sigma_j^x=|e_{{\mathbf r}_j}\rangle\langle m_{{\mathbf r}_j}|+|m_{{\mathbf r}_j}\rangle\langle e_{{\mathbf r}_j}|$, $\sigma_j^y=-i|e_{{\mathbf r}_j}\rangle\langle m_{{\mathbf r}_j}|+i|m_{{\mathbf r}_j}\rangle\langle e_{{\mathbf r}_j}|$ and $\sigma_j^z=|e_{{\mathbf r}_j}\rangle\langle e_{{\mathbf r}_j}|-|m_{{\mathbf r}_j}\rangle\langle m_{{\mathbf r}_j}|$. \section{Berry connection and Berry curvature} The topological properties of the wavefunctions are determined by the Berry connection $\mathbf{A}$ and Berry curvature $\mathbf{B}$, \begin{equation} \mathbf{A}=i\la\psi|\nabla|\psi\ra, \end{equation} \begin{equation} \mathbf{B}=\nabla\times\mathbf{A}, \end{equation} where $\nabla$ and $\nabla\times$ are the gradient and curl operators with respect to $\mathbf{r}$. Similar to a spin-1/2 in a magnetic field ${\mathbf h}$, the eigenenergies of the SL eigenstates of the effective Hamiltonian in equation () are \begin{equation} \epsilon_\pm=\pm h \end{equation} where $\pm$ are for the upper and lower bands and $h$ is the magnitude of $\mathbf{h}$. The eigen wavefunction in the upper band can be written as \begin{equation} |\psi_+(\mathbf{r})\ra=\frac{1}{\sqrt{2h(h+h_z)}} \left( \begin{array}{ccc} h+h_z\\ h_x+ih_y \end{array} \right). \end{equation} This wavefunction is well defined except for the south pole of the Bloch sphere where $\mathbf{h}=(0,0,h_z)$ with $h_z=-h$. The eigen wavefunction near the south pole can be written as \begin{equation} |\psi_+^\prime(\mathbf{r})\ra=e^{i\phi(\mathbf{r})}|\psi_+(\mathbf{r})\ra=\frac{1}{\sqrt{2h(h-h_z)}} \left( \begin{array}{ccc} h_x-ih_y \\ h-h_z \end{array} \right), \end{equation} where the gauge transformation \begin{equation} e^{i\phi({\mathbf{r}})}=\frac{h_x-ih_y}{|h_x-ih_y|}. \end{equation} This way, the whole Bloch sphere is fully covered by two gauges . The Berry connection in the upper band (except for the south pole of the Bloch sphere) is \begin{equation} \begin{aligned} \mathbf{A}_+=&i\la\psi_+|\nabla|\psi_+\ra\\ =&-\frac{1}{2h(h+h_z)}\left(h_x\nabla h_y-h_y\nabla h_x\right), \end{aligned} \end{equation} and near the south pole we can use \begin{align} \mathbf{A}^\prime_+=\mathbf{A}_+-\nabla\phi(\mathbf{r}). \end{align} The Berry curvature is \begin{equation} \begin{aligned} \mathbf{B}_+=&\nabla\times\mathbf{A}_+\\ =&-2\hat{z}\text{Im}\la\partial_x\psi_+|\partial_y\psi_+\ra\\ =&-\frac{\hat{z}}{2h^3}\epsilon_{abc}h_a\partial_xh_b\partial_yh_c, \end{aligned} \end{equation} where $\epsilon_{abc}$ with $a,b,c\in{x,y,z}$ is the Levi-Civita symbol. The Chern number is defined as the total Berry curvature in the whole first Brillouin zone \begin{equation} \begin{aligned} C=&\frac{1}{2\pi}\oiint_{\text{BZ}} \mathbf{B}_+\cdot\,d\mathbf{S}\\ =&-\frac{1}{4\pi}\oiint_{\text{BZ}} \frac{1}{h^3}\epsilon_{abc}h_a\partial_xh_b\partial_yh_c \,dx \,dy, \end{aligned} \end{equation} which counts the winding number of the effective magnetic field $\mathbf{h}$ wrapping around the Bloch sphere in the whole Brillouin zone . In Fig., we plot $\mathbf{h}$ in the Brillouin zone for a topological nontrivial SL where the band gap is opened by the NNN hopping. If we go from the $K$ point $\mathbf{r}_+=-\frac{4\sqrt{3}\pi}{9k_c}\hat{y}$ to the $K^\prime$ point $\mathbf{r}_-=\frac{4\sqrt{3}\pi}{9k_c}\hat{y}$, $\mathbf{h}$ moves from the north pole to the south pole of the Bloch sphere. The Chern number is one. If the band gap is opened by the on-site offset, $\Delta_c$, $h_z$ is a constant in the Brillouin zone. $\mathbf{h}$ can only cover a patch on the Bloch sphere and the Chern number is zero. The two bands of $H_{\text{eff}}$ have the smallest gap at the $K$ and $K^\prime$ points $\mathbf{r}_\pm$, where \begin{equation} \sum\limits_{l=1}^{3} e^{i\mathbf{r}_\pm\cdot\mathbf{k}_l}=0, \end{equation} and hence $h_x=h_y=0$. At these symmetry points of the Brillouin zone the Hamiltonian is diagonal, \begin{equation} \begin{aligned} \mathcal{H}(\mathbf{r}_\pm)=[\frac{\hbar\Delta_c}{2}\pm 2\sqrt{3}\Omega^\prime\hbar\alpha]\sigma_z, \end{aligned} \end{equation} where $\alpha=-\sum_{l=1}^3\sin(\phi_{l+1}-\phi_l)$. Let us consider the specific case that $\Delta_c=0$, $\phi_l=(l-1)4\pi/3$, and thus $\alpha=3\sqrt{3}/2$. In this case $\mathcal{H}(\mathbf{r}_\pm)=\pm 9\hbar\Omega^\prime\sigma_z$ has opposite signs at $\mathbf{r}_\pm$, as shown in Fig. (a). It is obvious that at $\mathbf{r}_+$, the eigenstate in the upperband with eigenenergy $9\hbar\Omega^\prime$ is $|e\ra$, while at $\mathbf{r}_-$, $|e\ra$ is the eigenstate in the lower band, as shown in Fig. (a). Near $\mathbf{r}_+$, we can use equation ({}) to describe the wavefunction. However, equation ({}) cannot describe all the wavefunctions in the upper band. It has a singularity at $\mathbf{r}_-$ where the magnetic field ${\mathbf h}$ points to the south pole. We can remove the singularity by tuning $\Delta_c$. When $\Delta_c>4\sqrt{3}\Omega^\prime\alpha$, the effective magnetic field $\mathbf{h}$ does not experience the whole Bloch sphere [see Fig. (b)]. In this case, $h_z$ in equation () has the same sign for $\mathbf{r}_\pm$, $|e\ra$ is in the upper band at both $\mathbf{r}_\pm$, as shown in Fig. (b). The wavefunction can be described by the same gauge. Thus the single-valued Berry connection $\mathbf{A}$ has no singularities on a closed surface and the Chern number is zero. Generally, the Chern number of the upper band can be written as \begin{equation} C=\frac{1}{2}\left[\text{sign}\left(\Delta_c+4\sqrt{3}\Omega^\prime\alpha\right)-\text{sign}\left(\Delta_c-4\sqrt{3}\Omega^\prime\alpha\right)\right]. \end{equation} Specifically, when $\Delta_c=0$, $C=\text{sign}(\alpha)$. \section{Dynamic evolution} We can solve the dynamics of the atomic ensemble in real space with the total Hamiltonian including the probe field, \begin{equation} \begin{aligned} H_t=&\sum\limits_{j=1}^{N}(h_z-\hbar\tilde{\nu}_p)|e_{\mathbf{r}_j}\ra \la e_{\mathbf{r}_j}|-(h_z+\hbar\tilde{\nu}_p)|m_{\mathbf{r}_j}\ra | m_{\mathbf{r}_j}|\\ &+[(h_x-ih_y)|e_{\mathbf{r}_j}\ra \la m_{\mathbf{r}_j}|+h.c.]\\ &-[\hbar\Omega_p e^{i\mathbf{k}_p\cdot\mathbf{r}_j}|e_{\mathbf{r}_j}\ra \la G|+h.c.], \end{aligned} \end{equation} with $\tilde{\nu}_p=\Delta_c/2-\Delta_p$ being the probe detuning with respect to the middle of the band gap and $\Omega_p$ the probe field Rabi frequency. The wavefunction in real space \begin{equation} |\Psi\ra=c_G|G\ra+\sum\limits_{j=1}^{N}c_e({\mathbf{r}_j})|e_{\mathbf{r}_j}\ra+c_m({\mathbf{r}_j})|m_{\mathbf{r}_j}\ra. \end{equation} The dynamic equations of the probability amplitudes are \begin{equation} \begin{aligned} \dot{c}_e(\mathbf{r}_j)=&\left[-\frac{i}{\hbar}(h_z-\hbar\tilde{\nu}_p)-\gamma_e\right]c_e(\mathbf{r}_j)\\ &-\frac{i}{\hbar}(h_x-ih_y)c_m(\mathbf{r}_j)+i\Omega_p e^{i\mathbf{k}_p\cdot\mathbf{r}_j}c_G, \end{aligned} \end{equation} \begin{equation} \begin{aligned} \dot{c}_m(\mathbf{r}_j)=&\left[\frac{i}{\hbar}(h_z+\hbar\tilde{\nu}_p)-\gamma_m\right]c_m(\mathbf{r}_j)\\ &-\frac{i}{\hbar}(h_x+ih_y)c_e(\mathbf{r}_j). \end{aligned} \end{equation} where $\gamma_{e/m}$ is the decoherence rate of $|e/m\ra$ states. In the limit of weak probe field $\Omega_p\ll\gamma_e$, we have $c_G\approx 1$. In the steady state, $\dot{c}_e(\mathbf{r}_j)=\dot{c}_m(\mathbf{r}_j)=0$. From Eq.(), we obtain \begin{equation} c_m(\mathbf{r}_j)=\frac{h_x+ih_y}{h_z+\hbar\tilde{\nu}_p+i\hbar\gamma_m}c_e(\mathbf{r}_j). \end{equation} Substituting Eq.() in Eq.(), we obtain \begin{equation} c_e(\mathbf{r}_j)=\frac{\hbar\Omega_p e^{i\mathbf{k}_p\cdot\mathbf{r}_j}}{h_z-\hbar\tilde{\nu}_p-i\hbar\gamma_e+\frac{h_x^2+h_y^2}{h_z+\hbar\tilde{\nu}_p+i\hbar\gamma_m}}. \end{equation} In the SL coordinates, the wavefunction can be written as \begin{equation} |\Psi\ra=c_G|G\ra+\sum\limits_{\mathbf{k}}c_\mathbf{k}|e_\mathbf{k}\ra+c_{\mathbf{k}-\mathbf{k}_1}|m_{\mathbf{k}-\mathbf{k}_1}\ra. \end{equation} where the probability amplitude \begin{equation} c_\mathbf{k}=\la e_\mathbf{k}|\Psi\ra=\frac{1}{\sqrt{N}}\sum\limits_{j=1}^{N}e^{-i\mathbf{k}\cdot\mathbf{r}_j}c_e(\mathbf{r}_j). \end{equation} For uniformly distributed atoms, in the limit $N\gg\Omega_s/\gamma_e$, we can assume all points in the Brillouin zone are occupied by atoms, so the summation can be written as integration in the first Brillouin zone, \begin{equation} c_\mathbf{k}=\frac{\sqrt{N}}{S}\oiint_{\text{BZ}}c_e(\mathbf{r})e^{-i\mathbf{k}\cdot\mathbf{r}}\,dx\,dy, \end{equation} where $S$ is the area of the first Brillouin zone. \section{Diffraction Contrast and topology} In this section, we quantitatively compare the contrast between $|c_{\mathbf{k}_+}|^2$ and $|c_{\mathbf{k}_-}|^2$ in topological and trivial SL's. We assume $\tilde{\nu}_p=0$ and $\gamma_m=0$. Then Eq.() becomes \begin{equation} c_e(\mathbf{r})=\frac{ h_z}{h^2-i\hbar h_z\gamma_e}\hbar\Omega_p e^{i\mathbf{k}_p\cdot\mathbf{r}}. \end{equation} We assume $\Omega^\prime,\Delta_c\ll\Omega_s$ and thus $c_e(\mathbf{r})$ is highly centred at $K$ and $K^\prime$ points, where we have $h_x=h_y=0$ and \begin{equation} c_e(\mathbf{r})=\frac{\hbar\Omega_p}{h_z-i\hbar\gamma_e} e^{i\mathbf{k}_p\cdot\mathbf{r}}. \end{equation} According to Eq.(), the probability amplitude $c_\mathbf{k}$ can be approximately calculated by the integration of $c_e(\mathbf{r})$ in small areas near $\mathbf{r}_\pm$, \begin{equation} \begin{aligned} c_\mathbf{k}\approx&\frac{\sqrt{N}\hbar\Omega_p}{S}[e^{i(\mathbf{k}_p-\mathbf{k})\cdot\mathbf{r}_+}\iint\limits_{\mathbf{r}_+}(h_z-i\hbar\gamma_e)^{-1}\,dx\,dy\\ &+e^{i(\mathbf{k}_p-\mathbf{k})\cdot\mathbf{r}_-}\iint\limits_{\mathbf{r}_-}(h_z-i\hbar\gamma_e)^{-1}\,dx\,dy], \end{aligned} \end{equation} where the phase factors have been taken out of the integration since they do not change appreciably in those small areas. For a topological SL where $\Omega^\prime\ne 0$ and $\Delta_c=0$, $h_z$ has opposite signs at $\mathbf{r}_\pm$. We denote $\iint_{\mathbf{r}_+}(h_z-i\hbar\gamma_e)^{-1}\,dx\,dy=p+iq$, then $\iint_{\mathbf{r}_-}(h_z-i\hbar\gamma_e)^{-1}\,dx\,dy=-p+iq$. Thus \begin{equation} \begin{aligned} c_\mathbf{k}\approx 2i\frac{\sqrt{N}\hbar\Omega_p}{S}\{p\sin[{(\mathbf{k}_p-\mathbf{k})\cdot\mathbf{r}_+}]+q\cos[{(\mathbf{k}_p-\mathbf{k})\cdot\mathbf{r}_+}]\}. \end{aligned} \end{equation} For $\mathbf{k}_p=-\mathbf{k}_1$, $\mathbf{k}_+=-\mathbf{k}_3$ and $\mathbf{k}_-=-\mathbf{k}_2$, we have \begin{equation} \begin{aligned} \eta\equiv\frac{|c_{\mathbf{k}_+}|^2-|c_{\mathbf{k}_-}|^2}{|c_{\mathbf{k}_+}|^2+|c_{\mathbf{k}_-}|^2}\approx\frac{2\sqrt{3}pq}{3p^2+q^2}. \end{aligned} \end{equation} For a trivial SL where $\Omega^\prime= 0$ and $\Delta_c\ne 0$, $h_z$ has the same sign at $\mathbf{r}_\pm$. We have $\iint_{\mathbf{r}_+}(h_z-i\hbar\gamma_e)^{-1}\,dx\,dy=\iint_{\mathbf{r}_-}(h_z-i\hbar\gamma_e)^{-1}\,dx\,dy=p+iq$ and \begin{equation} \begin{aligned} c_\mathbf{k}\approx 2\frac{\sqrt{N}\hbar\Omega_p}{S}(p+iq)\cos[{(\mathbf{k}_p-\mathbf{k})\cdot\mathbf{r}_+}]. \end{aligned} \end{equation} Therefore, $c_{\mathbf{k}_+}=c_{\mathbf{k}_-}$ and $\eta=0$. For a trivial SL with $\Omega^\prime\ne 0$ and $\Delta_c\ne 0$, it is difficult to analytically calculate the results. Nevertheless, as shown in Fig.3 (b) in the main text, $\eta$ reduces to a small value in the $C=0$ phases especially for large $\Omega^\prime$. \section{Coupled- wave equations} In this section, we calculate the propagation of a weak probe pulse. We set $\mathbf{k}_p=-\mathbf{k}_1$, $\mathbf{k}_+=-\mathbf{k}_3$ and $\mathbf{k}_-=-\mathbf{k}_2$. There will be directional emission in $\mathbf{k}_\pm$ by TDS $|e_{\mathbf{k}_\pm}\rangle$. The emitted photons in $\mathbf{k}_\pm$ interact with the atomic ensemble the same way as the probe field $\mathbf{k}_p$, creating TDS and resulting in directional emission in the other two modes. For example, emission at ${\mathbf k}_+$ excites the atom ensemble and leads to emission at ${\mathbf k}_-$ and ${\mathbf k}_{p}$. Therefore, the three optical fields in $\mathbf{k}_{p,\pm}$ are coupled via the atoms. In the following, we derive the coupling equations. The expectation value of the dipole moment of the $j$th atom at position $\mathbf{r}_j$ is \begin{equation} \begin{aligned} \la\Psi|\hat{\mu}_j|\Psi\ra=&\sum\limits_{\mathbf{k}}c_\mathbf{k}\la G|\hat{\mu}_j|e_\mathbf{k}\ra+c.c.\\ =&\frac{\mu}{\sqrt{N}}\sum\limits_{\mathbf{k}}{c}_\mathbf{k}e^{i\mathbf{k}\cdot\mathbf{r}_j}+c.c., \end{aligned} \end{equation} where $\hat{\mu}_j$ is the dipole operator of the $j$th atom and $\mu=\la g_j|\hat{\mu}_j|e_j\ra$. Since the atoms are homogeneously distributed, the polarization density as a function of positions is, \begin{equation} \begin{aligned} P(\mathbf{r})=&\frac{N}{V}\frac{\mu}{\sqrt{N}}\sum\limits_{\mathbf{k}}{c}_\mathbf{k}e^{i\mathbf{k}\cdot\mathbf{r}}+c.c., \end{aligned} \end{equation} where $V$ is the volume of the atomic ensemble. The probe field Rabi frequency is $\Omega_p=\mu E_p/\hbar$ with $E_p$ being the probe field strength. The polarization density contains the Fourier components in $\mathbf{k}_\pm$, $P_{\pm, p}=\epsilon_0\chi_\pm E_p$ with $\chi_\pm=\sqrt{N} \mu{c}_{\mathbf{k}_\pm}/V\epsilon_0E_p$ and $\epsilon_0$ is the permittivity in vacuum. The notation $P_{\pm, p}$ means the $\mathbf{k}_\pm$ Fourier component of polarization density generated by optical field in $\mathbf{k}_p$. Once modes $\mathbf{k}_\pm$ are excited and the corresponding fields $E_\pm$ are generated, they also polarize the atoms. For $\phi_1=0$, $\phi_2=4\pi/3$ and $\phi_3=2\pi/3$, the three fields have a cyclic relation between each other. We have $P_{p,+}=\epsilon_0\chi_- E_+$, $P_{-,+}=\epsilon_0\chi_+ E_+$, $P_{p,-}=\epsilon_0\chi_+ E_-$ and $P_{+,-}=\epsilon_0\chi_- E_-$. We assume the quasi-static approximation in which the atoms are assumed in steady sates, i.e., $c_{\mathbf k}$ is given by equation (). This approximation is justified when the pulse duration is much longer than the decoherence time $1/\gamma_e$. We also use the slowly-varying-envelop approximation. The fields in the three relevant modes $\mathbf{k}_{j}$ with $j=p,+$ and $-$ are denoted as $E_{j}(\mathbf{r})e^{-i\nu_pt+i\mathbf{k}_{j}\cdot\mathbf{r}}$ where $|\mathbf{k}_{j}|=\nu_p/c\equiv k_p$. The coupled-wave Maxwell equations for the three modes are \begin{equation} \left(\hat{\mathbf{k}}_p\cdot\nabla+\frac{1}{v_g}\frac{\partial}{\partial t}\right)E_p=\frac{ik_p}{2}\left(\chi_0 E_p+\chi_-E_++\chi_+E_-\right), \end{equation} \begin{equation} \left(\hat{\mathbf{k}}_+\cdot\nabla+\frac{1}{v_g}\frac{\partial}{\partial t}\right)E_+=\frac{ik_p}{2}\left(\chi_0 E_+ +\chi_-E_-+\chi_+E_p\right), \end{equation} \begin{equation} \left(\hat{\mathbf{k}}_-\cdot\nabla+\frac{1}{v_g}\frac{\partial}{\partial t}\right)E_-=\frac{ik_p}{2}\left(\chi_0 E_-+\chi_-E_p+\chi_+E_+\right), \end{equation} where $\hat{\mathbf{k}}_{p/+/-}$ are the unit vectors in the directions of $\mathbf{k}_{p/+/-}$ and $v_g$ is the group velocity of the pulses. \section{Local and global effects} In the main text, we use the chiral topological current to explain the consistence between $\eta$ and the topological order. The topological current is a global effect for which we must treat the lattice as a whole. Since the three superradiant TDS $|e_{\mathbf{k}_{p,\pm}}\ra$ are next nearest neighbours in the lattice, the local effect induced by the transitions inside a unit cell is also important. The direct NNN transition strength from $|e_{\mathbf{k}_p}\ra$ to $|e_{\mathbf{k}_+}\ra$ or $|e_{\mathbf{k}_-}\ra$ is $\Omega_{13}$ or $\Omega_{12}$, respectively. If the dynamics is dominated by these direct transitions, we expect the contrast of the probability $\eta$ should be similar to $(|\Omega_{13}|-|\Omega_{12}|)/(|\Omega_{13}|+|\Omega_{12}|)$, which is plotted in Fig.. We plot the contrast $\eta$ in Fig. for different $\Omega^\prime/\gamma_m$. The larger $\Omega^\prime$ is, the closer $\eta$ is to Fig. and the more important is the local effect, as shown in Fig. (a) and (b). For $\Omega^\prime\le\gamma_m$, the effect of the direct NNN transitions is suppressed by the decoherence rate $\gamma_m$. The global topological effect dominates for most areas and $\eta$ is closer to Fig.2 (a) in the main text, as shown Fig. (c) and (d). \begin{thebibliography}{10} \expandafter\ifx\csname url\endcsname\relax \def\url#1{\texttt{#1}}\fi \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi \providecommand{\bibinfo}[2]{#2} \providecommand{\eprint}[2][]{} \bibitem{2014arXiv1403.7097W} \bibinfo{author}{D.-W. Wang}, \bibinfo{author}{R.-B. Liu, R.-B.}, \bibinfo{author}{S.-Y. Zhu} and \bibinfo{author}{M. O. Scully}, \newblock \bibinfo{title}{Superradiance lattice}. \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{114}}, \bibinfo{pages}{043602} (\bibinfo{year}{2015}). \bibitem{Shirley1965} \bibinfo{author}{J. H. Shirley}, \newblock \bibinfo{title}{Solution of the Schr\"odinger equation with a Hamiltonian periodic in time}. \newblock \emph{\bibinfo{journal}{Physical Review}} \textbf{\bibinfo{volume}{138}}, \bibinfo{pages}{B979--B987} (\bibinfo{year}{1965}). \bibitem{Kitagawa2011} \bibinfo{author}{T. Kitagawa}, \bibinfo{author}{T. Oka}, \bibinfo{author}{A. Brataas}, \bibinfo{author}{L. Fu} and \bibinfo{author}{E. Demler}, \newblock \bibinfo{title}{Transport properties of nonequilibrium systems under the application of light: Photoinduced quantum Hall insulators without Landau levels}. \newblock \emph{\bibinfo{journal}{Physical Review B}} \textbf{\bibinfo{volume}{84}}, \bibinfo{pages}{235108} (\bibinfo{year}{2011}). \bibitem{Liu2000} \bibinfo{author}{R.-B. Liu} and \bibinfo{author}{B.-F. Zhu}, \newblock \bibinfo{title}{Dynamic Fano resonance of Floquet-state excitons in superlattices}. \newblock \emph{\bibinfo{journal}{Journal of Physics: Condensed Matter}} \textbf{\bibinfo{volume}{12}}, \bibinfo{pages}{L741} (\bibinfo{year}{2000}). \bibitem{Gomez-Leon2013} \bibinfo{author}{A. G\'omez-Le\'on} and \bibinfo{author}{G. Platero}, \newblock \bibinfo{title}{Floquet-Bloch theory and topology in periodically driven lattices}. \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{110}}, \bibinfo{pages}{200403} (\bibinfo{year}{2013}). \bibitem{Wu1975} \bibinfo{author}{T. T. Wu} and \bibinfo{author}{C. N. Yang}, \newblock \bibinfo{title}{Concept of nonintegrable phase factors and global formulation of gauge fields}. \newblock \emph{\bibinfo{journal}{Physical Review D}} \textbf{\bibinfo{volume}{12}}, \bibinfo{pages}{3845--3857} (\bibinfo{year}{1975}). \bibitem{Qi2006} \bibinfo{author}{X.-L. Qi}, \bibinfo{author}{Y.-S. Wu} and \bibinfo{author}{S.-C. Zhang}, \newblock \bibinfo{title}{Topological quantization of the spin Hall effect in two-dimensional paramagnetic semiconductors}. \newblock \emph{\bibinfo{journal}{Physical Review B}} \textbf{\bibinfo{volume}{74}}, \bibinfo{pages}{085308} (\bibinfo{year}{2006}). \bibitem{Katan2013} \bibinfo{author}{Y. T. Katan} and \bibinfo{author}{D. Podolsky}, \newblock \bibinfo{title}{Modulated Floquet topological insulators}. \newblock \emph{\bibinfo{journal}{Physical Review Letters}} \textbf{\bibinfo{volume}{110}} (\bibinfo{year}{2013}). \end{thebibliography} |
1501.04100 | Title: Spatial Interpolants
Abstract: We propose Splinter, a new technique for proving properties of
heap-manipulating programs that marries (1) a new separation logic-based
analysis for heap reasoning with (2) an interpolation-based technique for
refining heap-shape invariants with data invariants. Splinter is property
directed, precise, and produces counterexample traces when a property does not
hold. Using the novel notion of spatial interpolants modulo theories, Splinter
can infer complex invariants over general recursive predicates, e.g., of the
form all elements in a linked list are even or a binary tree is sorted.
Furthermore, we treat interpolation as a black box, which gives us the freedom
to encode data manipulation in any suitable theory for a given program (e.g.,
bit vectors, arrays, or linear arithmetic), so that our technique immediately
benefits from any future advances in SMT solving and interpolation.
Body: \section{From Proofs of Paths to Proofs of Programs} In this section, we describe how \alg constructs a proof of correctness (i.e., unreachability of the error location $v_{\rm e}$) of a program $\prog = \tuple{V,E,v_{\rm i},v_{\rm e}}$ from proofs of individual paths. We note that our algorithm is an extension of \impact~ to \logic proofs; we refer the reader to~ for optimizations and heuristics. The main difference is the procedure \ppath, which constructs an \logic proof for a given program path. \paragraph{The Main Algorithm} Figure~ shows the main algorithm of \alg. Elements of the set $S$ are program paths from $v_{\rm i}$ to $v_{\rm e}$, annotated with $\logic$ formulas. For example, $$(a_1,v_1),(a_2,v_2),\ldots,(a_n,v_n)$$ is an annotated path where (1) $\{a_j\}_j$ are $\logic$ formulas; (2) $v_1 = v_{\rm i}$ and $v_n = v_{\rm e}$; (3) for $j \in [1,n-1]$, $(v_j,v_{j+1}) \in E$; (4) for each edge $e = (v_j,v_{j+1})$, $\{a_j\}\;e^{\rm c}\;\{a_{j+1}\}$ is valid; and (5) $a_n$ is $\false$ (since we are interested in proving unreachability of $v_{\rm e}$). \alg uses the procedure $\cproof$ to check if the set of annotated paths $S$ represents a proof of correctness of the whole program. If not, $\cproof$ samples a new program path ($\pi$) that does not appear in $S$. Using the procedure $\ppath$, it tries to construct an annotation/proof (using spatial($\theory$) interpolants) of $\pi$. If no proof is constructed, \alg concludes that $\pi$ is a feasible execution to $v_{\rm e}$ (or it performs unsafe memory operations). Otherwise, it uses the procedure $\conj$ to strengthen the proofs of all program paths in $S$, adds the annotated path to $S$, and restarts the loop. Note that the annotated paths in $S$ represent an \emph{Abstract Reachability Tree} (ART), as used in~ as well as other software model checking algorithms. The tree in this case is rooted at $v_{\rm i}$, and branching represents when two paths diverge. We will now describe \alg's components in more detail. \paragraph{\cproof: Checking Proofs} Given a set of annotated paths $S$, we use the procedure $\cproof(S)$ to check if the annotations represent a proof of the whole program. Specifically, for each $v \in V$, we use $I(v)$ to denote the formula $\bigvee\{a_j \mid (a_j,v) \in \kappa, \kappa \in S\}$. In other words, for each location $v$ in $\prog$, we \emph{hypothesize} an inductive invariant $I(v)$ from our current annotations of paths passing through $v$. We can then check if our hypotheses indeed form an inductive invariant of the program. If so, the program is safe: since $I(v_{\rm e})$ is always $\false$ (by definition), the fact that $I$ is an invariant implies that the post-state of any execution which reaches $v_{\rm e}$ must satisfy $\false$, and therefore the error location $v_{\rm e}$ is unreachable. Otherwise, $\cproof$ returns a new program path on which our hypothesized invariant does not hold, which we use to refine our hypotheses. In practice, one can perform $\cproof$ implicitly and lazily by maintaining a \emph{covering} (entailment) relation~ over the nodes of the ART. \paragraph{\ppath: Constructing a Proof of a Path} Figure~ shows an algorithm for computing a proof of a path $\pi$. First, in lines~-, we check if $\pi$ is feasible (i.e., whether $\pi$ corresponds to a real program execution). We do this by computing the strongest postconditions along $\pi$ (using $\post$) and then attempting to strengthen the annotation (using \refine) with theory interpolants to prove that \textit{false} is a postcondition of $\pi$. If no such strengthening is found, we know that $\pi$ is feasible (by Theorem~). Note that if $\pi$ is memory-infeasible (lines~-), then we only compute spatial interpolants along the memory-feasible prefix of the path and return the result. This is because when the path is memory-infeasible, we do not need data refinements along the path to prove it cannot reach $v_{\rm e}$. The function $\textsf{Spatial}((v_1,\ldots,v_n),P)$ takes a program path $\pi = v_1,\ldots,v_n$ and a \sep formula $P$ and returns the path annotated with spatial interpolants w.r.t the postcondition $P$. The function $\textsf{Refine}(\kappa,\varphi)$ takes an annotated program path $\kappa$ with $\sep$ formulas (from spatial interpolants) and a $\varphi \in \textsf{DFormula}$ and returns a refined annotation of $\kappa$ that proves the postcondition $\varphi$ (using theory interpolants). If the path $\pi$ is infeasible, we proceed by constructing spatial($\theory$) interpolants for it (lines~-). We use the function \textsf{Spatial} (Section~) to construct spatial interpolants, which we then refine with theory interpolants using the function \textsf{Refine} (Section~). Spatial path interpolants are computed with respect to the postcondition $\etrue : \true$, indicating that we are looking for a memory safety proof. Note that we might not be able to find theory interpolants if the spatial interpolants computed are too weak and \emph{hide important data elements} (in which case, on line~, we use the result of $\post$ as the spatial interpolants -- the strongest possible spatial interpolants). To illustrate how this could happen, consider Figure~: a modification to our illustrative example from Figure~. Here, a list of length 2 is constructed. The second loop checks that nodes at odd positions in the linked list have odd data elements. Suppose \alg samples the path that goes through the first loop twice and enters the second loop arriving at the assertion. Our spatial interpolation procedure will introduce a list segment $\ls(x,\nil)$ at location \texttt{P}. As a result, we cannot find a refinement that proves the assertion, since using the list segment predicate definition from Section~ we cannot specify that the first element is odd and the second is even. This is because refinements must apply to all elements of $\ls$, and cannot mention specific positions. In this case, we use the symbolic heaps as our spatial interpolants. That is, we annotate location \texttt{P} with $x\mapsto [d_1',e'] * e' \mapsto [d_2',\nil]$. Consequently, theory interpolants are able to refine this by specifying that $d_1'$ is odd. \paragraph{\conj: Conjoining Proofs of Paths} When a new annotated path $\kappa$ is computed, we strengthen the proofs of all annotated paths $\kappa'$ in $S$ that share a prefix with $\kappa$ using an operation $\conj(\kappa,\kappa')$, defined in the following. (This is analogous to strengthening annotations of a path in an ART -- all other paths sharing a prefix with the strengthened path also get strengthened.) Let $\kappa = (a_1,v_1),\ldots,(a_n,v_n)$ and $\kappa' = (a_1',v_1'),\ldots,(a_{m}',v_{m}')$. Let $k$ be the largest integer such that for all $j \leq k$, $v_j = v_j'$ (i.e., $k$ represents the longest shared prefix of $\kappa$ and $\kappa'$). $\conj$ returns a pair $(\ol{\kappa},\ol{\kappa}')$ consisting of the strengthened annotations: \begin{align*} \ol{\kappa} &\gets (a_1 \land a_1', v_1),\ldots,(a_k \land a_k', v_k), (a_{k+1},v_{k+1}),\ldots,(a_n,v_n)\\ \ol{\kappa}' &\gets (a_1 \land a_1', v_1),\ldots,(a_k \land a_k', v_k), (a_{k+1}',v_{k+1}'),\ldots,(a_{m}',v_{m}') \end{align*} The issue here is that $\logic$ is not closed under logical conjunction ($\land$), since we do not allow logical conjunction of spatial conjunctions, e.g., $(\ls(x,y)*\true) \land (\ls(y,z)*\true)$. In practice, we heuristically under-approximate the logical conjunction, using the strongest postcondition of the shared prefix to guide the under-approximation. Any under-approximation which over-approximates the strongest postcondition (including the strongest postcondition itself) is sound, but overly strong annotations may not generalize to a proof of the whole program. Note that the above transformation maintains the invariant that all paths in $S$ are annotated with valid Hoare triples. \section{Proofs} In this section, we present proof sketches for the theorems appearing in this paper. \subsection{Proof of Theorem~} \begin{theorem} Let $S$ and $I'$ be \logic formulas and let $c$ be a command such that $\post(c,S) \models I'$. Then \begin{enumerate} \item[(I)] $S \models \interp{S}{c}{I'}$ \item[(II)] $\hoare{\interp{S}{c}{I'}}{c}{I'}$ \end{enumerate} \end{theorem} \begin{proof} We prove one case to give intuition on why the theorem holds. Suppose \texttt{c} is an allocation statement \texttt{x := new($n$,$m$)}. Recall that we defined \[ \interp{S}{c}{I'} = \qex{x}{A} \] where \[\abduce{\post(c,S)}{\emptyset}{\qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}}{I'}\] Define \[ S' = \post(c,S) = S[x'/x] * x \mapsto [\vec{a},\vec{z}] \] for $x',\vec{a},\vec{z}$ fresh. First we show (I). By the properties of bounded abduction, we have \[ S' = S[x'/x] * x \mapsto [\vec{a},\vec{z}] \models A * (\qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}) \] from which we can see that \[ S[x'/x] \models A \models \qex{x}{A} \] and thus \[ S \models \qex{x}{A} = \interp{S}{c}{I'} \] Next we show (II). We compute \begin{align*} \post(c,\qex{x}{A}) &= \qex{x',\vec{a},\vec{z}}{(\qex{x}{A})[x'/x] * x \mapsto [\vec{a},\vec{z}]}\\ & \hspace*{10pt}\hfill \emph{where $x',\vec{a},\vec{z}$ are fresh.}\\ & \equiv \qex{x',\vec{a},\vec{z}}{(\qex{x}{A}) * x \mapsto [\vec{a},\vec{z}]}\\ & \equiv (\qex{x}{A}) * (\qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}) \end{align*} Since $S'$ is of the form $S[x'/x] * x \mapsto [\vec{a},\vec{z}]$ and $S' \models A * \true$, the only place where $x$ may appear in $A$ is in a disequality with some other allocated variable. It follows that \[ (\qex{x}{A}) * (\qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}) \equiv A * (\qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}) \] By the properties of bounded abduction, we have \[ A * (\qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}) \models I \] and thus \[ \post(c,\qex{x}{A}) \models I \] and finally \[ \hoare{\interp{S}{c}{I'}}{c}{I'} \] \end{proof} \subsection{Proof of Theorem~} \begin{theorem}[Soundness] Suppose that $\pi$ is a path and that $\zeta$ is a proof of the judgement $\cjudge{\hoare{\Pi:\Sigma}{$\pi$}{\Pi':\Sigma'}}{\mathcal{C}}$, and that $\sigma$ is a solution to $\mathcal{C}$. Then $\zeta^\sigma$, the proof obtained by applying the substitution $\sigma$ to $\zeta$, is a (refined) separation logic proof of \[ \hoare{(\Pi:\Sigma)^\sigma}{$\pi$}{(\Pi':\Sigma')^\sigma}\ .\] \end{theorem} \begin{proof} The proof proceeds by induction on $\zeta$. We will give an illustrative example using an entailment judgement. Suppose that $\zeta$ is an entailment proof consisting of a single application of the \textsc{Predicate} rule: \begin{mathpar} \inferrule*[lab=Predicate]{ \Pi \models \Pi' \\ }{ \Phi' \gets \Phi; \Psi_1' \gets \Psi_1 \land \Phi; \dotsi; \Psi_{|\vec{\tau}|}' \gets \Psi_{|\vec{\tau}|} \land \Phi \triangleright\\ \Pi \land \Phi : Z(\vec{\tau},\vec{E}) \vdash \Pi' \land \Phi' : Z(\vec{\tau'},\vec{E}) } \end{mathpar} (where $\tau_i = \lda{\vec{a}_i. \Psi_i}$ and $\tau_i' = \lda{\vec{a}_i. \Psi_i'}$). Suppose that $\sigma$ is a solution to the constraint system \[ \mathcal{C} = \Phi' \gets \Phi; \Psi_1' \gets \Psi_1 \land \Phi; \dotsi; \Psi_{|\vec{\tau}|}' \gets \Psi_{|\vec{\tau}|} \land \Phi \] Since $\sigma$ is a solution to $\mathcal{C}$, we have that \[ \Phi'^\sigma \Leftarrow \Phi^\sigma \text{ and for all $i$, } \Psi_i'^\sigma \Leftarrow \Psi_i^\sigma \land \Phi^\sigma \] and thus (noting that $\Pi \models \Pi'$) \[ \Pi \land \Phi^\sigma \models \Pi' \land \Phi'^\sigma \text{ and for all $i$, } \Psi_i^\sigma \land \Pi \land \Phi^\sigma \models \Psi_i'^\sigma\] It follows that $\zeta^\sigma$, given below, is a valid derivation: \begin{mathpar} \inferrule*[lab=Predicate]{ \Pi \land \Phi'^\sigma \models \Pi' \land \Phi'^\sigma \\ \Psi_1^\sigma \land \Pi \land \Phi^\sigma \models \Psi_1'\\ \dotsi\\ \Psi_n^\sigma \land \Pi \land \Phi^\sigma \models \Psi_n'^\sigma }{ \Pi \land \Phi^\sigma : Z(\vec{\tau}_0^\sigma,\vec{E}) \vdash \Pi' \land \Phi'^\sigma : Z(\vec{\tau}_1^\sigma,\vec{E}) } \end{mathpar} \end{proof} \subsection{Proof of Theorem~} \begin{theorem}[Completeness] Suppose that $\pi$ is a memory-safe path and $\zeta$ is the proof of the judgement \[\cjudge{\hoare{R_0(\vec{v}):\emp}{$\pi$}{R_1(\vec{v}) : \true}}{\mathcal{C}}\] obtained by symbolic execution. If $\phi$ is a data formula such that $\hoare{\textit{true}:\emp}{$\pi$}{\phi : \true}$ holds, then there is a solution $\sigma$ to $\mathcal{C}$ such that $R_1^\sigma(\vec{v}) \Rightarrow \phi$. \end{theorem} \begin{proof} Consider that for each formula $\qs{X}{\Pi}{\Sigma}$ in a symbolic execution sequence, $\Sigma$ is a *-conjunction of (finitely many) points-to predicates. The constraints we generate in this situation are the same as the ones that would be generated for a program which does not access the heap (but which has additional variables corresponding to data-typed heap fields). \end{proof} \section{Formalization of \logic and \sep} In this section we present the full proof systems for \logic and \sep, as well as the full set of constraint generation rules which we described in Section~. The syntax and semantics of \logic formulas, in terms of stacks and heaps, are shown in Figures~ and~. \begin{figure*}[t] \begin{minipage}[b]{0.45\linewidth} \textbf{\emph{Syntax}} \begin{align*} x,y \in \textsf{HVar}& \hspace*{60pt}\text{Heap variables} \\ a,b \in \dvar& \hspace*{60pt}\text{Data variables} \\ A \in \textsf{DTerm} & \hspace*{60pt} \text{First-order term that evaluates to value in $\mathds{D}$}\\ \phi \in \textsf{DFormula} & \hspace*{60pt} \text{Data formulas}\\ Z \in \pred & \hspace*{60pt} \text{Recursive predicates}\\ \theta \in \textsf{Refinement} &::= \lambda \vec{a}. \phi\\ X \subseteq \textsf{Var}&::= x \mid a\\ E,F \in \textsf{HTerm}&::= \nil \mid x\\ \AE &::= A \ \mid E\\ \Pi \in \textsf{Pure}&::= \etrue \mid E = E \mid E \neq E \mid \varphi \ \mid \Pi \land \Pi\\ \Atom \in \textsf{Heaplet}&::= \true \mid \emp \mid E \mapsto [\vec{A},\vec{E}] \mid Z(\vec{\theta},\vec{E})\\ \Sigma \in \textsf{Spatial}&::= \Atom \mid \Atom * \Sigma\\ P \in \logic&::= \qs{X}{\Pi}{\Sigma} \end{align*} \end{minipage} \caption{Syntax of \logic formulas.} \end{figure*} \begin{figure*}[t] \begin{minipage}[b]{0.45\linewidth} \textbf{\emph{Semantic Domains}} \begin{align*} \textsf{Var} &= \textsf{HVar} + \dvar\\ \textsf{Val} &= \textsf{Loc} + \domain\\ \textsf{Stack} &= \textsf{Var} \rightarrow \textsf{Val}\\ \textsf{Heap} &= \textsf{Loc} \rightharpoonup_{\textsf{fin}} \rec\\ \rec &= \tuple{\mathds{N} \rightharpoonup_{\textsf{fin}} \domain, \mathds{N} \rightharpoonup_{\textsf{fin}} \textsf{Loc}}\\ \textsf{State} &= \textsf{Stack} \times \textsf{Heap} \end{align*} \end{minipage} \begin{minipage}[b]{0.45\linewidth} ~\\ \textbf{\emph{Satisfaction Semantics}} \begin{align*} s,h \models E = F &\iff \sem{E}(s) = \sem{F}(s)\\ s,h \models E \neq F &\iff \sem{E}(s) \neq \sem{F}(s)\\ s,h \models \varphi &\iff \sem{\varphi}(s)\\ s,h \models \Pi_1 \land \Pi_2 &\iff (s,h \models \Pi_1) \text{ and } (s,h \models \Pi_2)\\ s,h \models Z(\vec{\tau},\vec{E}) &\iff \exists P \in \mathit{cases}(Z(\vec{R},\vec{x})). s,h \models P[\vec{\tau}/\vec{R},\vec{E}/\vec{x}]\\ s,h \models \emp &\iff \dom(h) = \emptyset\\ s,h \models E \mapsto [\vec{A}, \vec{F}] & \iff \dom(h) = \{\sem{E}(s)\} \\ &\hspace*{1cm}\text{and } h(\sem{E}(s)) = &\\ &\hspace*{1.25cm} \tuple{\{i \mapsto \sem{A_i}(s) | i \in [1,|\vec{A}|]\}, \{i \mapsto \sem{F_i}(s) | i \in [1,|\vec{F}|]\}}\\ s,h \models \Sigma * \Sigma' &\iff \text{there exists } h_0, h_1 \text{ s.t. } h_0 \uplus h_1 = h \\ &\hspace*{1cm} \text{and } (s,h_0 \models \Sigma) \text{ and } (s,h_1 \models \Sigma')\\ s,h \models \Pi \colon \Sigma &\iff (s,h \models \Pi) \text{ and } (s,h \models \Sigma)\\ s,h \models \qs{X}{\Pi}{\Sigma} &\iff \text{there exists } \overline{s} : X \rightarrow \textsf{Val} \text{ s.t. } s\oplus \overline{s},h \models \Pi:\Sigma \end{align*} \end{minipage} ~\\ Note that we model records (\rec) as two finite maps representing data fields and heap fields.\\ We use $\uplus$ to denote union of functions with disjoint domains, and $\oplus$ to denote overriding union of functions. \caption{Stack/heap semantics of \logic formulas.} \end{figure*} \begin{figure*}[t] \footnotesize \figsep{Entailment rules} \begin{mathpar} \inferrule*[lab=Empty]{\Pi \models \Pi'}{ \Pi : \emp \vdash \Pi' : \emp} \inferrule[$\exists$-left]{ P[x'/x] \vdash Q }{ \qex{x}{P} \vdash Q } \inferrule[$\exists$-right]{ P \vdash Q[\mbox{\AE}/x] }{ P \vdash \qex{x}{Q} } \inferrule*[lab=Predicate,right={\rm\begin{minipage}{2.85cm} Where $\tau_i = \lda{\vec{a}_i. \phi_i}$\\ and $\tau_i' = \lda{\vec{a}_i. \phi_i'}$ \end{minipage}}]{ \Pi \models \Pi' \\ \phi_1 \land \Pi \models \phi_1'\\ \dotsi\\ \phi_n \land \Pi \models \phi_n' }{ \Pi : Z(\vec{\tau},\vec{E}) \vdash \Pi' : Z(\vec{\tau}',\vec{E}) } \inferrule*[lab=True]{\Pi \models \Pi'}{ \Pi : \Sigma \vdash \Pi' : \true } \inferrule[Points-to]{\Pi \models \Pi'}{ \Pi : E \mapsto [\vec{A},\vec{F}] \vdash \Pi' : E \mapsto [\vec{A},\vec{F}] } \inferrule[Star]{ \Pi : \Sigma_0 \vdash \Pi' : \Sigma_0' \\ \Pi : \Sigma_1 \vdash \Pi' : \Sigma_1' }{ \Pi : \Sigma_0 * \Sigma_1 \vdash \Pi' : \Sigma_0' * \Sigma_1' } \inferrule[Substitution]{ \Pi[E/x] : \Sigma[E/x] \vdash \Pi'[E/x] : \Sigma'[E/x] \\ \Pi \models x = E }{ \Pi : \Sigma \vdash \Pi' : \Sigma' } \inferrule[\nil-not-Lval]{ \Pi \land E \neq \nil : \Sigma * E \mapsto [\vec{A},\vec{F}] \vdash \Pi' : \Sigma' }{ \Pi : \Sigma * E \mapsto [\vec{A},\vec{F}] \vdash \Pi' : \Sigma' } \inferrule[*-Partial]{ \Pi \land E \neq F : \Sigma * E \mapsto [\vec{A},\vec{E}] * F \mapsto [\vec{B},\vec{F}] \vdash \Pi' : \Sigma' }{ \Pi : \Sigma * E \mapsto [\vec{A},\vec{E}] * F \mapsto [\vec{B},\vec{F}] \vdash \Pi' : \Sigma' } \inferrule*[lab=Fold,right={\rm $P \in \mathit{cases}(Z(\vec{R},\vec{x}))$}]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * P[\vec{\tau}/\vec{R},\vec{E}/\vec{x}] }{ \Pi : \Sigma \vdash \Pi' : \Sigma' * Z(\vec{\tau},\vec{E}) } \inferrule*[lab=Unfold,right={\rm $\{P_1,\dotsc,P_n\} = \mathit{cases}(Z(\vec{R},\vec{x}))$}]{ \Pi : \Sigma * P_1[\vec{\tau}/\vec{R},\vec{E}/\vec{x}] \vdash \Pi' : \Sigma'\\ \dotsi\\ \Pi : \Sigma * P_n[\vec{\tau}/\vec{R},\vec{E}/\vec{x}] \vdash \Pi' : \Sigma' }{ \Pi : \Sigma * Z(\vec{\tau},\vec{E}) \vdash \Pi' : \Sigma' } \end{mathpar} \caption{\logic Proof System} \end{figure*} \begin{figure*}[t] \figsep{Execution rules} \begin{mathpar} \inferrule[Assign]{ }{ \hoare{\Pi \colon \Sigma} {x := \AE} {\qex{x'}{\Pi[x'/x] \land x = \mbox{\AE}[x'/x] : \Sigma[x'/x]}} } \inferrule[Assume]{ }{ \hoare{\Pi \colon \Sigma}{assume($\Pi'$)}{\Pi \land \Pi' : \Sigma} } \inferrule[Sequence]{ \hoare{P}{$\pi_0$}{O}\\ \hoare{O}{$\pi_1$}{Q} }{ \hoare{P}{$\pi_0;\pi_1$}{Q} } \inferrule[Data-Store]{ \Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}] }{ \hoare{\Pi \colon \Sigma} {x->D$_i$ := A} {\Pi' : \Sigma' * x \mapsto [\vec{A}[A/A_i],\vec{E}]} } \inferrule[Heap-Store]{ \Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}] }{ \hoare{\Pi \colon \Sigma}{x->N$_i$ := E}{\Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}[E/E_i]]} } \inferrule[Data-Load]{ \Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}] }{ \hoare{\Pi \colon \Sigma} {y := x->D$_i$} {\qs{y'} {\Pi'[y'/y] \land y = A_i[y'/y]} { (\Sigma' * x \mapsto [\vec{A},\vec{E}])[y'/y]}} } \inferrule[Free]{ \Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}] }{ \hoare{\Pi \colon \Sigma}{free(x)}{\Pi' : \Sigma'} } \inferrule[Heap-Load]{ \Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}] }{ \hoare{\Pi \colon \Sigma} {y := x->N$_i$} {\qs{y'} {\Pi'[y'/y] \land y = E_i[y'/y]} { (\Sigma' * x \mapsto [\vec{A},\vec{E}])[y'/y]}} } \inferrule[Consequence]{ P' \vdash P\\ \hoare{P}{c}{Q}\\ Q \vdash Q' }{ \hoare{P'}{c}{Q'} } \inferrule[Alloc]{ }{ \hoare{\Pi \colon \Sigma} {x := new(n,m)} {\qs{x',\vec{a},\vec{y}}{\Pi[x'/x] }{ \Sigma[x'/x] * x \mapsto [\vec{a},\vec{y}]}} } \inferrule*[lab=Exists,right={\rm $x \notin \Vars(Q) \cup \Vars(c)$}]{ \hoare{P}{c}{Q} }{ \hoare{\qex{x}{P}}{c}{Q} } \end{mathpar} \caption{\logic proof system.} \end{figure*} \begin{figure*}[t] \begin{mathpar} \small \inferrule*[lab=Predicate]{ \Pi \models \Pi' }{ \Pi : Z(\vec{E}) \vdash \Pi' : Z(\vec{E}) } \inferrule*[lab=Fold,right={\rm $P \in \mathit{cases}(Z(\vec{R},\vec{x}))$}]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * P[\vec{E}/\vec{x}] }{ \Pi : \Sigma \vdash \Pi' : \Sigma' * Z(\vec{E}) } \inferrule*[lab=Unfold,right={\rm $\{P_1,\dotsc,P_n\} = \mathit{cases}(Z(\vec{R},\vec{x}))$}]{ \Pi : \Sigma * P_1[\vec{E}/\vec{x}] \vdash \Pi' : \Sigma'\\\\ \dotsi\\\\ \Pi : \Sigma * P_n[\vec{E}/\vec{x}] \vdash \Pi' : \Sigma' }{ \Pi : \Sigma * Z(\vec{E}) \vdash \Pi' : \Sigma' } \end{mathpar} \caption{\sep proof system. All other entailment and execution rules are as in Figure~.} \end{figure*} \begin{figure*}[t] \figsep{Entailment rules} \vspace{-.15in} \begin{mathpar} \footnotesize \inferrule[Empty]{ \Pi \models \Pi' }{ \cjudge{\Pi \land \Phi : \emp \vdash \Pi' \land \Phi' : \emp} {\Phi' \gets \Phi} } \inferrule[True]{\Pi \models \Pi'}{ \cjudge{\Pi \land \Phi : \Sigma \vdash \Pi' \land \Phi' : \true} {\Phi' \gets \Phi} } \inferrule[Inconsistent]{ \Pi \models \textit{false} }{ \cjudge{\Pi \land \Phi : \Sigma \vdash \Pi' \land \Phi' : \Sigma'} { [] } } \inferrule[$\exists$-left]{ \cjudge{P[x'/x] \vdash Q}{\mathcal{C}} }{ \cjudge{\qex{x}{P} \vdash Q}{\mathcal{C}} } \inferrule[Substitution]{ \cjudge{\Pi[E/x] \land \Phi : \Sigma[E/x] \vdash \Pi'[E/x] \land \Phi' : \Sigma'[E/x]} {\mathcal{C}}\\ \Pi \models x = E }{ \cjudge{\Pi \land \Phi : \Sigma \vdash \Pi' \land \Phi' : \Sigma'} {\mathcal{C}} } \inferrule[\nil-not-Lval]{ \cjudge{\Pi \land \Phi \land E \neq \nil : \Sigma * E \mapsto [\vec{A},\vec{F}] \vdash \Pi' \land \Phi' : \Sigma'} {\mathcal{C}} }{ \cjudge{\Pi \land \Phi : \Sigma * E \mapsto [\vec{A},\vec{F}] \vdash \Pi' \land \Phi' : \Sigma'} {\mathcal{C}} } \inferrule[$\exists$-right]{ \cjudge{P \vdash Q[\mbox{\AE}/x]}{\mathcal{C}} }{ \cjudge{P \vdash \qex{x}{Q}}{\mathcal{C}} } \inferrule[*-Partial]{ \cjudge{\Pi \land E \neq F \land \Phi : \Sigma * E \mapsto [\vec{A},\vec{E}] * F \mapsto [\vec{B},\vec{F}] \vdash \Pi' \land \Phi' : \Sigma'} {\mathcal{C}} }{ \cjudge{\Pi \land \Phi : \Sigma * E \mapsto [\vec{A},\vec{E}] * F \mapsto [\vec{B},\vec{F}] \vdash \Pi' \land \Phi' : \Sigma'} {\mathcal{C}} } \inferrule[Star]{ \cjudge{\Pi \land \Phi : \Sigma_0 \vdash \Pi' \land \Phi' : \Sigma_0'} { \mathcal{C}_0} \\ \cjudge{\Pi \land \Phi : \Sigma_1 \vdash \Pi' \land \Phi' : \Sigma_1'} { \mathcal{C}_1 } }{ \cjudge{\Pi \land \Phi : \Sigma_0 * \Sigma_1 \vdash \Pi' \land \Phi' : \Sigma_0' * \Sigma_1'} {\mathcal{C}_0;\mathcal{C}_1} } \inferrule[Points-to]{ \Pi \models \Pi' } { \cjudge{\Pi \land \Phi : E \mapsto [\vec{A},\vec{F}] \vdash \Pi' \land \Phi' : E \mapsto [\vec{A},\vec{F}]} { \Phi' \gets \Phi } } \inferrule*[lab=Fold,right={\rm $P \in \mathit{cases}(Z(\vec{R},\vec{x}))$}]{ \cjudge{\Pi : \Sigma \vdash \Pi' : \Sigma' * P[\vec{\tau}/\vec{R},\vec{E}/\vec{x}]} {\mathcal{C}} }{ \cjudge{\Pi : \Sigma \vdash \Pi' : \Sigma' * Z(\vec{\tau},\vec{E}) } {\mathcal{C}} } \inferrule*[lab=Unfold,right={\rm $\{P_1,\dotsc,P_n\} = \mathit{cases}(Z(\vec{R},\vec{x}))$}]{ \cjudge{\Pi : \Sigma * P_1[\vec{\tau}/\vec{R},\vec{E}/\vec{x}] \vdash \Pi' : \Sigma' } {\mathcal{C}_1}\\ \dotsi\\ \cjudge{\Pi : \Sigma * P_n[\vec{\tau}/\vec{R},\vec{E}/\vec{x}] \vdash \Pi' : \Sigma'} {\mathcal{C}_n} }{ \cjudge{\Pi : \Sigma * Z(\vec{\tau},\vec{E}) \vdash \Pi' : \Sigma'} {\mathcal{C}_1; \dotsc{;}\, \mathcal{C}_n} } \inferrule*[lab=Predicate,right={\rm\begin{minipage}{2.85cm} Where $\tau_i = \lda{\vec{a}_i. \Psi_i}$\\ and $\tau_i' = \lda{\vec{a}_i. \Psi_i'}$ \end{minipage}}]{ \Pi \models \Pi' \\ }{ \Phi' \gets \Phi; \Psi_1' \gets \Psi_1 \land \Phi; \dotsc{;}\, \Psi_{|\vec{\tau}|}' \gets \Psi_{|\vec{\tau}|}' \land \Phi \triangleright \Pi \land \Phi : Z(\vec{\tau},\vec{E}) \vdash \Pi' \land \Phi' : Z(\vec{\tau'},\vec{E}) } \end{mathpar} \caption{Constraint Generation: Entailment Rules} \end{figure*} \begin{figure*}[t] \figsep{Execution rules} \vspace{-.15in} \begin{mathpar} \footnotesize \inferrule[Data-Assume]{ \cjudge{P \land \phi \vdash Q} {\mathcal{C}} }{ \cjudge{\hoare{P} {assume($\phi$)} {Q}} {\mathcal{C}} } \inferrule[Free]{ \cjudge{P \vdash \Pi \land \Phi : \Sigma * x \mapsto [\vec{A},\vec{E}]} {\mathcal{C}} }{ \cjudge{\hoare{P}{free(x)}{\Pi \land \Phi : \Sigma}} {\mathcal{C}} } \inferrule[Sequence]{ \cjudge{\hoare{P} {$\pi_0$} {\widehat{O}}} {\mathcal{C}_0}\\ \cjudge{\hoare{\widehat{O}} {$\pi_1$} {Q}} {\mathcal{C}_1} }{ \cjudge{\hoare{P} {$\pi_0;\pi_1$} {Q}} {\mathcal{C}_0;\mathcal{C}_1} } \inferrule[Data-Load]{ \cjudge{P \vdash \qex{X}{\Pi \land \widehat{\Phi} : \widehat{\Sigma} * x \mapsto [\vec{A},\vec{E}]}} {\mathcal{C}_0}\\\\ \cjudge{\qex{X,a'}{\Pi[a'/a] \land \widehat{\Phi}[a'/a] \land a = A_i[a'/a] \colon (\widehat{\Sigma} * x \mapsto [\vec{A},\vec{E}])[a'/a]} \vdash Q} {\mathcal{C}_1} }{ \cjudge{\hoare{P} {a := x->D$_i$} {Q}} {\mathcal{C}_0; \mathcal{C}_1 } } \inferrule[Data-Assign]{ \cjudge{\qex{a'}{\Pi \land \Phi[a'/a] \land a = A[a'/a] : \Sigma[a'/a] \vdash Q}} {\mathcal{C}} }{ \cjudge{\hoare{\Pi \land \Phi \colon \Sigma}{a := A}{Q}} { \mathcal{C} } } \inferrule[Data-Store]{ \cjudge{P \vdash \qex{X}{\Pi \land \widehat{\Phi} : \widehat{\Sigma} * x \mapsto [\vec{A},\vec{E}]}} {\mathcal{C}_0}\\\\ \cjudge{\qex{X,a'}{\Pi \land \widehat{\Phi} \land a' = A : \widehat{\Sigma} * x \mapsto [\vec{A}[a'/A_i],\vec{E}]} \vdash Q} {\mathcal{C}_1} }{ \cjudge{\hoare{P}{x->D$_i$ := A}{Q}} {\mathcal{C}_0; \mathcal{C}_1} } \inferrule[Alloc]{ \cjudge{\qex{x',\vec{a},\vec{x}}{\Pi[x'/x] \land \Phi : \Sigma[x'/x] * x \mapsto [\vec{a},\vec{x}]} \vdash Q} {\mathcal{C}} }{ \cjudge{\hoare{\Pi \land \Phi \colon \Sigma} {x := new($n$,$m$)} {Q}} {\mathcal{C}} } \inferrule[Heap-Store]{ \cjudge{\Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}]} {\mathcal{C}} }{ \cjudge{\hoare{\Pi \colon \Sigma}{x->N$_i$ := E}{\Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}[E/E_i]]}} {\mathcal{C}} } \inferrule[Consequence]{ \cjudge{P' \vdash \widehat{P}}{\mathcal{C}_1}\\ \cjudge{\hoare{\widehat{P}}{c}{\widehat{Q}}}{\mathcal{C}_2}\\ \cjudge{\widehat{Q} \vdash Q'}{\mathcal{C}_3} }{ \cjudge{\hoare{P'}{c}{Q'}}{\mathcal{C}_1;\mathcal{C}_2;\mathcal{C}_3} } \inferrule[Heap-Load]{ \cjudge{\Pi \colon \Sigma \vdash \Pi' : \Sigma' * x \mapsto [\vec{A},\vec{E}]} {\mathcal{C}} }{ \cjudge{\hoare{\Pi \colon \Sigma} {y := x->N$_i$} {\qs{y'} {\Pi'[y'/y] \land y = E_i[y'/y]} { (\Sigma' * x \mapsto [\vec{A},\vec{E}])[y'/y]}}} {\mathcal{C}} } \end{mathpar} \caption{Constraint Generation: Execution Rules} \end{figure*} \section{Spatial interpolation for assumptions} In the spatial interpolation rules for \texttt{assume} presented in Section~, we encountered the following problem: we have an equality or disequality assertion $\Pi$, a symbolic heap $S$, and a formula $M$ such that $S \land \Pi \vdash M$, and we need to compute a formula $M'$ such that $S \vdash M'$ and $M' \land \Pi \vdash M$. Moreover, we wish $M'$ to be as weak as possible (i.e., $M'$ should be ``close to $M$'' rather than ``close to $S$''). In this section, we will define a recursive procedure $\textsf{pitp}(S,\Pi,M)$ which takes as input a symbolic heap $S$, an equality or disequality formula $\Pi$, and a $\sep$ formula $M$ such that $S \land \Pi \vdash M$ and computes a formula $M' = \textsf{pitp}(S,\Pi,M)$ such that $S \vdash M'$ and $M' \land \Pi \vdash M$. We will assume that $S = \Pi_S : \Sigma_S$ is saturated in the sense that for any assertion $\Pi_0$, if $S \vdash \Pi_0$ then $\Pi_S \vdash \Pi_0$, and that $S \land \Pi$ is satisfiable. We observe that if $M$ has existentially quantified variables, they can be instantiated using the proof of $S \land \Pi \vdash M$. Thus we may assume that $M$ is quantifier-free, and write $M$ as \[ M = \Pi_M : H_1 * \dotsi * H_n \] The proof of $S \land \Pi \vdash M$ also induces an $n$-colouring on $S$ which colours each points-to predicate in $S$ with the index $i$ of the corresponding heaplet $H_i$ (cf. step 1 of the bounded abduction procedure presented in Section~). We may thus write $S$ as follows: \[ S = \Pi_S : \coloured{\Sigma_1}{1} * \dotsi * \coloured{\Sigma_n}{n} \] (such that for each $i$, $\Pi_S \land \Pi : \Sigma_i \vdash \Pi_M : H_i $). We will compute a pure formula $\Pi_M'$ such that $\Pi_S \vdash \Pi_M'$ and $\Pi_M' \land \Pi \vdash \Pi_M$ and for each $i$ we will compute a formula $P_i$ such that $\Pi_S : \Sigma_i \vdash P_i$ and $P_i \land \Pi \vdash H_i$. We then take $\textsf{pitp}(S,\Pi,M)$ to be $\Pi_M' : P_1 * \dotsi * P_n$. First, we show how to compute $\Pi_M'$. Note that note that since $S$ is saturated, the fact that $S \land \Pi \vdash M$ implies $\Pi_S \land \Pi \vdash \Pi_M$. We will assume that $\Pi_M$ consists of a single equality or disequality: the procedure can be extended to arbitrary conjunction by applying it separately for each conjunct and conjoining the results. If $\Pi_S \vdash \Pi_M$ then we simply take $\Pi_M'$ to be $\Pi_M$. Otherwise, assume $w,x,y,z$ are such that $\Pi$ is an equality or disequality $w = x$ / $w \neq x$ and $\Pi_M$ is an equality or disequality $y = z$ / $y \neq z$. Since $\Pi_S \land \Pi \vdash \Pi_M$ and $\Pi_S \not\vdash \Pi_M$, there is some $y',z' \in \{w,x\}$ such that $\Pi_S \vdash y = y' \land z = z'$ (to see why, consider each of the four cases for $\Pi_S$ and $\Pi_M$). We take $\Pi_M'$ to be $y = y' \land z = z'$. Finally, we show how to compute $P_i$ (for each $i \in [1,n]$). If $\Pi_S : \Sigma_i \vdash H_i$, then we simply take $P_i$ to be $H_i$. Otherwise, suppose that $H_i$ is $Z(\vec{E})$ for some predicate $Z$ and vector of heap terms $\vec{E}$ (the case that $H_i$ is a points-to predicate is essentially a special case). First, we attempt to find a vector of heap terms $\vec{E'}$ such that $\Pi_S : \Sigma_i \vdash Z(\vec{E'})$ and $\Pi_S \land \Pi \vdash E_j = E_j'$ for each $j$ (noting that there are finitely many such $\vec{E'}$ to choose from). If we succeed, we may take $P_i$ to be $\Pi' : Z(\vec{E'})$, where $\Pi' = \textsf{pitp}(S, \Pi, \bigwedge_j E_j = E_j')$. If we fail, then since $\Pi_S \land \Pi : \Sigma_i \vdash Z(\vec{E})$, there is some $Q \in \mathit{cases}(Z(\vec{R},\vec{x}))$ such that $\Pi_S \land \Pi : \Sigma_i \vdash \ul{Q}[\vec{E}/\vec{x}]$. We may take $P_i$ to be $\textsf{pitp}(\Pi_S : \Sigma_i, \Pi, Q[\vec{E}/\vec{x}])$. \section{Bounded Abduction} \newcommand{\coloured}[2]{ \ifthenelse{\equal{#2}{r}}{ {\color{BrickRed}{\ensuremath[#1]^{\rm #2}}} }{\ifthenelse{\equal{#2}{b}}{ {\color{RoyalBlue}\ensuremath[#1]^{\rm #2}} }{ \ensuremath[#1]^{#2} } } } \newcommand{\hiding}[2]{\ensuremath\langle#1 \unlhd #2\rangle} \begin{figure*}[t] \setlength{\abovecaptionskip}{-4pt} \setlength{\belowcaptionskip}{4pt} \centering \begin{mathpar} \scriptsize \inferrule[Empty]{\Pi \models \Pi'}{ \Pi : \coloured{\emp}{c} \vdash \Pi' : \hiding{\coloured{\emp}{c}}{\emp} } \inferrule[Star]{ \Pi : \Sigma_0 \vdash \Pi' : \Sigma_0' \\ \Pi : \Sigma_1 \vdash \Pi' : \Sigma_1' }{ \Pi : \Sigma_0 * \Sigma_1 \vdash \Pi' : \Sigma_0' * \Sigma_1' } \inferrule[Points-to]{\Pi \models \Pi'}{ \Pi : \coloured{E \mapsto [a,F]}{c} \vdash \Pi' : \hiding{\coloured{E \mapsto [a, F]}{c}}{E \mapsto [a, F]} } \inferrule[True]{\Pi \models \Pi'}{ \Pi : \Sigma \vdash \Pi' : \hiding{\coloured{\true}{c}}{\true} } \inferrule[Substitution]{ \Pi[E/x] : \Sigma[E/x] \vdash \Pi'[E/x] : \Sigma'[E/x] \\ \Pi \models x = E }{ \Pi : \Sigma \vdash \Pi' : \Sigma' } \inferrule[$\exists$-right]{ P \vdash Q[\AE/x] }{ P \vdash \qex{x}{Q} } \end{mathpar} \caption{Coloured strengthening. All primed variables are chosen fresh.} \end{figure*} In this section, we discuss our algorithm for bounded abduction. Given a bounded abduction problem \[L \vdash \qex{X}{M * [\ ]} \vdash R\] we would like to find a formula $A$ such that $L \vdash \qex{X}{M * A} \vdash R$. Our algorithm is sound but not complete: it is possible that there exists a solution to the bounded abduction problem, but our procedure cannot find it. In fact, there is in general no complete procedure for bounded abduction, as a consequence of the fact that we do not pre-suppose that our proof system for entailment is complete, or even that entailment is decidable. \paragraph{High level description} Our algorithm proceeds in three steps: \begin{compactenum} \item Find a \emph{colouring} of $L$. This is an assignment of a colour, either \emph{red} or \emph{blue}, to each heaplet appearing in $L$. Intuitively, red heaplets are used to satisfy $M$, and blue heaplets are left over. This colouring can be computed by recursion on a proof of $L \vdash \qex{X}{M * \true}$. \item Find a \emph{coloured strengthening} $\Pi : \coloured{M'}{r} * \coloured{A}{b}$ of $R$. (We use the notation $\coloured{\Sigma}{r}$ or $\coloured{\Sigma}{b}$ to denote a spatial formula $\Sigma$ of red or blue colour, respectively.) Intuitively, this is a formula that (1) entails $R$ and (2) is coloured in such a way that the red heaplets correspond to the red heaplets of $L$, and the blue heaplets correspond to the blue heaplets of $L$. This coloured strengthening can be computed by recursion on a proof of $L \vdash R$ using the colouring of $L$ computed in step 1. \item Check $\Pi' : M * A \models R$, where $\Pi'$ is the strongest pure formula implied by $L$. This step is necessary because $M$ may be weaker than $M'$. If the entailment check fails, then our algorithm fails to compute a solution to the bounded abduction problem. If the entailment check succeeds, then $\Pi'' : A$ is a solution, where $\Pi''$ is the set of all equalities and disequalities in $\Pi'$ which were actually used in the proof of the entailment $\Pi' : M * A \models R$ (roughly, all those equalities and disequalities which appear in the leaves of the proof tree, plus the equalities that were used in some instance of the {\sc Substitution} rule). \shortenpar \end{compactenum} First, we give an example to illustrate these high-level steps: \begin{example} Suppose we want to solve the following bounded abduction problem: \[ L \vdash \ls(x,y) * [\ ] \vdash R \] where $L = x \mapsto [a,y] * y \mapsto [b,\nil]$ and $R = \qex{z}{x\mapsto [a,z] * \ls(y,\nil)}$. Our algorithm operates as follows: \begin{compactenum} \item Colour $L$: $\coloured{x \mapsto [a,y]}{r} * \coloured{y \mapsto [b,\nil]}{b}$ \item Colour $R$: $\qex{z}{\coloured{x \mapsto [a,z]}{r} * \coloured{\ls(y,\nil)}{b}}$ \item Prove the entailment \[ x \neq \nil \land y \neq \nil \land x \neq y : \ls(x,y) * \ls(y,\nil) \models R\] This proof succeeds, and uses the pure assertion $x \neq y$. \end{compactenum} Our algorithm computes $x \neq y : \ls(y,\nil)$ as the solution to the bounded abduction problem. \eoe\end{example} We now elaborate our bounded abduction algorithm. We assume that $L$ is quantifier free (without loss of generality, since quantified variables can be Skolemized) and \emph{saturated} in the sense that for any pure formula $\Pi'$, if $L \vdash \Pi'$, where $L = \Pi : \Sigma$, then $\Pi \vdash \Pi'$. \paragraph{Step 1} The first step of the algorithm is straightforward. If we suppose that there exists a solution, $A$, to the bounded abduction problem, then by definition we must that have $L \models \qex{X}{M * A}$. Since $\qex{X}{M * A} \models \qex{X}{M * \true}$, we must also have $L \models \qex{X}{M * \true}$. We begin step 1 by computing a proof of $L \vdash \qex{X}{M * \true}$. If we fail, then we abort the procedure and report that we cannot find a solution to the abduction problem. If we succeed, then we can colour the heaplets of $L$ as follows: for each heaplet $E \mapsto [\vec{A},\vec{F}]$ in $L$, either $E \mapsto [\vec{A},\vec{F}]$ was used in an application of the {\sc Points-to} axiom in the proof of $L \vdash \qex{X}{M * \true}$ or not. If yes, we colour $E \mapsto [\vec{A},\vec{F}]$ red; otherwise, we colour it blue. We denote a heaplet $H$ coloured by a colour $c$ by $\coloured{H}{c}$. \shortenpar \paragraph{Step 2} The second step is to find a coloured strengthening of $R$. Again, supposing that there is some solution $A$ to the bounded abduction problem, we must have $L \models \qex{X}{M * A} \models R$, and therefore $L \models R$. We begin step 2 by computing a proof of $L \vdash R$. If we fail, then we abort. If we succeed, then we define a coloured strengthening of $R$ by recursion on the proof of $L \vdash R$. Intuitively, this algorithm operates by inducing a colouring on points-to predicates in the leaves of the proof tree from the colouring of $L$ (via the {\sc Points-to} rule in Fig.~) and then only folding recursive predicates when all the folded heaplets have the same colour. \shortenpar More formally, for each formula $P$ appearing as the consequent of some sequent in a proof tree, our algorithm produces a mapping from heaplets in $P$ to coloured spatial formulas. The mapping is represented using the notation $\hiding{\Sigma}{H}$, which denotes that the heaplet $H$ is mapped to the coloured spatial formula $\Sigma$. For each recursive predicate $Z$ and each $\qex{X}{\Pi : H_1 * \dotsi * H_n} \in \mathit{cases}(Z(\vec{R},\vec{x}))$, we define two versions of the fold rule, corresponding to when $H_1,\ldots,H_n$ are coloured homogeneously ({\sc Fold1}) and heterogeneously ({\sc Fold2}): \begin{mathpar} \scriptsize \inferrule[Fold1]{ (\Pi : \Sigma \vdash \Pi' : \Sigma' * \hiding{\coloured{H_1}{c}}{H_1} * \dotsi * \hiding{\coloured{H_n}{c}}{H_n})[\vec{E}/\vec{x}] }{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \hiding{\coloured{Z(\vec{E})}{c}}{Z(\vec{E})} } \inferrule[Fold2]{ (\Pi : \Sigma \vdash \Pi' : \Sigma' * \hiding{\Sigma_1'}{H_1} * \dotsi * \hiding{\Sigma_n'}{H_n})[\vec{E}/\vec{x}] }{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \hiding{\Sigma_1'* \dotsi * \Sigma_n'}{Z(\vec{E})} } \end{mathpar} The remaining rules for our algorithm are presented formally in Fig.~.\footnote{Note that some of the inference rules are missing. This is because these rules are inapplicable (in the case of {\sc Unfold} and {\sc Inconsistent}) or unnecessary (in the case of {\sc \nil-not-Lval} and {\sc*-Partial}), given our assumptions on the antecedent.} To illustrate how this algorithm works, consider the {\sc Fold1} and {\sc Fold2} rules. If a given (sub-)proof finishes with an instance of {\sc Fold} that folds $H_1 *\cdots* H_n$ into $Z(\vec{E})$, we begin by colouring the sub-proof of \[\Pi : \Sigma \vdash \Pi' : \Sigma' * H_1 * \dotsi * H_n \] This colouring process produces a coloured heaplet $\Sigma_i$ for each $H_i$. If there is some colour $c$ such that each $\Sigma_i'$ is $\coloured{H_i}{c}$, then we apply {\sc Fold1} and $Z(\vec{E})$ gets mapped to $\coloured{Z(\vec{E})}{c}$. Otherwise (if there is some $i$ such that $\Sigma_i$ is not $H_i$ or there is some $i,j$ such that $\Sigma_i$ and $\Sigma_j$ have different colours), we apply {\sc Fold2}, and map $Z(\vec{E})$ to $\Sigma_1* \dotsi *\Sigma_n$. After colouring a proof, we define $A$ to be the blue part of $R$. That is, if the colouring process ends with a judgement of\\ $\Pi : \coloured{\Sigma_1}{r} * \coloured{\Sigma_2}{b} \vdash \Pi' : \hiding{\coloured{\Sigma_{11}'}{r}*\coloured{\Sigma_{12}}{b}}{H_1} * \cdots * \hiding{\coloured{\Sigma_{n1}'}{r}*\coloured{\Sigma_{n2}}{b}}{H_n}$\\ (where for any coloured spatial formula $\Sigma$, its partition into red and blue heaplets is denoted by $\coloured{\Sigma_1}{r}*\coloured{\Sigma_2}{b}$), we define $A$ to be $\Pi' : \Sigma_{12} * \dotsi * \Sigma_{n2}$. This choice is justified by the following lemma: \begin{lemma} Suppose that\\ $\Pi : \coloured{\Sigma_1}{r} * \coloured{\Sigma_2}{b} \vdash \Pi' : \hiding{\coloured{\Sigma_{11}'}{r}*\coloured{\Sigma_{12}}{b}}{H_1} * \cdots * \hiding{\coloured{\Sigma_{n1}'}{r}*\coloured{\Sigma_{n2}}{b}}{H_n}$\\ is derivable using the rules of Fig.~, and that the antecedent is saturated. Then the following hold: \begin{compactitem} \item $\Pi' : \Sigma_{11}*\Sigma_{12}*\dotsi*\Sigma_{n2} \models \Pi' : H_1 * \dotsi * H_n$; \item $\Pi : \Sigma_1 \models \Pi' : \Sigma_{11} * \dotsi * \Sigma_{n1}$; and \item $\Pi : \Sigma_2 \models \Pi' : \Sigma_{12} * \dotsi * \Sigma_{n2}$. \end{compactitem} \end{lemma} \paragraph{Step 3} The third step of our algorithm is to check the entailment $\Pi : M * A \models R$. To illustrate why this is necessary, consider the following example: \begin{example} Suppose we want to solve the following bounded abduction problem: \[x \neq y : x \mapsto [a,y] \vdash \ls(x,y)*[\ ] \vdash x \mapsto [a,y]\ .\] In Step 1, we compute the colouring $x \neq y : \coloured{x \mapsto [a,y]}{r}*\coloured{\emp}{b}$ of the left hand side. In step 2, we compute the colouring $\coloured{x \mapsto [a,y]}{r}*\coloured{\emp}{b}$ of the right hand side. However, $\emp$ is not a solution to the bounded abduction problem. In fact, there is no solution to the bounded abduction problem. Intuitively, this is because $M$ is too weak to entail the red part of the right hand side. \eoe\end{example} \section{Discussion} We have presented \alg, a new technique for proving safety properties of programs requiring heap and data reasoning. \alg combines a new path-based separation logic analysis with first-order interpolation techniques for inferring intricate invariants. By bringing the advantages of the path-based refinement approach that has proven to be extremely effective in the domain of numerical and control-sensitive property verification to the domain of combined heap and data verification, we believe \alg is an important step in creating precise and generic automatic heap/data analyses. In this work we compute spatial($\mathcal{T}$) interpolants in a two-tiered manner: computing spatial interpolants followed by theory interpolants. This suffers from the problem that the computed spatial interpolants might not have theory interpolants (even though refinable spatial interpolants may exist), in which case our algorithm reverts to the strongest spatial interpolants. In the future, we would like to study systematically searching the space of spatial interpolants, until a set of spatial interpolants that can be refined with theory interpolants is found. We believe this can be performed in a counterexample-guided manner, where an unsolvable set of Horn clauses can inform us how to modify our spatial interpolants. \section{Putting it All Together} \begin{itemize} \item Describe how from path interpolants we can go to full proofs -- impact style. \item Describe how we need to take prefixes of paths in cases where the whole path does not have a proof of correctness. \item Finally, maybe describe how we can perform conjunction -- or we can handwave it if it's too much work. \end{itemize} \section{Implementation and Evaluation} Our primary goal is to study the feasibility of our proposed algorithm. To that end, we implemented an instantiation of our generic algorithm with the linked list recursive predicate $\ls$ (as defined in Sec.~) and refinements in the theory of linear arithmetic (QF\_LRA). The following describes our implementation and evaluation of \alg in detail. \paragraph{Implementation} We implemented \alg in the T2 safety and termination verifier~. Specifically, we extended T2's front-end to handle heap-manipulating programs, and used its safety checking component (which implements McMillan's \impact algorithm) as a basis for our implementation of \alg. To enable reasoning in separation logic, we implemented an entailment checker for \logic along with a bounded abduction procedure. We implemented a constraint-based solver using the linear rational arithmetic interpolation techniques of Rybalchenko and Stokkermans~ to solve the non-recursive Horn clauses generated by \alg. Although many off-the-shelf tools for interpolation exist (e.g.,~) we implemented our own solver for experimentation and evaluation purposes to allow us more flexibility in controlling the forms of interpolants we are looking for. We expect that \alg would perform even better using these highly tuned interpolation engines. \shortenpar Our main goal is to evaluate the feasibility of our proposed extension of interpolation-based verification to heap and data reasoning, and not necessarily to demonstrate performance improvements against other tools. Nonetheless, we note that there are two tools that target similar programs: (1) \textsc{Thor}~, which computes a memory safety proof and uses off-the-shelf numerical verifiers to strengthen it, and (2) \textsc{Xisa}~, which combines shape and data abstract domains in an abstract interpretation framework. \textsc{Thor} cannot compute arbitrary refinements of recursive predicates (like the ones demonstrated here and required in our benchmarks) unless they are manually supplied with the required theory predicates. Instantiated with the right abstract data domains, \textsc{Xisa} can in principle handle most programs we target in our evaluation. (\textsc{Xisa} was unavailable for comparison~.) Sec.~ provides a detailed comparison with related work. \paragraph{Benchmarks} To evaluate \alg, we used a number of linked list benchmarks that require heap and data reasoning. First, we used a number of simple benchmarks: \texttt{listdata} is similar to Fig.~, where a linked list is constructed and its data elements are later checked; \texttt{twolists} requires an invariant comparing data elements of two lists (all elements in list $A$ are greater than those in list $B$); \texttt{ptloop} tests our spatial interpolation technique, where the head of the list must not be folded in order to ensure its data element is accessible; and \texttt{refCount} is a reference counting program, where our goal is to prove memory safety (no double free). For our second set of benchmarks, we used a cut-down version of BinChunker (), a Linux utility for converting between different audio CD formats. BinChunker maintains linked lists and uses their data elements for traversing an array. Our property of interest is thus ensuring that all array accesses are within bounds. To test our approach, we used a number of modifications of BinChunker, \texttt{bchunk\_a} to \texttt{bchunk\_f}, where \texttt{a} is the simplest benchmark and \texttt{f} is the most complex one. \paragraph{Heuristics} We employed a number of heuristics to improve our implementation. First, given a program path to prove correct, we attempt to find a similar proof to previously proven paths that traverse the same control flow locations. This is similar to the \emph{forced covering} heuristic of~ to force path interpolants to generalize to inductive invariants. Second, our Horn clause solver uses Farkas' lemma to compute linear arithmetic interpolants. We found that minimizing the number of non-zero \emph{Farkas coefficients} results in more generalizable refinements. A similar heuristic is employed by~. \begin{table}[t] \centering \scalebox{0.8}{ \begin{tabular}{c|c|c|c|c} Benchmark & \#\ppath & Time (s) & $\theory$ Time & Sp. Time \\ \hline \hline \texttt{listdata} & 5 & 1.37 & 0.45 & 0.2 \\ \texttt{twolists} & 5 & 3.12 & 2.06 & 0.27 \\ \texttt{ptloop} & 3 & 1.03 & 0.28 & 0.15\\ \texttt{refCount} & 14 & 1.6 & 0.59 & 0.14\\ \hline \texttt{bchunk\_a} & 6 & 1.56 & 0.51 & 0.25\\ \texttt{bchunk\_b} & 18 & 4.78 & 1.7 & 0.2 \\ \texttt{bchunk\_c} & 69 & 31.6 & 14.3 & 0.26\\ \texttt{bchunk\_d} & 23 & 9.3 & 4.42 & 0.27 \\ \texttt{bchunk\_e} & 52 & 30.1 & 12.2 & 0.25 \\ \texttt{bchunk\_f} & 57 & 22.4 & 12.0 & 0.25 \\ \end{tabular}} \caption{Results of running \alg on our benchmark set. } \end{table} \paragraph{Results} Table~ shows the results of running \alg on our benchmark suite. Each row shows the number of calls to $\ppath$ (number of paths proved), the total time taken by \alg in seconds, the time taken to generate Horn clauses and compute theory interpolants ($\theory$ Time), and the time taken to compute spatial interpolants (Sp. Time). \alg proves all benchmarks correct w.r.t. their respective properties. As expected, on simpler examples, the number of paths sampled by \alg is relatively small (3 to 14). In the \texttt{bchunk\_*} examples, \alg examines up to 69 paths (\texttt{bchunk\_c}). It is important to note that, in all benchmarks, almost half of the total time is spent in theory interpolation. We expect this can be drastically cut with the use of a more efficient interpolation engine. The time taken by spatial interpolation is very small in comparison, and becomes negligible in larger examples. The rest of the time is spent in checking entailment of \logic formulas and other miscellaneous operations. Our results highlight the utility of our proposed approach. Using our prototype implementation of \alg, we were able to verify a set of realistic programs that require non-trivial combinations of heap and data reasoning. We expect the performance of our prototype implementation of \alg can greatly improve with the help of state-of-the-art Horn clause solvers, and more efficient entailment checkers for separation logic. \section{Overview} In this section, we demonstrate the operation of \alg (Fig.~) on the simple linked list example shown in Fig.~. We assume that integers are unbounded (i.e., integer values are drawn from $\mathbb{Z}$ rather than machine integers) and \setlength{\intextsep}{0pt} \setlength{\columnsep}{5pt} \begin{wrapfigure}{r}{4.6cm} \hfill \begin{minipage}{4.6cm} \begin{lstlisting} 1: int i = nondet(); node* x = null; 2: while (i != 0) node* tmp = malloc(node); tmp->N = x; tmp->D = i; x = tmp; i--; 3: while (x != null) 4: assert(x->D >= 0); x = x->N; \end{lstlisting} \end{minipage} \caption{Illustrative Example } \end{wrapfigure} that there is a \texttt{struct} called \texttt{node} denoting a linked list node, with a next pointer \texttt{N} and an integer (data) element \texttt{D}. The function \texttt{nondet()} returns a nondeterministic integer value. This program starts by building a linked list in the loop on location \texttt{2}. The loop terminates if the initial value of \texttt{i} is $\geq 0$, in which case a linked list of size \texttt{i} is constructed, where data elements \texttt{D} of list nodes range from \texttt{1} to \texttt{i}. Then, the loop at location \texttt{3} iterates through the linked list asserting that the data element of each node in the list is $\geq 0$. Our goal is to prove that the assertion at location \texttt{4} is never violated. \paragraph{Sample a Program Path} To start, we need a path $\pi$ through the program to the assertion at location \texttt{4}. Suppose we start by sampling the path \texttt{1,2,2,3,4}, that is, the path that goes through the first loop once, and enters the second loop arriving at the assertion. This path is illustrated in Fig.~ (where \texttt{2a} indicates the second occurrence of location \texttt{2}). Our goal is to construct a Hoare-style proof of this path: an annotation of each location along the path with a formula describing reachable states, such that location \texttt{4} is annotated with a formula implying that \texttt{x->D >= 0}. This goal is accomplished in two phases. First, we use \emph{spatial interpolation} to compute a memory safety proof for the path $\pi$ (Fig.~(b)). Second, we use \emph{theory refinement} to strengthen the memory safety proof and establish that the path satisfies the post-condition \texttt{x->D >= 0} (Fig.~(c)). \paragraph{Compute Spatial Interpolants} The first step in constructing the proof is to find \emph{spatial interpolants}: a sequence of separation logic formulas \emph{approximating} the shape of the heap at each program location, and forming a Hoare-style memory safety proof of the path. Our spatial interpolation procedure is a two step process that first symbolically executes the path in a forward pass and then derives a weaker proof using a backward pass. The backward pass can be thought of as an under-approximate weakest precondition computation, which uses the symbolic heap from the forward pass to guide the under-approximation. We start by showing the \emph{symbolic heaps} in Fig.~(a), which are the result of the forward pass obtained by symbolically executing \emph{only} heap statements along this program path (i.e., the strongest postcondition along the path). The separation logic annotations in Fig.~ follow standard notation (e.g.,~), where a formula is of the form $\Pi:\Sigma$, where $\Pi$ is a Boolean first-order formula over heap variables (pointers) as well as data variables (e.g., $x = \nil$ or $i > 0$), and $\Sigma$ is a \emph{spatial conjunction} of \emph{heaplets} (e.g., \textsf{emp}, denoting the empty heap, or $Z(x,y)$, a recursive predicate, e.g., that denotes a linked list between $x$ and $y$). For the purposes of this example, we assume a recursive predicate $\ls(x,y)$ that describes linked lists. In our example, the symbolic heap at location \texttt{2a} is $true:x\mapsto[d',\nil],$ where the heap consists of a node, pointed to by variable $x$, with $\nil$ in the \texttt{N} field and the (implicitly existentially quantified) variable $d'$ in the \texttt{D} field (since so far we are only interested in heap shape and not data). \shortenpar The symbolic heaps determine a memory safety proof of the path, but it is too strong and would likely not generalize to other paths. The goal of spatial interpolation is to find a sequence of annotations that are weaker than the symbolic heaps, but that still prove memory safety of the path. A sequence of spatial interpolants is shown in Fig.~(b). Note that all spatial interpolants are implicitly spatially conjoined with $\true$; for clarity, we avoid explicitly conjoining formulas with $\true$ in the figure. For example, location \texttt{2} is annotated with $true : \ls(x,\nil) * \true$, indicating that there is a list on the heap, as well as other potential objects not required to show memory safety. We compute spatial interpolants by going backwards along the path and asking questions of the form: \emph{how much can we weaken the symbolic heap while still maintaining memory safety?} We will describe how to answer such questions in Section~. \paragraph{Refine with Theory Interpolants} Spatial interpolants give us a memory safety proof as an approximate heap shape at each location. Our goal now is to strengthen these heap shapes with data refinements, in order to prove that the assertion at the end of the path is not violated. To do so, we generate a system of Horn clause constraints from the path in some first-order theory admitting interpolation (e.g., linear arithmetic). These Horn clauses carefully encode the path's data manipulation along with the spatial interpolants, which tell us heap shape at each location along the path. A solution of this constraint system, which can be solved using off-the-shelf interpolant generation techniques (e.g.,~), is a \emph{refinement} (strengthening) of the memory safety proof. \shortenpar In this example, we encode program operations over integers in the theory of linear integer arithmetic, and use Craig interpolants to solve the system of constraints. A solution of this system is a set of linear arithmetic formulas that refine our spatial interpolants and, as a result, imply the assertion we want to prove holds. One possible solution is shown in Fig.~(c). For example, location \texttt{2a} is now labeled with $true \color{black} : \mathsf{ls}(\color{OliveGreen} \lda{\nu. \nu \geq i} \color{black} ,x,\mathsf{null}),$ where the {\color{OliveGreen}green} parts of the formula are those added by refinement. Specifically, after refinement, we know that \emph{all} elements in the list from $x$ to $\nil$ after the first loop have data values greater than or equal to $i$, as indicated by the predicate $\color{OliveGreen} \lda{\nu. \nu \geq i}$. (In Section~, we formalize recursive predicates with data refinements.) Location \texttt{4} is now annotated with $\color{OliveGreen}d' \geq 0 \color{black} : x \mapsto [d',n'] * \true$, which implies that \texttt{x->D >= 0}, thus proving that the path satisfies the assertion. \paragraph{From Proofs of Paths to Proofs of Programs} We go from proofs of paths to whole program proofs implicitly by building an \emph{abstract reachability tree} as in \textsc{Impact}~. To give a flavour for how this works, consider that the assertions at \texttt{2} and \texttt{2a} are identical: this implies that this assertion is an inductive invariant at line \texttt{2}. Since this assertion also happens to be strong enough to prove safety of the program, we need not sample any longer unrollings of the first loop. However, since we have not established the inductiveness of the assertion at \texttt{3}, the proof is not yet complete and more traces need to be explored (in fact, exploring one more trace will do: consider the trace that unrolls the second loop once and shows that the second time \texttt{3} is visited can also be labeled with $\color{black} true \color{black} : \mathsf{ls}(\color{OliveGreen} \lda{\nu. \nu \geq 0} \color{black} ,x,\mathsf{null})$). \shortenpar Since our high-level algorithm is virtually the same as \textsc{Impact}~, we will not describe it further in the paper. For the remainder of this paper, we will concentrate on the novel contribution of our algorithm: computing spatial interpolants with theory refinements for program paths. \section{Introduction} Since the problem of determining whether a program satisfies a given property is undecidable, every verification algorithm must make some compromise. There are two classical schools of program verification, which differ in the compromise they make: the \emph{static analysis} school gives up refutation soundness (i.e., may report \emph{false positives}); and the \emph{software model checking} school gives up the guarantee of termination. In the world of integer program verification, both schools are well explored and enjoy cross-fertilization of ideas: each has its own strengths and uses in different contexts. In the world of heap-manipulating programs, the static analysis school is well-attended , while the software model checking school has remained essentially vacant. This paper initiates a program to rectify this situation, by proposing one of the first path-based software model checking algorithms for proving combined shape-and-data properties. The algorithm we propose, \alg, marries two celebrated program verification ideas: McMillan's \emph{lazy abstraction with interpolants} (\impact) algorithm for software model checking , and \emph{separation logic}, a program logic for reasoning about shape properties . \alg (like \impact) is based on a path-sampling methodology: given a program $P$ and safety property $\varphi$, \alg constructs a proof that $P$ is memory safe and satisfies $\varphi$ by sampling a finite number of paths through the control-flow graph of $P$, proving them safe, and then assembling proofs for each sample path into a proof for the whole program. The key technical advance which enables \alg is an algorithm for \emph{spatial interpolation}, which is used to construct proofs in \emph{separation logic} for the sample traces (serving the same function as \emph{Craig interpolation} for first-order logic in \impact). \shortenpar \alg is able to prove properties requiring integrated heap and data (e.g., integer) reasoning by strengthening separation logic proofs with \emph{data refinements} produced by classical Craig interpolation, using a technique we call \emph{spatial interpolation modulo theories}. Data refinements are \emph{not tied to a specific logical theory}, giving us a rather generic algorithm and freedom to choose an appropriate theory to encode a program's data. Fig.~ summarizes the high-level operation of our algorithm. Given a program with no heap manipulation, \alg only computes theory interpolants and behaves exactly like \impact, and thus one can thus view \alg as a proper extension of \impact to heap manipulating programs. At the other extreme, given a program with no data manipulation, \alg is a new shape analysis that uses path-based relaxation to construct memory safety proofs in separation logic. \shortenpar There is a great deal of work in the static analysis school on shape analysis and on combined shape-and-data analysis, which we will discuss further in Sec.~. We do not claim superiority over these techniques (which have had the benefit of 20 years of active development). \alg, as the first member of the software model checking school, is not \emph{better}; however, it \emph{is} fundamentally \emph{different}. Nonetheless, we will mention two of the features of \alg (not enjoyed by any previous verification algorithm for shape-and-data properties) that make our approach worthy of exploration: path-based refinement and property-direction. \begin{compactitem} \item \emph{Path-based refinement}: This supports a progress guarantee by tightly correlating program exploration with refinement, and by avoiding imprecision due to lossy join and widening operations employed by abstract domains. \alg does not report false positives, and produces counterexamples for violated properties. This comes, as usual, at the price of potential divergence. \item \emph{Property-direction}: Rather than seeking the strongest invariant possible, we compute one that is \emph{just strong enough} to prove that a desired property holds. Property direction enables scalable reasoning in rich program logics like the one described in this paper, which combines separation logic with first-order data refinements. \end{compactitem} We have implemented an instantiation of our generic technique in the \textsc{T2} verification tool~, and used it to prove correctness of a number of programs, partly drawn from open source software, requiring combined data and heap invariants. Our results indicate the usability and promise of our approach. \paragraph{Contributions} We summarize our contributions as follows: \begin{enumerate} \item A generic property-directed algorithm for verifying and falsifying safety of programs with heap and data manipulation. \item A precise and expressive separation logic analysis for computing memory safety proofs of program paths using a novel technique we term \emph{spatial interpolation}. \item A novel interpolation-based technique for strengthening separation logic proofs with data refinements. \item An implementation and an evaluation of our technique for a fragment of separation logic with linked lists enriched with linear arithmetic refinements. \end{enumerate} \section{Introduction} The past decade has witnessed significant advances in the quest for efficient automated software verification techniques. For example, advances such as predicate abstraction~, interpolation~, and SMT solving~ have facilitated efficient verification of control-sensitive and numerical properties. Within the same arena of numerical properties, techniques employing numerical abstract domains~ have been successfully applied to large safety-critical systems~. On the other hand, in the heap world, scalable techniques based on separation logic~ have been applied to proving memory safety of large low-level software~. However, combined heap and data reasoning is comparatively under-explored. For instance, memory safety analyses generally either try to prove array accesses are within bounds assuming pointers to arrays are valid, or to prove accesses through pointers are valid assuming accesses to valid arrays are within bounds. This paper is concerned with verifying safety properties that require heap as well as data reasoning. Consider, for example, a program that stores array indices in a linked list and then uses them to access the array, or a program that uses reference counting to manage memory. In order to prove memory safety of such programs, we need to be able to infer intricate invariants that involve both data and heap shape. The aforementioned techniques are unable to prove safety of such programs: software model checking techniques track only very shallow heap information (e.g., using traditional pointer analysis), and most separation logic and shape analysis techniques do not track data invariants of heap data structures. In this paper, we propose \alg, a novel safety verification algorithm that is able to prove properties requiring integrated heap and data reasoning. \alg marries (1) a new form of rich separation logic reasoning, for \emph{lazily} inferring heap invariants, with (2) interpolation-based reasoning, for strengthening heap invariants with \emph{data refinements}. Given a program $P$ and safety property $\varphi$, \alg constructs a proof that $P$ is memory safe and satisfies $\varphi$ by \emph{sampling paths} through the control-flow graph of $P$ and proving them safe. That is, by proving safety of a finite number of paths (samples), we construct a proof of the whole program. This is inspired by McMillan's \emph{lazy abstraction with interpolants} (\impact) verification technique~, where \emph{Craig interpolants} are used to prove individual paths. The \emph{path sampling} methodology (e.g., as in~) enables the following features: \vspace{-.05in} \begin{enumerate} \item \emph{Path-based refinement}: This supports a progress guarantee by tightly correlating program exploration with refinement, and by avoiding imprecision due to lossy join and widening operations employed by abstract domains. \alg does not report false positives, and produces counterexamples for violated properties. This comes, as usual, at the price of potential divergence. \item \emph{Property-direction}: Rather than seeking the strongest invariant possible, we compute one that is \emph{just strong enough} to prove that a desired property holds. Property direction enables scalable reasoning in rich program logics (like the one described in this paper), which is necessary for proving properties combining data and the heap. \end{enumerate} Figure~ summarizes the high-level operation of our algorithm. Notice that, at one extreme, given a program with no heap manipulation, \alg only computes theory interpolants and behaves exactly like \impact. One can thus view \alg as a proper extension of \impact to heap manipulating programs. At the other extreme, given a program with no data manipulation, our algorithm is a new shape analysis that uses path-based relaxation to construct memory safety proofs in separation logic. Thus, \alg is a step forward in the direction towards fully automatic verification: since it integrates both numerical and heap reasoning, it obviates the need to choose which category of verification tool to apply to a program of interest. Moreover, it can be used to prove properties that lie beyond the scope of either category. There is some prior work on combined heap and data analysis, discussed in Section~, but none enjoys \alg's two key features: path-based refinement and property-direction. At the core of \alg is the idea of \emph{spatial interpolants modulo theories}, which are used to construct a proof of correctness of a given program path. \paragraph{Spatial Interpolants} To prove a program path $\pi$ safe, we first construct a Hoare-style proof of memory safety of $\pi$ using a new separation logic--based technique. We call the annotations along the path \emph{spatial path interpolants}, since, in the style of interpolation-based verification, their logical strength is \emph{in between} that of the sequences of strongest postconditions and of weakest preconditions. Effectively, spatial interpolants tell us which parts of the heap, and their structure, we should remember at each point along the path in order to ensure safe execution thereafter. Consequently, our technique does not suffer from the imprecision incurred by forward-running shape analyses, which might \emph{abstract too much} (e.g., by widening) and report unsafe memory operations in a safe program. Our spatial interpolation procedure is a two step process that first symbolically executes the path in a forward pass and then derives a weaker proof using a backward pass. The backward pass can be thought of as an under-approximate weakest precondition computation, which uses the symbolic heap from the forward pass to guide the under-approximation. This second pass is the principal point at which heuristic approximation is involved. The approximation strives toward the weakest preconditions while, in the limit, ``unsuccessful'' approximation yields the overly-strong proof from the first pass. \paragraph{Spatial Interpolants Modulo Theories} Given a memory safety proof of a path $\pi$ (spatial interpolants), we aim to strengthen it with data invariants such that the proof establishes that the path satisfies the safety property $\varphi$. To do so, we generate a system of Horn clause constraints from the path in some first-order theory admitting interpolation (e.g., linear arithmetic). These Horn clauses carefully encode the path's data manipulation along with the spatial interpolants, which tell us heap shape at each location along the path. The solution of this constraint system, which can be solved using off-the-shelf first-order interpolant generation techniques (e.g.,~), is a \emph{refinement} (strengthening) of the memory safety proof. For example, we might transform an assertion $\ls(x,\nil)$ -- there is a linked list from $x$ to $\nil$ -- into the stronger formula $\ls(\lda{a.a > 0}, x, \nil)$ -- all elements of the list from $x$ to $\nil$ are greater than $0$. We call the resulting annotations \emph{spatial interpolants modulo theories} (or \emph{spatial($\theory$)} interpolants for short). Note that our data refinements are \emph{not tied to a specific logical theory} $\theory$, giving us a rather generic algorithm and freedom to choose an appropriate theory to encode a program's data. Heap shapes are expressed using a fragment of separation logic with general recursive predicates. The (second-order) recursive predicates are parameterized by data refinement predicates that at each unfolding are able to constrain a finite history or window of the data stored in memory described by the recursive predicate. The recursive predicate definitions contribute to the Horn constraints, and first-order interpolants are used to refine the data refinement predicate parameters of recursive predicates. We have implemented an instantiation of our generic technique in the \textsc{T2} safety and termination verification tool~, and used it to prove correctness of a number of programs, partly drawn from open source software, requiring combined data and heap invariants. Our results indicate the usability and promise of our approach. \paragraph{Contributions} We summarize our contributions as follows: \begin{compactitem} \item A generic property-directed algorithm for verifying and falsifying safety of programs with heap and data manipulation. \item A precise and expressive separation logic analysis for computing memory safety proofs of program paths using a novel technique we term \emph{spatial interpolation}. \item A novel interpolation-based technique for strengthening separation logic proofs with data refinements. \item An implementation and an evaluation of our technique for a fragment of separation logic with linked lists enriched with linear arithmetic refinements. \end{compactitem} \emph{The \textbf{appendix} contains an extended version (A) a complete description of \alg, (B) proofs, (C) complete semantics and proof system of our program logic, and (D) other extended clarifications. } \section{Preliminaries} We deal with list-manipulating programs, where each list node contains a data field. Our separation logic fragment is the standard one with lists segments, points-to predicates, and existential quantifiers. \section{Memory Safety Proof Relaxation} Let $\pi = p_1, \ldots, p_n$ be a program path and $C = c_0,\ldots, c_n$ be the result of symbolically executing this path (without applying abstraction rules), where $c_0 \equiv true : \emp$. We assume that symbolic execution does not perform an unsafe memory operation. As a result, we would like to find a weaker proof $W = w_0,\ldots,w_n$ that still preserves memory safety. Each symbolic state $c_i$ is of the form $\Pi_i:\Sigma_i$, where $\Sigma_i$ is treated as a set of points-to predicates (since no abstraction is performed). An empty set $\Sigma_i$ denotes $\emp$. Computing $w_0, \ldots, w_n$ is performed backwards starting from $w_n$. We describe the algorithm as follows: \textbf{[Note:The following assumes that the only non-determinism in the program is in the control flow. That is, symbolic execution without abstraction entails either the condition of the assume statement or its negation.]} \subsection{Initialization ($w_n$)} $w_n = \ltrue:\true$. We also create a function $f_n$ that maps each points-to predicate in $\Sigma_n$ to the predicate $\true$ in $w_n$. Effectively, $f_n$ specifies a fine-grained entailment relation between heaplets in $c_n$ and $w_n$. \subsection{Computing $w_i$ (without abstraction)} We now show how to compute $w_i$, where $i < n$. There are a number of cases, depending on the command $p_{i+1}$ along the path. \paragraph{Assignment} When $p_{i+1}$ is an assignment of the form $\texttt{x := y}$, then $$w_{i} = \Pi[y/x]:\Sigma[y/x],$$ where $w_{i+1} = \Pi:\Sigma$. The map $f_i$ from $c_i$ to $w_i$ is the same as $f_{i+1}$, since the list and points-to predicates have not changed. \paragraph{Data-field assignment} When $p_{i+1}$ is an assignment of the form $\texttt{d := y->D}$ or $\texttt{y->D := d}$, then we have to enforce that $w_i$ entails the existence of a cell $y \mapsto [\_]$, thus ensuring a safe memory operation. Let points-to predicate $X = x \mapsto [d,n]$ such that $X \in \Sigma_{i+1}$ and $\Pi_{i+1} \vdash x = y$. Then, $$w_i = \Pi \land (x = y) : (\Sigma - f_{i+1}(X)) * x \mapsto [d,n] * \sub(f_{i+1}(X), x \mapsto [d,n]),$$ where $w_{i+1} = \Pi:\Sigma$, and $\sub(Pred, x \mapsto [d,n])$ is defined as follows: \begin{itemize} \item If $Pred = \true$, then result is $\true$. \item If $Pred = ls(z,w)$, then result is $ls(z,x) * ls(n,w)$, where $n$ is a fresh (existentially quantified) variable. \item If $Pred = y \mapsto [d,n]$, then result is $\emp$. \end{itemize} $f_{i}$ is set to $f_{i+1}$, with $X$ mapping to the points-to predicate $x \mapsto [d,n]$ and the rest of the predicates mapping to $f_{i+1}(X)$ now map to the result of $\sub$. Note that in case $\sub$ returns two lists (case 2), then we have to split the predicates mapping to the result of subtract to those that can reach $X$, and those that $X$ can reach, the former and the latter predicates map to $ls(z,x)$ and $ls(n,w)$, respectively, in $f_i$. \paragraph{Allocation} When $p_{i+1}$ is an allocation of the form $\texttt{alloc(y)}$, Let points-to predicate $X = y \mapsto [d,n]$ such that $X \in \Sigma_{i+1}$. Then, $$w_i = (\Pi:\Sigma - f_{i+1}(X) * \sub(f_{i+1}(X), y \mapsto [d,n]))[y'/y],$$ where $w_{i+1} = \Pi : \Sigma$. $f_i$ is the same as $f_{i+1}$, where $y \mapsto [d,n]$ where $X$ is not in the domain. \paragraph{De-allocation} When $p_{i+1}$ is a de-allocation statement of the form $\texttt{free(y)}$, then there exists a predicate $X = x \mapsto [d,n]$ such that $X \in \Sigma_{i}$ and $\Pi_i \vdash x = y$. Thus, $$w_i = \Pi \land (x = y) : x \mapsto [d,n] * \Sigma,$$ where $w_{i+1} = \Pi : \Sigma$. $f_i$ is the same as $f_{i+1}$, with the difference that $X$ maps to the predicate $x \mapsto [d,n]$ in $f_i$. \paragraph{Next-field assignment} When $p_{i+1}$ is of the form $\texttt{z := y->n}$, let points-to predicate $X = x \mapsto [d,n]$ such that $X \in \Sigma_{i+1}$ and $\Pi_{i+1} \vdash x = y$. Then, $$w_i = \Pi[n/z] \land (x = y):$$ $$(\Sigma[n/z] - f_{i+1}(X)) * x \mapsto [d,n] * \sub(f_{i+1}(X), x \mapsto [d,n])[n/z],$$ Similarly, if $p_{i+1}$ is of the form $\texttt{y->n := z}$, then $$w_i = \Pi[z/n] \land (x = y):$$ $$(\Sigma[z/n] - f_{i+1}(X)) * x \mapsto [d,n] * \sub(f_{i+1}(X), x \mapsto [d,n])[z/n],$$ $f_i$ is set as described for data-field assignment statements. \paragraph{Assumptions} When $p_{i+1}$ is of the form $\texttt{assume(X)}$, where $\texttt{X}$ is $\texttt{x = y}$ or $\texttt{x != y}$, then $$w_i = (X \Rightarrow \Pi):\Sigma,$$ where $w_{i+1} = \Pi:\Sigma$. Note that we've added a disjunction here. \subsection{Transforming $w_i$ (list introduction)} In the previous rules, we always only exposed the heaplet that is required for ensuring memory safety. In order to get a proof $W$ that is more likely to be inductive, we need to be more aggressive, by not only introducing heaplets, but folding sequences of heaplets into lists. The following transformation rules are used to introduce list segments in some $w_i$. Note that given a $w_i$, applying the following rules doesn't necessarily produce a new $w_i'$ that is weaker than $w_i$. The only guarantee is that the resulting $w_i'$ satisfies $\{w_i'\}p_{i+1}\{w_{i+1}\}$. The following assumes $w_i$ is of the form $\Pi:\Sigma$. \paragraph{Carving a list out of $\true$} Let $x\mapsto[n,d]$ be a predicate in $w_i$ produced by the aforementioned rules. Let the set of predicates $$S = \{e_0 \mapsto[e_1,\_], e_1' \mapsto[e_2,\_], e_2' \mapsto [e_3,\_], \ldots, e_{n-1}' \mapsto [e_n,\_]\} \subseteq \Sigma_i$$ such that for all $1 \leq j \leq n-1$, $\Pi_{i}:\Sigma_i \vdash e_j = e_j'$, and $\Pi_i:\Sigma_i \vdash x = e_0$ and $e_n$ doesn't equal any of the other $e_i$ or $e'_i$ variables (i.e., the sequence of predicates forms an acyclic list). Assuming that all predicates $e_i' \mapsto [e_{i_1},\_]$ map to $\true$ in $f_i$, then $$w_i' = \Pi \land x \neq e' : (\Sigma - x\mapsto[n,d]) * ls(x,e'),$$ where $e'$ is a fresh existentially quantified variable. \emph{Note: If $\Pi_i:\Sigma_i \vdash e_n = null$, then we can replace $e'$ with $null$ in $w_i'$, which is often desirable when dealing with null-terminated lists.} \zak{Making the decision to always introduce an existentially quantified variable as a list endpoint is probably OK - I think it just shifts the burden of choosing how to instantiate that quantifier to later (e.g., we can probably delay the choice to the cutpoint, and we will have collected a bunch of disequalities in the antecedent of $\Sigma$ by passing through the assumptions between the cutpoint and the error location).} \paragraph{Carving lists through other lists} Now assume the above predicates in $S$ are not all mapped to $\true$ in $f_i$, but some of them map to list segments in $w_i$. If we make the additional assumption that for all $X \in S$, if $f_i(X) = ls(e',y)$, then $f_i^{-1}(ls(e',y)) \subset S$, no program variable $v$ aliases $e'$ in $w_i$, and $ls(e',y)$ could be empty in $w_{i+1}$, then $$w_i' = \Pi \land x \neq e' : (\Sigma - lists) * ls(x,e'),$$ where $lists = \{f_i(X) \mid X \in S \text{ and } f_i(X) \text{ is a list segment}\}$. As before, $e'$ can be replaced by $null$ or a program variable $v$ that aliases $e_n$. Of course, after these transformations, all predicates in the set $S$ now map to the newly introduced list segment $ls(x,e')$ in $f_i$. \section{Formalism} \subsection{Syntax} \begin{align*} x,y \in \textsf{HVar}&& \text{Heap variables} \\ a,b \in \textsf{AVar}&& \text{Arithmetic variables} \\ X \subseteq \textsf{Var}&::= x | a\\ E,F \in \textsf{HTerm}&::= \nil \mid x\\ t,u \in \textsf{ATerm}&::= k \mid a \mid t + t \mid t - t \mid t \cdot t\\ \Pi \in \textsf{Pure}&::= \true \mid E = E \mid E \neq E \mid t \leq t \mid \Pi \land \Pi\\ \Sigma \in \textsf{Spatial}&::= \true \mid \emp \mid E \mapsto [t,E] \mid \ls(\Pi,E,E) \mid \Sigma * \Sigma\\ P \in \textsf{Formula}&::= (\exists X)(\Pi : \Sigma) \end{align*} \subsection{Semantics} We use $+$ to denote disjoint union of sets and $\oplus$ to denote disjoint union of functions. \begin{align*} \textsf{Var} &= \textsf{HVar} + \textsf{AVar}\\ \textsf{Val} &= \textsf{Loc} + \mathbb{Z}\\ \textsf{Stack} &= \textsf{Var} \rightarrow \textsf{Val}\\ \textsf{Heap} &= \textsf{Loc} \rightharpoonup_{\textsf{fin}} \mathbb{Z} \times \textsf{Loc}\\ \textsf{State} &= \textsf{Stack} \times \textsf{Heap} \end{align*} \begin{align*} s,h \models E = F &\iff \sem{E}(s) = \sem{F}(s)\\ s,h \models E \neq F &\iff \sem{E}(s) \neq \sem{F}(s)\\ s,h \models t \leq u &\iff \sem{t}(s) \leq \sem{u}(s)\\ s,h \models \phi \land \psi &\iff (s,h \models \phi) \land (s,h \models \psi)\\ s,h \models \emp &\iff \dom(h) = \emptyset\\ s,h \models E \mapsto [t, F] & \iff \dom(h) = \{\sem{E}(s)\} \land h(\sem{E}(s)) = \tuple{\sem{t}(s), \sem{F}(s)}\\ s,h \models \ls(\phi,E,F) & \iff (s,h \models E = F \land \emp) \\&\hspace*{1cm}\lor (\exists k,E'. s,h \models E \neq F \land \phi[d/k] : E \mapsto [k,E'] * \ls(\phi,E',F))\\ s,h \models \Sigma * \Sigma' &\iff \exists h_0,h_1. h_0 \oplus h_1 = h \land (s,h_0 \models \Sigma) \land (s,h_1 \models \Sigma')\\ s,h \models \Pi : \Sigma &\iff (s,h \models \Pi) \land (s,h \models \Sigma)\\ s,h \models (\exists X)(P) &\iff \exists \overline{s} : X \rightarrow \in \textsf{Val} \text{ such that } s\oplus \overline{s},h \models P\\ \end{align*} \subsection{Predicate transformer} \begin{definition}[Witness] Let $\Sigma = \Sigma_1 * \dotsi * \Sigma_N$ be a spatial formulae such that each $\Sigma_i$ is an atom, and let $s \in \textsf{Stack}$ and $h \in \textsf{Heap}$. A \emph{witness} for the state $\tuple{s,h}$ and the formula $(\exists X)(\Pi : \Sigma)$ is a pair $\omega = \tuple{\rho, \overline{s}}$ consisting of a map $\rho : \dom(h) \rightarrow [1,N]$ and a stack $\overline{s} : X \rightarrow \textsf{Val}$ such that \begin{enumerate} \item $s \oplus \overline{s},h_i \models \Pi$ \item For all $i \in [1,N]$, $s \oplus \overline{s},h_i \models \Sigma_i$\\ where $h_i = h|_{\{ x \in \dom(h) : \rho(x) = i\}}$. \end{enumerate} If $\omega$ is a witness for $\tuple{s,h}$ and $(\exists X)(\Pi : \Sigma)$, we write \[s,h \models_\omega (\exists X)(\Pi : \Sigma)\] \end{definition} \begin{lemma}[Partition] Let $s,h$ be a state and $(\exists X)(\Pi:\Sigma)$ be a formula. Then $s,h \models (\exists X)(\Pi : \Sigma)$ iff there exists a witness $\omega = \tuple{\rho, \overline{s}}$ such that \[s,h \models_\omega (\exists X)(\Pi : \Sigma)\] \end{lemma} Let $Q = \Pi : \Sigma_1 * \dotsi * \Sigma_N$, $c \in \Cmd$, $s,s'$ be stores, $h,h'$ be heaps such that $\tuple{s,h}\sem{c}\tuple{s',h'}$, and $\omega' = \tuple{\rho', \overline{s}'}$ be such that $s',h' \models_{\omega'} Q$. We define the \emph{precondition} of $Q$ along $c$ as follows: \begin{itemize} \item Case: $c$ is \texttt{x := y->next} \begin{itemize} \item Case: $\Sigma_{\rho'(\sem{\texttt{y}}(s'))} = \texttt{z} \mapsto [d,n]$ \[\pre(c,Q,\rho,s,h,s',h') = (\exists X)(\texttt{y} = \texttt{z} \land \Pi[\texttt{n}/\texttt{x}]: \Sigma_1' * \dotsi * \Sigma_N') \] where for each $i$ \[\Sigma_i' = \Sigma_i[\texttt{n}/\texttt{x}] \] We define $\omega = \omega'$. \item Case: $\Sigma_{\rho'(\sem{\texttt{y}}(s'))} = \ls(\phi,E,F)$ Let $\tuple{d, n} = h(s(\texttt{y}))$, and let $\texttt{d},\texttt{n}$ be fresh variable symbols. \[\pre(c,Q,\omega',s,h,s',h') = (\exists X \cup \{\texttt{d},\texttt{n}\})(\Pi[\texttt{n}/\texttt{x}]: \Sigma_1' * \dotsi * \Sigma_N') \] where for each $i$ \[\Sigma_i' = \begin{cases} (\ls(\phi,E,\texttt{y}) * \texttt{y} \mapsto [\texttt{d},\texttt{n}] * \ls(\phi,\texttt{n},F))[\texttt{n}/\texttt{x}] & \text{if } i=\rho'(\sem{\texttt{y}}(s'))\\ \Sigma_i[\texttt{n}/\texttt{x}] & \text{otherwise} \end{cases}\] We define $\omega = \tuple{\rho, \overline{s}}$ by \[ \rho(\ell) = \begin{cases} \rho'(\ell) & \text{if } \rho'(\ell) < \rho'(\sem{\texttt{y}}(s'))\\ \rho'(\sem{\texttt{y}}(s)) & \text{if } \rho'(\ell) = \rho'(\sem{\texttt{y}}(s')) \land \ell \in Between(s',h',\sem{E}(s'),\sem{\texttt{y}}(s')) \\ \rho'(\sem{\texttt{y}}(s)) + 1 & \text{if } \ell = \sem{\texttt{y}}(s')\\ \rho'(\ell) + 2 & \text{otherwise} \end{cases}\] \[ \overline{s}' = \overline{s}[\texttt{d} \gets d, \texttt{n} \gets n] \] \item Case: $\Sigma_{\rho'(\sem{\texttt{y}}(s'))} = \true$ Let $\tuple{d, n} = h(s(\texttt{y}))$, and let $\texttt{d},\texttt{n}$ be fresh variable symbols. \[\pre(c,Q,\omega',s,h,s',h') = (\exists X \cup \{\texttt{d},\texttt{n}\})(\Pi[\texttt{n}/\texttt{x}] : \Sigma_1' * \dotsi * \Sigma_N') \] where for each $i$ \[\Sigma_i' = \begin{cases} \texttt{y} \mapsto [\texttt{d},\texttt{n}] * \true & \text{if } i=\rho'(\sem{\texttt{y}}(s'))\\ \Sigma_i[\texttt{n}/\texttt{x}] & \text{otherwise} \end{cases}\] We define $\omega = \tuple{\rho, \overline{s}}$ by \[ \rho(\ell) = \begin{cases} \rho'(\ell) & \text{if } \rho'(\ell) < \rho'(\sem{\texttt{y}}(s'))\\ \rho'(\sem{\texttt{y}}(s)) & \text{if } \ell = \sem{\texttt{y}}(s')\\ \rho'(\sem{\texttt{y}}(s)) & \text{if } \ell \neq \sem{\texttt{y}}(s') \land \rho'(\ell) = \rho'(\sem{\texttt{y}}(s'))\\ \rho'(\ell) + 1 & \text{otherwise} \end{cases}\] \[ \overline{s}' = \overline{s}[\texttt{d} \gets d, \texttt{n} \gets n] \] \end{itemize} \item ... \end{itemize} \begin{lemma} Let $Q = \Pi : \Sigma_1 * \dotsi * \Sigma_N$, $c \in \Cmd$, $s,s'$ be stores, $h,h'$ be heaps such that $\tuple{s,h} \sem{c} \tuple{s',h'}$, and $\omega' : \dom(h') \rightarrow [1,N]$ be such that $s',h' \models_\omega Q$. Let $\omega,P$ be such that $\pre(c,Q,\rho,s,h,s',h') = \tuple{\omega,P}$. Then the following hold: \begin{enumerate} \item $s,h \models_\omega P$ \item $\hoare{P}{c}{Q}$ \end{enumerate} \end{lemma} \begin{proposition} Let $\tau = \tuple{s_0,h_0}\texttt{c}_0\tuple{s_1,h_1}\texttt{c}_1 \dotsi \tuple{s_n,h_n}$ be a program path, and let $\Pi,\Sigma,\omega$ be such that $s_n,h_n \models_\omega \Pi:\Sigma$. Define a sequence of predicates $\{P_i\}$ by \begin{itemize} \item $P_n = Q_n$, $\omega_n = \omega$ \item $\tuple{P_i,\omega_i} = \pre(\texttt{c}_{i},P_{i+1},s_i,h_i,s_{i+1},h_{i+1},\omega_{i+1}))$. \end{itemize} Then $\{ P_0 \} \texttt{c}_0 \{ P_1 \} \dotsi \{ P_{n-1} \} \texttt{c}_{n-1} \{ P_n \}$ is a valid Hoare proof. \end{proposition} \subsection{Proof system} \begin{figure*} \begin{mathpar} \inferrule[Arith-Weak]{ \Pi \vDash \Pi' }{ \Pi : \Sigma \vdash \Pi' : \Sigma } \inferrule[Refine-Weak]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,y) \\ \Pi' \land \phi \vDash \psi }{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\psi,x,y) } \inferrule[Rearrangement]{ \Pi : \Sigma \vdash \Pi' : \Sigma_0 * \Sigma_1}{ \Pi : \Sigma \vdash \Pi' : \Sigma_1 * \Sigma_0 } \inferrule[Fold/Base]{ \Pi : \Sigma \vdash \Pi' : \Sigma'}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,x) } \inferrule[Fold/Rec]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,y] * \ls(\phi,y,z) \\ \Pi' \land \nu = d \vDash \phi}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,z) } \inferrule[Fold/Seg-Null]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,y) * \ls(\phi,y,\nil)}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,\nil) } \inferrule[Fold/Seg-Pt]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,y) * \ls(\phi,y,z) * z \mapsto [d,n]}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,z) * z \mapsto [d,n] } \inferrule*[lab=Unfold/Rec,right={\rm $d,n$ fresh}]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,y) \\ \Pi' \vDash x \neq y }{ \Pi : \Sigma \vdash \Pi' \land \phi[d/\nu] : \Sigma' * x \mapsto [d,n] * \ls(\phi,h,y) } \inferrule[Unfold/Base]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,y) \\ \Pi' \vDash x = y}{ \Pi : \Sigma \vdash \Pi' : \Sigma' } \inferrule*[lab=Drop/Pt,right={\rm $d$ not free in $\Pi':\Sigma'$}]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n]}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \true } \inferrule[Drop/Ls]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(\phi,x,y)}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \true } \end{mathpar} \caption{Entailment rules} \end{figure*} \begin{figure*} \begin{mathpar} \inferrule[Assign]{\Pi : \Sigma \vdash \Pi' : \Sigma'}{\hoare{\Pi : \Sigma}{x := E}{\Pi'[x'/x] \land x = E[x'/x] : \Sigma'[x'/x]}} \inferrule[Assume]{\Pi : \Sigma \vdash \Pi' : \Sigma'}{\hoare{\Pi : \Sigma}{assume($\phi$)}{\Pi' \land \phi : \Sigma'}} \inferrule[Arith-Store]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n]}{\hoare{\Pi : \Sigma}{x->D := E}{\Pi'[d'/d] \land d=E[d'/d] : \Sigma'[d'/d]}} \inferrule[Arith-Load]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n]}{\hoare{\Pi : \Sigma}{y := x->D}{\Pi'[y'/y] \land y = d : \Sigma'[y'/y]}} \inferrule[Heap-Store]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n]}{\hoare{\Pi : \Sigma}{x->D := E}{\Pi'[d'/d] \land d=E[d'/d] : \Sigma'[d'/d]}} \inferrule[Heap-Load]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n]}{\hoare{\Pi : \Sigma}{y := x->D}{\Pi'[y'/y] \land y = d : \Sigma'[y'/y]}} \inferrule*[lab=Alloc,right={\rm $d,n$ free in $\Pi':\Sigma'$}]{\Pi : \Sigma \vdash \Pi' : \Sigma'}{\hoare{\Pi : \Sigma}{x := new list}{\Pi'[x'/x] : \Sigma'[x'/x] * x \mapsto [d,n]}} \inferrule[Free]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n]}{\hoare{\Pi : \Sigma}{free(x)}{\Pi' : \Sigma'}} \end{mathpar} \caption{Execution rules} \end{figure*} \newpage \begin{figure*} \begin{mathpar} \inferrule[Arith-Weak]{\Pi \vDash \Pi'}{ \Pi \land R(\vec{x}) : \Sigma \vdash \Pi' \land R'(\vec{x}) : \Sigma \triangleright R'(\vec{x}) \leftarrow R(\vec{x}) } \inferrule[Rearrangement]{ \Pi : \Sigma \vdash \Pi' : \Sigma_0 * \Sigma_1 \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma_1 * \Sigma_0 \triangleright \mathcal{C} } \inferrule[Fold/Base]{ \Pi : \Sigma \vdash \Pi' : \Sigma' \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R,x,x) \triangleright \mathcal{C} } \inferrule[lab=Fold/Rec]{ \Pi : \Sigma \vdash \Pi' \land P(\vec{x}) : \Sigma' * x \mapsto [d,y] * \ls(R,y,z) \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R,x,z) \triangleright \mathcal{C}; R(\nu, \vec{x}) \leftarrow P(\vec{x}) \land \nu = d } \inferrule[Fold/Seg-Null]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R_0,x,y) * \ls(R_1,y,\nil) \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R,x,\nil)\triangleright \mathcal{C}; R(\nu,\vec{x}) \gets R_0(\nu,\vec{x}) \lor R_1(\nu,\vec{x})} \inferrule[Fold/Seg-Pt]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R_0,x,y) * \ls(R_1,y,z) * z \mapsto [d,n] \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R,x,z) * z \mapsto [d,n]\triangleright \mathcal{C}; R(\nu,\vec{x}) \gets R_0(\nu,\vec{x}) \lor R_1(\nu,\vec{x}) } \inferrule*[lab=Unfold/Rec,right={\rm $d,n$ fresh}]{ \Pi : \Sigma \vdash \Pi' \land R_p(\vec{x}): \Sigma' * \ls(R_{ls},x,y) \triangleright \mathcal{C} \\ \Pi' \vDash x \neq y }{ \Pi : \Sigma \vdash R_p'(\vec{x},d) : \Sigma' * x \mapsto [d,n] * \ls(R_{ls},h,y)\triangleright \mathcal{C}; R_p'(\vec{x},d) \leftarrow R_p(\vec{x}) \land R_{ls}(d,\vec{x}) } \inferrule[Unfold/Base]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R,x,y) \triangleright \mathcal{C} \\ \Pi' \vDash x = y}{ \Pi : \Sigma \vdash \Pi' : \Sigma' \triangleright \mathcal{C} } \inferrule*[lab=Drop/Pt,right={\rm $d$ not free in $\Pi':\Sigma'$}]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n] \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \true \triangleright \mathcal{C}} \inferrule[Drop/Ls]{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \ls(R,x,y) \triangleright \mathcal{C}}{ \Pi : \Sigma \vdash \Pi' : \Sigma' * \true \triangleright \mathcal{C} } \end{mathpar} \caption{Constraint generation for entailment rules} \end{figure*} \begin{figure*} \begin{mathpar} \inferrule[Assign]{\Pi : \Sigma \vdash \Pi' \land R(\vec{x}) : \Sigma' \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{x := E}{\Pi' : \Sigma'[x'/x]} \triangleright \mathcal{C}; R'(\vec{x}[x'/x]) \leftarrow R(\vec{x}) \land x' = E} \inferrule[Assume]{\Pi : \Sigma \vdash \Pi' \land R(\vec{x}) : \Sigma' \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{assume($\phi$)}{\Pi' \land R'(\vec{x}) : \Sigma'} \triangleright \mathcal{C} ; R'(\vec{x}) \leftarrow R(\vec{x}) \land \phi} \inferrule[Arith-Store]{\Pi : \Sigma \vdash \Pi' \land R(\vec{x}) : \Sigma' * x \mapsto [d,n]\triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{x->D := E}{\Pi' \land R'(\vec{x}) : \Sigma'[d'/d]} \triangleright \mathcal{C}; R'(\vec{x}[d'/d]) \leftarrow R(\vec{x}) \land d' = E} \inferrule[Arith-Load]{\Pi : \Sigma \vdash \Pi' \land R(\vec{x}) : \Sigma' * x \mapsto [d,n] \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{y := x->D}{\Pi' \land R'(\vec{x}) : \Sigma'[y'/y]} \triangleright \mathcal{C}; R'(\vec{x}[d/y]) \leftarrow R(\vec{x})} \inferrule[Heap-Store]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n] \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{x->D := E}{\Pi'[d'/d] \land d=E[d'/d] : \Sigma'[d'/d]} \triangleright \mathcal{C}} \inferrule[Heap-Load]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n] \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{y := x->D}{\Pi'[y'/y] \land y = d : \Sigma'[y'/y]} \triangleright \mathcal{C}} \inferrule*[lab=Alloc,right={\rm $d,n$ free in $\Pi':\Sigma'$}]{\Pi : \Sigma \vdash \Pi' : \Sigma' \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{x := new list}{\Pi'[x'/x] : \Sigma'[x'/x] * x \mapsto [d,n]} \triangleright \mathcal{C}} \inferrule[Free]{\Pi : \Sigma \vdash \Pi' : \Sigma' * x \mapsto [d,n] \triangleright \mathcal{C}}{\hoare{\Pi : \Sigma}{free(x)}{\Pi' : \Sigma'} \triangleright \mathcal{C}} \end{mathpar} \caption{Constraint generation for execution rules} \end{figure*} \section{Characterization of the precondition operation} \begin{definition} Let $(s,h)$ be a state. The \emph{underlying graph} of $(s,h)$ is a triple $G_{s,h} = (V,E,\lambda)$ where $V = \dom(h) \cup \{ \ell' : \exists \ell \in \dom(h). \exists d \in \mathbb{Z}. h(\ell) = [d,\ell']\}$ is a set of vertices, $E \subseteq V \times V$ is a set of edges defined by \[ (v,v') \in E \text{ iff there exists some } d \text{ such that } h(v) = [d,v'] \] and $\lambda : \textsf{HVar} \rightarrow V$ is defined by \[ \lambda(x) = s(x) \] \end{definition} \begin{definition} Let $(\exists X)(\Pi : \Sigma)$ be a separation logic formula with $\Sigma = \Sigma_1 * ... * \Sigma_n$, and such that each $\Sigma_i$ is a points-to formula. The \emph{underlying graph} of $(\exists X)(\Pi : \Sigma)$ is a triple $G_{(\exists X)(\Pi : \Sigma)} = (V,E,\lambda)$ where the set of vertices $V$ is the set of variables appearing to the left or right of a points-to predicate (quotiented by equivalence), and $E \subseteq V \times V$ is a set of edges defined by \[ (v,v') \in E \text{ iff there exists some } i,d \text{ such that } \Sigma_i = v \mapsto [d,v'] \] and $\lambda : \textsf{HVar} \rightarrow V$ is defined by setting $\lambda(x) = y$, where $y$ is the representative of the equivalence class of $x$. \end{definition} \begin{definition}[Subdivision] Let $G = (V,E,\lambda)$. For an edge $(v,v') \in E$, the \emph{$(v,v')$-subdivision} of $G$ is $G^{(v,v')} = (V \cup \{u\}, (E \setminus \{(v,v')\}) \cup \{(v,u), (u,v')\}, \lambda)$. A graph $G'$ is a \emph{subdivision} of $G$ if it is the result of a sequence of edge subdivisions. \end{definition} \begin{definition}[Homeomorphism] Let $G = (V,E,\lambda)$ and $G' = (V',E',\lambda')$. $G$ and $G'$ are \emph{homeomorphic} if there exists a subdivision $\overline{G}$ of $G$ and $\overline{G'}$ of $G'$ such that $\overline{G}$ and $\overline{G'}$ are isomorphic. \end{definition} \begin{definition}[Topological entailment] Let $S$ be a separation logic formula without list-segment predicates, and let $P$, $P'$ be (arbitrary) separation logic formulae. We write $P \vDash_S P'$ if for all $(s,h)$ such that $(s,h) \models P$ and the underlying graph of $(s,h)$ is homeomorphic to the underlying graph of $S$, we have $(s,h) \models P'$. \end{definition} \begin{proposition} Let $P$, $P'$ be separation logic formulae. If $P \vDash P'$, then $P \vDash_S P'$ for any $S$. \end{proposition} \begin{conjecture} For any $S,S',I',c$, the precondition rules compute a formula $I$ such that \begin{enumerate} \item $S \models I$ \item $\hoare{I}{c}{I'}$ \end{enumerate} and for any $J$ such that the above two conditions hold, we have \[ J \vDash_S I \] \end{conjecture} \title{Spatial Interpolants} \author{Aws Albargouthi\inst{1} \and Josh Berdine\inst{2} \and Byron Cook\inst{3} \and Zachary Kincaid\inst{4}} \institute{University of Wisconsin-Madison$^1$, Microsoft Research$^2$, University College London$^3$, University of Toronto$^4$} \maketitle \begin{abstract} We propose \alg, a new technique for proving properties of heap-manipulating programs that marries (1) a new \emph{separation logic--based} analysis for heap reasoning with (2) an \emph{interpolation-based} technique for refining heap-shape invariants with data invariants. \alg is \emph{property directed}, \emph{precise}, and produces counterexample traces when a property does not hold. Using the novel notion of \emph{spatial interpolants modulo theories}, \alg can infer complex invariants over general recursive predicates, e.g., of the form \emph{all elements in a linked list are even} or a \emph{binary tree is sorted}. Furthermore, we treat interpolation as a black box, which gives us the freedom to encode data manipulation in any suitable theory for a given program (e.g., bit vectors, arrays, or linear arithmetic), so that our technique immediately benefits from any future advances in SMT solving and interpolation. \end{abstract} \input{intro2} \input{example} \input{prelims} \input{spint} \input{snint} \input{bounded_abduction} \input{eval} \input{related} \bibliographystyle{splncs03} \bibliography{paper} \iflong \clearpage \appendix \newpage \input{alg} \input{appendix} \fi \usepackage{amsmath,amssymb,stmaryrd} \usepackage{array} \usepackage{mathpartir} \usepackage{listings} \usepackage[usenames,dvipsnames]{xcolor} \usepackage{graphicx} \usepackage{latexsym} \usepackage{dsfont} \usepackage[sort&compress,sectionbib,numbers,square]{natbib} \usepackage{fix-cm} \renewcommand{\bibfont}{\fontsize{8}{10}\selectfont} \let\llncssubparagraph\subparagraph \let\subparagraph\paragraph \usepackage[small,compact]{titlesec} \let\subparagraph\llncssubparagraph \usepackage[noend]{algpseudocode} \usepackage{algorithm} \usepackage{xparse} \newcommand{\figsep}[1]{ \null\hfill\rule[0.5ex]{2cm}{0.75pt} \small{{\textbf{\emph{#1}}}} \rule[0.5ex]{2cm}{0.75pt}\hfill\null\par } \usepackage{xspace} \usepackage{url} \usepackage{wrapfig} \usepackage{paralist} \usepackage{setspace} \renewcommand{\triangleright}{\blacktriangleright} \newcommand{\impact}{\textsc{Impact}\xspace} \newcommand{\ib}[1]{\emph{\textbf{#1}}} \newcommand{\lda}[1]{(\lambda #1)} \newcommand{\ls}{\textsf{ls}} \newcommand{\bt}{\textsf{bt}} \newcommand{\nil}{\textsf{null}} \newcommand{\true}{\textsf{true}} \newcommand{\etrue}{\emph{true}} \newcommand{\ltrue}{\emph{true}} \newcommand{\emp}{\textsf{emp}} \newcommand{\hoare}[3]{\{#1\}\;\texttt{#2}\;\{#3\}} \newcommand{\sem}[1]{\llbracket #1 \rrbracket} \newcommand{\tuple}[1]{\langle #1 \rangle} \renewcommand{\phi}{\varphi} \newcommand{\sub}{\textsf{sub}} \newcommand{\dom}{\text{dom}} \newcommand{\pre}{\textsf{pre}} \newcommand{\domain}{\mathds{D}} \newcommand{\Cmd}{\textsf{Cmd}} \newcommand{\logic}{\textsf{RSep}\xspace} \newcommand{\sep}{\textsf{Sep}\xspace} \newcommand{\Atom}{H} \newcommand{\rec}{\textsf{Rec}} \newcommand{\prog}{\mathcal{P}} \newcommand{\hvar}{\textsf{HVar}} \newcommand{\rvar}{\textsf{RVar}} \newcommand{\Vars}{\textsf{Var}} \newcommand{\dvar}{\textsf{DVar}} \newcommand{\aterm}{\textsf{DTerm}\xspace} \newcommand{\hterm}{\textsf{HTerm}\xspace} \newcommand{\dform}{\textsf{DFormula}\xspace} \newcommand{\pred}{\textsf{RPred}\xspace} \newcommand{\arity}{\textsf{arity}\xspace} \newcommand{\rarity}{\textsf{rArity}\xspace} \newcommand{\mproof}{\textsf{HExec}} \newcommand{\fent}{\textsf{F}} \newcommand{\geta}{\textsf{aliases}} \newcommand{\unalias}{\textsf{unaliased}} \newcommand{\getl}{\textsf{getList}} \newcommand{\geth}{\textsf{heapletAt}} \newcommand{\abst}{\textsc{ListFromTrue}\xspace} \newcommand{\absc}{\textsc{ListConcat}\xspace} \newcommand{\arule}{\textsc{ListIntro}\xspace} \newcommand{\alg}{\textsc{SplInter}\xspace} \newcommand{\theory}{\mathcal{T}} \newcommand{\interp}[3]{\textsf{itp}(#1, #2, #3)} \newcommand{\post}{\textsf{exec}} \newcommand{\refine}{\textsf{Refine}} \newcommand{\ppath}{\textsf{ProvePath}} \newcommand{\conj}{\textsf{Conj}\xspace} \newcommand{\cproof}{\textsf{IsProof}} \newcommand{\false}{\emph{false}} \newcommand{\bbox}[1]{\fbox{\color{blue}#1}} \newcommand{\qex}[2]{(\exists #1.\;#2)} \newcommand{\qs}[3]{\qex{#1}{#2 : #3}} \newcommand{\ul}[1]{\underline{#1}} \newcommand{\ol}[1]{\overline{#1}} \newcommand{\abduce}[4]{#1 \vdash \qex{#2}{#3 * [A]} \vdash #4} \newcommand{\cjudge}[2]{#2 \ \triangleright\ #1} \newcommand{\eoe}{ \hspace*{\fill}\ensuremath{\lrcorner} } \usepackage[scaled=0.80]{DejaVuSansMono} \lstdefinestyle{customc}{ belowcaptionskip=1\baselineskip, breaklines=true, xleftmargin=\parindent, language=C, showstringspaces=false, basicstyle=\footnotesize\ttfamily, keywordstyle=\ttfamily\color{blue}, commentstyle=\itshape\color{black}, identifierstyle=\ttfamily\color{black}, stringstyle=\color{orange}, morekeywords={assert,assume,node} } \lstset{escapechar=@,style=customc} \newif\iflong \longtrue \newcommand{\ifshort}[1]{\iflong\else#1\fi} \newcommand{\spacehack}[1]{\ifshort{#1}} \newcommand{\vspacehack}[1]{\spacehack{\vspace*{#1}}} \newcommand{\vsh}[1]{\spacehack{\vspace*{#1}}} \newcommand{\lb}{\ifshort{\linebreak}\xspace} \newcommand{\pb}{\ifshort{\pagebreak}\xspace} \newcommand{\npb}{\ifshort{~\nopagebreak}\xspace} \newcommand{\shortenpar}{ \ifshort{\looseness=-1} } \newcommand{\setshort}{ \longfalse \AtBeginDocument{ \setlength{\textfloatsep}{8pt plus 2pt minus 8pt} \setlength\abovedisplayskip{5pt plus 2pt minus 2pt} \setlength\belowdisplayskip{5pt plus 2pt minus 2pt} \setlength\abovedisplayshortskip{2pt plus 2pt minus 2pt} \setlength\belowdisplayshortskip{2pt plus 2pt minus 2pt} \renewcommand{\MathparLineskip}{\lineskiplimit=1.2em\lineskip=0.9em plus 0.15em} } } \renewcommand{\leq}{\leqslant} \renewcommand{\geq}{\geqslant} \definecolor{dgray}{gray}{0.35} \section{Preliminaries} \subsection{Separation Logic} We define \logic, a fragment of separation logic formulas featuring points-to predicates and general recursive predicates refined by theory propositions. Fig.~ defines the syntax of \logic formulas. In comparison with the standard list fragment used in separation logic analyses (e.g.,~), the differentiating features of \logic are: (1) General \emph{recursive predicates}, for describing unbounded pointer structures like lists, trees, etc. (2) Recursive predicates are augmented with a vector of \emph{refinements}, which are used to constrain the data values appearing on the data structure defined by the predicate, detailed below. (3) Each heap cell (points-to predicate), $E \mapsto [\vec{A},\vec{E}]$, is a \emph{record} consisting of \emph{data} fields (a vector $\vec{A}$ of \aterm) followed by \emph{heap} fields (a vector $\vec{E}$ of \hterm). (Notationally, we will use $d_i$ to refer to the $i$th element of the vector $\vec{d}$, and $\vec{d}[t/d_i]$ to refer to the vector $\vec{d}$ with the $i$th element modified to $t$.) (4) \textsf{Pure} formulas contain heap and first-order data constraints. Our definition is (implicitly) parameterized by a first-order theory $\theory$. $\dvar$ denotes the set of theory variables, which we assume to be disjoint from $\hvar$ (the set of heap variables). $\textsf{DTerm}$ and $\textsf{DFormula}$ denote the sets of theory terms and formulas, and we assume that heap variables do not appear in theory terms. For an \logic formula $P$, $\Vars(P)$ denotes its free (data and heap) variables. We treat a $\textsf{Spatial}$ formula $\Sigma$ as a multiset of heaplets, and consider formulas to be equal when they are equal as multisets. For \logic formulas $P = \qs{X_P}{\Pi_P}{\Sigma_P}$ and $Q = \qs{X_Q}{\Pi_Q}{\Sigma_Q}$, we write $P*Q$ to denote the \logic formula \shortenpar \[ P * Q = \qs{X_P \cup X_Q}{\Pi_P \land \Pi_Q}{\Sigma_P * \Sigma_Q} \] assuming that $X_P$ is disjoint from $\Vars(Q)$ and $X_Q$ is disjoint from $\Vars(P)$ (if not, then $X_P$ and $X_Q$ are first suitably renamed). For a set of variables $X$, we write $\qex{X}{P}$ to denote the \logic formula \[ \qex{X}{P} = \qs{X \cup X_P}{\Pi_P}{\Sigma_P} \] \paragraph{Recursive predicates} Each recursive predicate $Z \in \pred$ is associated with a definition that describes how the predicate is unfolded. Before we formalize these definitions, we will give some examples. \shortenpar The definition of the list segment predicate from Sec.~ is: \[\begin{split} \ls(R,x,y) \equiv {}& (x = y : \emp) \lor {}\\ & \qs{d,n'} {x \neq y \land R(d)} {x \mapsto [d,n'] * \ls(R,n',y)} \end{split}\] In the above, $R$ is a \emph{refinement variable}, which may be instantiated to a concrete refinement $\theta \in \textsf{Refinement}$. For example, $\ls(\lda{a . a \geq 0},x,y)$ indicates that there is a list from $x$ to $y$ where every element of the list is at least $0$. A refined binary tree predicate is a more complicated example: \begin{align*} \textsf{bt}(Q,L,R,x) =&\; (x = \nil : \emp)\\ \lor &\; \qex{d,l,r}{Q(d) : x \mapsto [d,l,r] \\ &\hspace*{0.5cm} * \textsf{bt}((\lambda a. Q(a) \land L(d,a)), L, R, l) \\ &\hspace*{0.5cm} * \textsf{bt}((\lambda a. Q(a) \land R(d,a)), L, R, r)} \end{align*} This predicate has three refinement variables: a unary refinement $Q$ (which must be satisfied by every node in the tree), a binary refinement $L$ (which is a relation that must hold between every node and its descendants to the left), and a binary refinement $R$ (which is a relation that must hold between every node and its descendants to the right). For example, \[ \textsf{bt}((\lambda a. \ltrue), (\lambda a, b. a \geq b), (\lambda a, b. a \leq b), x) \] indicates that $x$ is the root of a \emph{binary search tree}, and \[ \textsf{bt}((\lambda a. a \geq 0), (\lambda a, b. a \leq b), (\lambda a, b. a \leq b), x) \] indicates that $x$ is the root of a \emph{binary min-heap} with non-negative elements. To formalize these definitions, we first define \emph{refinement terms} and \emph{refined formulas}: a refinement term $\tau$ is either (1) a refinement variable $R$ or (2) an abstraction $(\lambda a_1,\dotsc,a_n. \Phi)$, where $\Phi$ is a refined formula. A \emph{refined formula} is a conjunction where each conjunct is either a data formula (\textsf{DFormula}) or the application $\tau(\vec{A})$ of a refinement term to a vector of data terms (\textsf{DTerm}). A \emph{predicate definition} has the form \[Z(\vec{R}, \vec{x}) \equiv \qex{X_1}{\Pi_1 \land \Phi_1 : \Sigma_1} \lor \dotsb \lor \qex{X_n}{\Pi_n \land \Phi_n : \Sigma_n}\] where $\vec{R}$ is a vector of refinement variables, $\vec{x}$ is a vector of heap variables, and where refinement terms may appear as refinements in the spatial formulas $\Sigma_i$. We refer to the disjuncts of the above formula as the \emph{cases} for $Z$, and define $\mathit{cases}(Z(\vec{R},\vec{x}))$ to be the set of cases of $Z$. $\vec{R}$ and $\vec{x}$ are bound in $\mathit{cases}(Z(\vec{R},\vec{x}))$, and we will assume that predicate definitions are closed, that is, for each case of $Z$, the free refinement variables belong to $\vec{R}$, the free heap variables belong to $\vec{x}$, and there are no free data variables. We also assume that they are well-typed in the sense that each refinement term $\tau$ is associated with an arity, and whenever $\tau(\vec{A})$ appears in a definition, the length of $\vec{A}$ is the arity of $\tau$. \shortenpar \paragraph{Semantics} The semantics of our logic, defined by a satisfaction relation $s,h \models Q$, is essentially standard. Each predicate $Z \in \pred$ is defined to be the least solution\footnote{Our definition does not preclude ill-founded predicates; such predicates are simply unsatisfiable, and do not affect the technical development in the rest of the paper.} to the following equivalence: \[ s,h \models Z(\vec{\theta},\vec{E}) \iff \exists P \in \mathit{cases}(Z(\vec{R},\vec{x})).\ s,h \models P[\vec{\theta}/\vec{R} , \vec{E}/\vec{x} ] \] Note that when substituting a $\lambda$-abstraction for a refinement variable, we implicitly $\beta$-reduce resulting applications. For example, $R(b)[(\lambda a. a \geq 0)/R] = b \geq 0$. Semantic entailment is denoted by $P \models Q$, and provable entailment by $P \vdash Q$. When referring to a proof that $P \vdash Q$, we will mean a sequent calculus proof. \subsection{Programs} A program $\prog$ is a tuple $\tuple{V,E,v_{\rm i},v_{\rm e}}$, where \begin{compactitem} \item $V$ is a set of control locations, with a distinguished \emph{entry} node $v_{\rm i} \in V$ and \emph{error} (exit) node $v_{\rm e} \in V$, and \item $E \subseteq V \times V$ is a set of directed edges, where each $e \in E$ is associated with a program command $e^{\rm c}$. \end{compactitem} We impose the restriction that all nodes $V \setminus \{v_{\rm i}\}$ are reachable from $v_{\rm i}$ via $E$, and all nodes can reach $v_{\rm e}$. The syntax for program commands appears below. Note that the allocation command creates a record with $n$ data fields, $D_1,\ldots,D_n$, and $m$ heap fields, $N_1,\ldots,N_m$. To access the $i$th data field of a record pointed to by \texttt{x}, we use \texttt{x->D$_i$} (and similarly for heap fields). We assume that programs are well-typed, but not necessarily memory safe. ~\\~\\ \noindent {\small \begin{minipage}{0.3\linewidth} \textbf{Assignment}: \texttt{x := \AE}\\ \textbf{Heap store}: \texttt{x->N$_i$ := E}\\ \textbf{Heap load}: \texttt{y := x->N$_i$} \end{minipage} \hfill \begin{minipage}{0.3\linewidth} \textbf{Assumption}: \texttt{assume($\Pi$)}\\ \textbf{Data store}: \texttt{x->D$_i$ := A}\\ \textbf{Data load}: \texttt{y := x->D$_i$} \end{minipage} \hfill \begin{minipage}{0.33\linewidth} \textbf{Allocation}: \texttt{x := new($n,m$)}\\ \textbf{Disposal}: \texttt{free(x)}\\ \end{minipage}} \mbox{}\\\noindent As is standard, we compile assert commands to reachability of $v_{\rm e}$. \section{Related Work} \paragraph{Abstraction Refinement for the Heap} To the best of our knowledge, the work of Botincan et al.~ is the only separation logic shape analysis that employs a form of abstraction refinement. It starts with a family of separation logic domains of increasing precision, and uses spurious counterexample traces (reported by forward fixed-point computation) to pick a more precise domain to restart the analysis and (possibly) eliminate the counterexample. Limitations of this technique include: (1) The precision of the analysis is contingent on the set of abstract domains it is started with. (2) The refinement strategy (in contrast to \alg) does not guarantee progress (it may explore the same path repeatedly), and may report false positives. On the other hand, given a program path, \alg is guaranteed to find a proof for the path or correctly declare it an unsafe execution. (3) Finally, it is unclear whether refinement with a powerful theory like linear arithmetic can be encoded in such a framework, e.g., as a set of domains with increasingly more arithmetic predicates. Podelski and Wies~ propose an abstraction refinement algorithm for a shape-analysis domain with a logic-based view of three-valued shape analysis (specifically, first-order logic plus transitive closure). Spurious counterexamples are used to either refine the set of predicates used in the analysis, or refine an imprecise abstract transformer. The approach is used to verify specifications given by the user as first-order logic formulas. A limitation of the approach is that refinement is syntactic, and if an important recursive predicate (e.g., there is a list from $x$ to $\nil$) is not explicitly supplied in the specification, it cannot be inferred automatically. Furthermore, abstract post computation can be expensive, as the abstract domain uses quantified predicates. Additionally, the analysis assumes a memory safe program to start, whereas, in \alg, we construct a memory safety proof as part of the invariant, enabling us to detect unsafe memory operations that lead to undefined program behavior. Beyer et al.~ propose using shape analysis information on demand to augment numerical predicate abstraction. They use shape analysis as a backup analysis when failing to prove a given path safe without tracking the heap, and incrementally refines TVLA's~ three-valued shape analysis~ to track more heap information as required. As with~,~ makes an \emph{a priori} assumption of memory safety and requires an expensive abstract post operator. Finally, Manevich et al.~ give a theoretical treatment of counterexample-driven refinement in power set (e.g., shape) abstract domains. \paragraph{Combined Shape and Data Analyses} The work of Magill et al.~ infers shape and numerical invariants, and is the most closely related to ours. First, a separation logic analysis is used to construct a memory safety proof of the whole program. This proof is then \emph{instrumented} by adding additional user-defined integer parameters to the recursive predicates appearing in the proof (with corresponding user-defined interpretations). A numerical program is generated from this instrumented proof and checked using an off-the-shelf verification tool, which need not reason about the heap. Our technique and 's are similar in that we both decorate separation logic proofs with additional information: in , the extra information is instrumentation variables; in this paper, the extra information is refinement predicates. Neither of these techniques properly subsumes the other, and we believe that they may be profitably combined. An important difference is that we synthesize data refinements automatically from program paths, whereas uses a fixed (though user-definable) abstraction. A number of papers have proposed abstract domains for shape and data invariants. Chang and Rival~ propose a separation logic--based abstract domain that is parameterized by programmer-supplied \emph{invariant checkers} (recursive predicates) and a data domain for reasoning about contents of these structures. McCloskey et al.~ also proposed a combination of heap and numeric abstract domains, this time using 3-valued structures for the heap. While the approaches to combining shape and data information are significantly different, an advantage of our method is that it does not lose precision due to limitations in the abstract domain, widening, and join. Bouajjani et al.~ propose an abstract domain for list manipulating programs that is parameterized by a data domain. They show that by varying the data domain, one can infer invariants about list sizes, sum of elements, etc. Quantified data automata (QDA)~ have been proposed as an abstract domain for representing list invariants where the data in a list is described by a regular language. In~, invariants over QDA have been synthesized using language learning techniques from concrete program executions. Expressive logics have also been proposed for reasoning about heap and data~, but have thus far been only used for invariant checking, not invariant synthesis. A number of decision procedures for combinations of the singly-linked-list fragment of separation logic with SMT theories have recently been proposed~. \paragraph{Path-based Verification} A number of works proposed path-based algorithms for verification. Our work builds on McMillan's \impact technique~ and extends it to heap/data reasoning. Earlier work~ used interpolants to compute predicates from spurious paths in a CEGAR loop. Beyer et al.~ proposed \emph{path invariants}, where infeasible paths induce program slices that are proved correct, and from which predicates are mined for full program verification. Heizmann et al.~ presented a technique that uses interpolants to compute path proofs and generalize a path into a visibly push-down language of correct paths. In comparison with \alg, all of these techniques are restricted to first-order invariants. \shortenpar Our work is similar to that of Itzhaky et al.~, in the sense that we both generalize from bounded unrollings of the program to compute ingredients of a proof. However, they compute proofs in a fragment of first-order logic that can only express linked lists and has not yet been extended to combined heap and data properties. \section{Spatial Interpolation Modulo Theories} \newcolumntype{L}[1]{>{\raggedright\let\newline\\\arraybackslash\hspace{0pt}}m{#1}} \newcolumntype{C}[1]{>{\centering\let\newline\\\arraybackslash\hspace{0pt}}m{#1}} \newcolumntype{R}[1]{>{\raggedleft\let\newline\\\arraybackslash\hspace{0pt}}m{#1}} We now consider the problem of \emph{refining} (or \emph{strengthening}) a given separation logic proof of memory safety with information about (non-spatial) data. This refinement procedure results in a proof of a conclusion stronger than can be proved by reasoning about the heap alone. In view of our example from Fig.~, this section addresses how to derive the third sequence (Spatial Interpolants Modulo Theories) from the second (Spatial Interpolants). The input to our spatial interpolation modulo theories procedure is a path $\pi$, a separation logic (\sep) proof $\zeta$ of the triple $\hoare{\textit{true}:\emp}{$\pi$}{\textit{true}:\true}$ (i.e., a memory safety proof for $\pi$), and a postcondition $\phi$. The goal is to transform $\zeta$ into an \logic proof of the triple $\hoare{\textit{true}:\emp}{$\pi$}{\phi:\true}$. The high-level operation of our procedure is as follows. First, we traverse the memory safety proof $\zeta$ and build (1) a corresponding \emph{refined} proof $\zeta'$ where refinements may contain second-order variables, and (2) a constraint system $\mathcal{C}$ which encodes logical dependencies between the second-order variables. We then attempt to find a solution to $\mathcal{C}$, which is an assignment of data formulas to the second-order variables such that all constraints are satisfied. If we are successful, we use the solution to instantiate the second-order variables in $\zeta'$, which yields a valid \logic proof of the triple $\hoare{\textit{true}:\emp}{$\pi$}{\phi:\true}$. \shortenpar \paragraph{Horn Clauses} The constraint system produced by our procedure is a recursion-free set of Horn clauses, which can be solved efficiently using existing first-order interpolation techniques (see~ for a detailed survey). Following , we define a \emph{query} to be an application $Q(\vec{a})$ of a second-order variable $Q$ to a vector of (data) variables, and define an \emph{atom} to be either a data formula $\phi \in \textsf{DFormula}$ or a query $Q(\vec{a})$. A \emph{Horn clause} is of the form \iflong\[\else $\fi h \gets b_1 \land \cdots \land b_N \iflong\]\else$ \fi where each of $h,b_1,\dotsc,b_N$ is an atom. In our constraint generation rules, it will be convenient to use a more general form which can be translated to Horn clauses: we will allow constraints of the form \iflong\[\else $\fi h_1 \land \cdots \land h_M \gets b_1 \land \cdots \land b_N \iflong\]\else$ \fi (shorthand for the set of Horn clauses $\{ h_i \gets b_1 \land \cdots \land b_N \}_{1 \leq i \leq M}$) and we will allow queries to be of the form $Q(\vec{A})$ (i.e., take arbitrary data terms as arguments rather than variables). If $\mathcal{C}$ and $\mathcal{C}'$ are sets of constraints, we will use $\mathcal{C};\mathcal{C}'$ to denote their union. A \emph{solution} to a system of Horn clauses $\mathcal{C}$ is a map $\sigma$ that assigns each second-order variable $Q$ of arity $k$ a $\textsf{DFormula}$ $Q^\sigma$ with free variables drawn from $\vec{\nu} = \tuple{\nu_1,\dotsc,\nu_k}$ such that for each clause \iflong\[\else $\fi h \gets b_1 \land \cdots \land b_N \iflong\]\else$ \fi in $\mathcal{C}$ the implication \iflong\[\else $\fi \forall A. (h^\sigma \Leftarrow (\exists B. b_1^\sigma \land \cdots \land b_N^\sigma)) \iflong\]\else$ \fi holds, where $A$ is the set of free variables in $h$ and $B$ is the set of variables free in some $b_i$ but not in $h$. In the above, for any data formula $\phi$, $\phi^\sigma$ is defined to be $\phi$, and for any query $Q(\vec{a})$, $Q(\vec{a})^\sigma$ is defined to be $Q^\sigma[a_1/\nu_1,\dotsc,a_k/\nu_k]$ (where $k$ is the arity of $Q$). \shortenpar \paragraph{Constraint Generation Calculus} We will present our algorithm for spatial interpolation modulo theories as a calculus whose inference rules mirror the ones of separation logic. The calculus makes use of the same syntax used in recursive predicate definitions in Sec.~. We use $\tau$ to denote a \emph{refinement term} and $\Phi$ to denote a \emph{refined formula}. The calculus has two types of judgements. An \emph{entailment judgement} is of the form \[ \cjudge{\qex{X}{\Pi \land \Phi : \Sigma} \vdash \qex{X'}{\Pi' \land \Phi' : \Sigma'}}{\mathcal{C}} \] where $\Pi,\Pi'$ are equational pure assertions over heap terms, $\Sigma,\Sigma'$ are refined spatial assertions, $\Phi$, $\Phi'$ are refined formulas, and $\mathcal{C}$ is a recursion-free set of Horn clauses. Such an entailment judgement should be read as ``for any solution $\sigma$ to the set of constraints $\mathcal{C}$, $\qex{X}{\Pi \land \Phi^\sigma : \Sigma^\sigma}$ entails $\qex{X'}{\Pi' \land \Phi'^\sigma : \Sigma'^\sigma}$,'' where $\Phi^\sigma$ is $\Phi$ with all second order variables replaced by their data formula assignments in $\sigma$ (and similarly for $\Sigma^\sigma$). Similarly, an \emph{execution judgement} is of the form \[ \cjudge{\hoare{\qex{X}{\Pi \land \Phi : \Sigma}}{$\pi$}{\qex{X'}{\Pi' \land \Phi' : \Sigma'}}}{\mathcal{C}} \] where $\pi$ is a path and $X,X',\Pi,\Pi',\Phi,\Phi',\Sigma,\Sigma'$, and $\mathcal{C}$ are as above. Such an execution judgement should be read as ``for any solution $\sigma$ to the set of constraints $\mathcal{C}$,\shortenpar \[\hoare{\qex{X}{\Pi \land \Phi^\sigma : \Sigma^\sigma}}{\ensuremath \pi}{\qex{X'}{\Pi' \land \Phi'^\sigma : \Sigma'^\sigma}}\] is a valid triple.'' Let $\pi$ be a path, let $\zeta$ be a separation logic proof of the triple $\hoare{\textit{true}:\emp}{$\pi$}{\textit{true}:\true}$ (i.e., a memory safety proof for $\pi$), and let $\phi \in \textsf{DFormula}$ be a postcondition. Given these inputs, our algorithm operates as follows. We use $\vec{v}$ to denote a vector of all data-typed program variables. The triple is \emph{rewritten with refinements} by letting $R$ and $R'$ be fresh second-order variables of arity $|\vec{v}|$ and conjoining $R(\vec{v})$ and $R'(\vec{v})$ to the pre and post. By recursing on $\zeta$, at each step applying the appropriate rule from our calculus in Fig.~, we derive a judgement\shortenpar \begin{mathpar} \text{\footnotesize{\( \inferrule{ \zeta' }{ \cjudge{\hoare{\textit{true} \land R(\vec{v}):\true}{$\pi$}{\textit{true} \land R'(\vec{v}):\true}}{\mathcal{C}} } \)}} \end{mathpar} and then compute a solution $\sigma$ to the constraint system \[\mathcal{C};\hspace*{0.25cm} R(\vec{v}) \gets \textit{true};\hspace*{0.25cm} \phi \gets R'(\vec{v})\] (if one exists). The algorithm then returns $\zeta'^\sigma$, the proof obtained by applying the substitution $\sigma$ to $\zeta'$. Intuitively, our algorithm operates by recursing on a separation logic proof, introducing refinements into formulas on the way down, and building a system of constraints on the way up. Each inference rule in the calculus encodes both the downwards and upwards step of this algorithm. For example, consider the {\sc Fold} rule of our calculus: we will illustrate the intended reading of this rule with a concrete example. Suppose that the input to the algorithm is a derivation of the following form: \begin{mathpar} \text{\footnotesize{\( \inferrule*[right=Fold]{ \inferrule*{ \zeta_0 }{ x \mapsto [a,\nil] \vdash \qex{b,y}{x \mapsto [b,y] * \ls(y,\nil)} } }{ Q(i) : x \mapsto [a,\nil] \vdash R(i) : \ls(\lda{a. S(x,a)}, x,\nil) } \)}} \end{mathpar} (i.e., a derivation where the last inference rule is an application of {\sc Fold}, and the conclusion has already been rewritten with refinements). We introduce refinements in the premise and recurse on the following derivation: \begin{mathpar} \text{\footnotesize{\( \inferrule*{ \zeta_0 }{ Q(i) : x \mapsto [a,\nil] \vdash \\\qex{b,y}{R(i) \land S(i,b) : x \mapsto [b,y] * \ls(\lda{a. S(x,a)},y,\nil)} } \)}} \end{mathpar} The result of this recursive call is a refined derivation $\zeta_0'$ as well as a constraint system $\mathcal{C}$. We then return both (1) the refined derivation obtained by catenating the conclusion of the {\sc Fold} rule onto $\zeta_0'$ and (2) the constraint system $\mathcal{C}$. A crucial point of our algorithm is hidden inside the hat notation in Fig.~ (e.g, $\widehat{O}$ in {\sc Sequence}): this notation is used to denote the introduction of fresh second-order variables. For many of the inference rules (such as {\sc Fold}), the refinements which appear in the premises follow fairly directly from the refinements which appear in the conclusion. However, in some rules entirely new formulas appear in the premises which do not appear in the conclusion (e.g., in the {\sc Sequence} rule in Fig.~, the intermediate assertion $\widehat{O}$ is an arbitrary formula which has no obvious relationship to the precondition $P$ or the postcondition $Q$). We refine such formula $O$ by introducing a fresh second-order variable for the pure assertion and for each refinement term that appears in $O$. The following offers a concrete example. \begin{example} Consider the trace $\pi$ in Fig.~. Suppose that we are given a memory safety proof for $\pi$ which ends in an application of the {\sc Sequence} rule: \begin{mathpar} \text{\footnotesize{\( \inferrule*[right=Sequence]{ \hoare{\textit{true} : \emp}{$\pi_0$}{\textit{true} : \ls(x,\nil)}\\ \hoare{\textit{true} : \ls(x,\nil)}{$\pi_1$}{\qex{b,y}{\textit{true} : x \mapsto [b,y]}} }{ \hoare{Q(i) : \emp}{$\pi_0;\pi_1$}{\qex{b,y}{R(i,b) : x \mapsto [b,y]}} } \)}} \end{mathpar} where $\pi$ is decomposed as $\pi_0;\pi_1$, $\pi_0$ is the path from 1 to 3, and $\pi_1$ is the path from 3 to 4. Let $O = \textit{true} : \ls(x,\nil)$ denote the intermediate assertion which appears in this proof. To derive $\widehat{O}$, we introduce two fresh second order variables, $S$ (with arity 1) and $T$ (with arity 2), and define $\widehat{O} = {S(i) : \ls(\lda{a. T(i,a)},x,\nil)}$. The resulting inference is as follows: \begin{mathpar} \text{\footnotesize{\( \inferrule*{ \hoare{Q(i) : \emp}{$\pi_0$}{S(i) : \ls(\lda{a.T(i,a)},x,\nil)}\\ \hoare{S(i) : \ls(\lda{a.T(i,a)},x,\nil)}{$\pi_1$}{\qex{b,y}{R(i,b) : x \mapsto [b,y]}} }{ \qquad \hoare{Q(i) : \emp}{$\pi_0;\pi_1$}{\qex{b,y}{R(i,b) : x \mapsto [b,y]}} \qquad\lrcorner } \)}} \end{mathpar} \end{example} The following example provides a simple demonstration of our constraint generation procedure: \begin{example} Recall the example in Fig.~ of Sec.~. The row of spatial interpolants in Fig.~ is a memory safety proof $\zeta$ of the program path. Fig.~ shows the refined proof $\zeta'$, which is the proof $\zeta$ with second-order variables that act as placeholders for data formulas. \textbf{\emph{For the sake of illustration, we have simplified the constraints by skipping a number of intermediate annotations in the Hoare-style proof.}} The constraint system $\mathcal{C}$ specifies the logical dependencies between the introduced second-order variables in $\zeta'$. For instance, the relation between $R_{2}$ and $R_{3}$ is specified by the Horn clause $R_{3}(i) \gets R_{2}(i) \land i = 0$, which takes into account the constraint imposed by $\texttt{assume (i == 0)}$ in the path. The Horn clause $d' \geq 0 \gets R_4(i,d')$ specifies the postcondition defined by the assertion $\texttt{assert(x->D >= 0)}$, which states that the value of the data field of the node $x$ should be $\geq 0$. \shortenpar Replacing second-order variables in $\zeta'$ with their respective solutions in $\sigma$ produces a proof that the assertion at the end of the path holds (last row of Fig.~). \eoe\end{example} \paragraph{Soundness and Completeness} The key result regarding the constraint systems produced by these judgements is that any solution to the constraints yields a valid refined proof. The formalization of the result is the following theorem. \begin{theorem}[Soundness] Suppose that $\pi$ is a path, $\zeta$ is a derivation of the judgement $\cjudge{\hoare{P}{$\pi$}{Q}}{\mathcal{C}}$, and that $\sigma$ is a solution to $\mathcal{C}$. Then $\zeta^\sigma$, the proof obtained by applying the substitution $\sigma$ to $\zeta$, is a (refined) separation logic proof of $\hoare{P^\sigma}{$\pi$}{Q^\sigma}$. \end{theorem} Another crucial result for our counterexample generation strategy is a kind of completeness theorem, which effectively states that the strongest memory safety proof always admits a refinement. \begin{theorem}[Completeness] Suppose that $\pi$ is a memory-feasible path and $\zeta$ is a derivation of the judgement $\cjudge{\hoare{R_0(\vec{v}):\emp}{$\pi$}{R_1(\vec{v}) : \true}}{\mathcal{C}}$ obtained by symbolic execution. If $\phi$ is a data formula such that $\hoare{\textit{true}:\emp}{$\pi$}{\phi : \true}$ holds, then there is a solution $\sigma$ to $\mathcal{C}$ such that $R_1^\sigma(\vec{v}) \Rightarrow \phi$. \end{theorem} \section{Spatial Interpolants} In this section, we first define the notion of spatial path interpolants, which serve as memory safety proofs of program paths. We then describe a technique for computing spatial path interpolants. This algorithm has two phases: the first is a (forwards) \emph{symbolic execution} phase, which computes the strongest memory safety proof for a path; the second is a (backwards) \emph{interpolation} phase, which weakens the proof so that it is more likely to generalize. Spatial path interpolants are bounded from below by the strongest memory safety proof, and (implicitly) from above by the weakest memory safety proof. Prior to considering the generation of inductive invariants using spatial path interpolants, consider what could be done with only one of the bounds, in general, with either a path-based approach or an iterative fixed-point computation. Without the upper bound, an interpolant or invariant could be computed using a standard forward transformer and widening. But this suffers from the usual problem of potentially widening too aggressively to prove the remainder of the path, necessitating the design of analyses which widen conservatively at the price of computing unnecessarily strong proofs. The upper bound neatly captures the information that must be preserved for the future execution to be proved safe. On the other hand, without the lower bound, an interpolant or invariant could be computed using a backward transformer (and lower widening). But this suffers from the usual problem that backward transformers in shape analysis explode, due to issues such as not knowing the aliasing relationship in the pre-state. The lower bound neatly captures such information, heavily reducing the potential for explosion. These advantages come at the price of operating over full paths from entry to error. Compared to a forwards iterative analysis, operating over full paths has the advantage of having information about the execution's past and future when weakening at each point along the path. A forwards iterative analysis, on the other hand, trades the information about the future for information about many past executions through the use of join or widening operations. The development in this section is purely spatial: we do not make use of data variables or refinements in recursive predicates. Our algorithm is thus of independent interest, outside of its context in this paper. We use \sep to refer to the fragment of \logic in which the only data formula (appearing in pure assertions and in refinements) is $\textit{true}$ (this fragment is equivalent to classical separation logic). An \logic formula $P$, in particular including those in recursive predicate definitions, determines a \sep formula $\ul{P}$ obtained by replacing all refinements (both variables and $\lambda$-abstractions) with $\lda{\vec{a}.\ltrue}$ and all $\textsf{DFormula}$s in the pure part of $P$ with $\ltrue$. Since recursive predicates, refinements, and \textsf{DFormula}s appear only positively, $\ul{P}$ is no stronger than any refinement of $P$. Since all refinements in $\sep$ are trivial, we will omit them from the syntax (e.g., we will write $Z(\vec{E})$ rather than $Z((\lambda \vec{a}. \ltrue), \vec{E})$). \subsection{Definition} We define a \emph{symbolic heap} to be a \sep formula where the spatial part is a *-conjunction of points-to heaplets and the pure part is a conjunction of pointer (dis)equalities. Given a command $\texttt{c}$ and a symbolic heap $S$, we use $\post(\texttt{c}, S)$ to denote the symbolic heap that results from symbolically executing $\texttt{c}$ starting in $S$ (the definition of $\post$ is essentially standard , and is shown in Fig.~). Given a program path $\pi = e_1,\ldots,e_n$, we obtain its strongest memory safety proof by symbolically executing $\pi$ starting from the empty heap \emp. We call this sequence of symbolic heaps the symbolic execution sequence of $\pi$, and say that a path $\pi$ is \emph{memory-feasible} if every formula in its symbolic execution sequence is consistent. The following proposition justifies calling this sequence the strongest memory safety proof. \begin{proposition} For a path $\pi$, if the symbolic execution sequence for $\pi$ is defined, then $\pi$ is memory safe. If $\pi$ is memory safe and memory-feasible, then its symbolic execution sequence is defined. \end{proposition} Recall that our strategy for proving program correctness is based on sampling and proving the correctness of several program paths (\emph{\'{a} la} {\sc Impact}~). The problem with \emph{strongest} memory safety proofs is that they do not generalize well (i.e., do not generate inductive invariants). One solution to this problem is to take advantage of property direction. Given a desired postcondition $P$ and a (memory-safe and -feasible) path $\pi$, the goal is to come up with a proof that is weaker than $\pi$'s symbolic execution sequence, but still strong enough to show that $P$ holds after executing $\pi$. Coming up with such ``weak'' proofs is how traditional path interpolation is used in {\sc Impact}. In light of this, we define \emph{spatial path interpolants} as follows: \begin{definition}[Spatial path interpolant] Let $\pi = e_1,\ldots, e_n$ be a program path with symbolic execution sequence $S_0,\ldots,S_n$, and let $P$ be a \sep formula (such that $S_n \models P$). A \emph{spatial path interpolant} for $\pi$ is a sequence $I_0,\ldots,I_n$ of \sep formulas such that \shortenpar \begin{compactitem} \item for each $i \in [0,n]$, $S_i \models I_i$; \item for each $i \in [1,n]$, $\{I_{i-1}\}\;e^{\rm c}_i\;\{I_i\}$ is a valid triple in separation logic; and \item $I_n \models P$ . \end{compactitem} \end{definition} Our algorithm for computing spatial path interpolants is a backwards propagation algorithm that employs a \emph{spatial interpolation} procedure at each backwards step. Spatial interpolants for a single command are defined as: \begin{definition}[Spatial interpolant] Given \sep formulas $S$ and $I'$ and a command $\texttt{c}$ such that $\post(\texttt{c},S) \models I'$, a \emph{spatial interpolant} (for $S$, $\texttt{c}$, and $I'$) is a \sep formula $I$ such that $S \models I$ and $\hoare{I}{c}{I'}$ is valid. \end{definition} Before describing the spatial interpolation algorithm, we briefly describe how spatial interpolation is used to compute path interpolants. Let us use $\interp{S}{\texttt{c}}{I}$ to denote a spatial interpolant for $S,\texttt{c},I$, as defined above. Let $\pi = e_1,\ldots,e_n$ be a program path and let $P$ be a \sep formula. First, symbolically execute $\pi$ to compute a sequence $S_0,\ldots,S_n$. Suppose that $S_n \vdash P$. Then we compute a sequence $I_0,\ldots,I_n$ by taking $I_n = P$ and (for $k < n$) $I_{k} = \interp{S_{k}}{e_{k+1}^{\rm c}}{I_{k+1}}$. The sequence $I_0,\ldots,I_n$ is clearly a spatial path interpolant. \shortenpar \subsection{Bounded Abduction} Our algorithm for spatial interpolation is based on an abduction procedure. Abduction refers to the inference of explanatory hypotheses from observations (in contrast to deduction, which derives conclusions from given hypotheses). The variant of abduction we employ in this paper, which we call \emph{bounded abduction}, is simultaneously a form of abductive and deductive reasoning. Seen as a variant of abduction, bounded abduction adds a constraint that the abduced hypothesis be at least weak enough to be derivable from a given hypothesis. Seen as a variant of deduction, bounded abduction adds a constraint that the deduced conclusion be at least strong enough to imply some desired conclusion. Formally, we define bounded abduction as follows: \begin{definition}[Bounded abduction] Let $L,M,R$ be \sep formulas, and let $X$ be a set of variables. A solution to the \emph{bounded abduction problem} \[ L \vdash \qex{X}{M * [\ ]} \vdash R \] is a \sep formula $A$ such that $ L \models \qex{X}{M * A} \models R$. \end{definition} Note how, in contrast to bi-abduction~ where a solution is a pair of formulas, one constrained from above and one from below, a solution to bounded abduction problems is a single formula that is simultaneously constrained from above and below. The fixed lower and upper bounds in our formulation of abduction give considerable guidance to solvers, in contrast to bi-abduction, where the bounds are part of the solution. Sec.~ presents our bounded abduction algorithm. For the remainder of this section, we will treat bounded abduction as a black box, and use $L \vdash \qex{X}{M * [A]} \vdash R$ to denote that $A$ is a solution to the bounded abduction problem. \subsection{Computing Spatial Interpolants} We now proceed to describe our algorithm for spatial interpolation. Given a command $\texttt{c}$ and \sep formulas $S$ and $I'$ such that $\post(\texttt{c},S) \vdash I'$, this algorithm must compute a \sep formula $\interp{S}{\texttt{c}}{I'}$ that satisfies the conditions of Definition~. Several examples illustrating this procedure are given in Fig.~. This algorithm is defined by cases based on the command $\texttt{c}$. We present the cases for the spatial commands; the corresponding data commands are similar. \paragraph{Allocate} Suppose \texttt{c} is \texttt{x := new($n,m$)}. We take $ \interp{S}{\texttt{c}}{I'} = \qex{x}{A}$, where $A$ is obtained as a solution to $\abduce{\post(\texttt{c},S)}{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}]}{I'}$, and $\vec{a}$ and $\vec{z}$ are vectors of fresh variables of length $n$ and $m$, respectively. \paragraph{Deallocate} Suppose \texttt{c} is \texttt{free(x)}. We take $\interp{S}{\texttt{c}}{I'} = \qex{\vec{a},\vec{z}}{I' * x \mapsto [\vec{a},\vec{z}]}$, where $\vec{a}$ and $\vec{z}$ are vectors of fresh variables whose lengths are determined by the unique heap cell which is allocated to $x$ in $S$. \paragraph{Assignment} Suppose \texttt{c} is \texttt{x := E}. We take $\interp{S}{\texttt{c}}{I'} = I'[E/x]$. \paragraph{Store} Suppose \texttt{c} is \texttt{x->N$_i$ := E}. We take $\interp{S}{\texttt{c}}{I'} = \qex{\vec{a},\vec{z}}{A * x \mapsto [\vec{a},\vec{z}]}$, where $A$ is obtained as a solution to $\post(\texttt{c},S) \vdash \qex{\vec{a},\vec{z}}{x \mapsto [\vec{a},\vec{z}[E/z_i]] * [A]} \vdash I'$ and where $\vec{a}$ and $\vec{z}$ are vectors of fresh variables whose lengths are determined by the unique heap cell which is allocated to $x$ in $S$. \begin{example} Suppose that $S$ is $t \mapsto [4,y,\nil] * x \mapsto [2,\nil,\nil]$ where the cells have one data and two pointer fields, \texttt{c} is \texttt{t->N$_0$ := x}, and $I'$ is $\textsf{bt}(t)$. Then we can compute $\post(\texttt{c},S) = t \mapsto [4,x,\nil] * x \mapsto [2,\nil,\nil]$, and then solve the bounded abduction problem \[ \post(\texttt{c},S) \vdash \qex{a,z_1}{t \mapsto [a,x,z_1] * [\ ]} \vdash I'\ .\] One possible solution is $A = \bt(x) * \bt(z_1)$, which yields \[ \interp{S}{\texttt{c}}{I'} = \qex{a,z_0,z_1}{t \mapsto [a,z_0,z_1] * \bt(z_1) * \bt(x)} \ .\tag*{\ensuremath{\lrcorner}} \] \end{example} \paragraph{Load} Suppose \texttt{c} is \texttt{y := x->N$_i$}. Suppose that $\vec{a}$ and $\vec{z}$ are vectors of fresh variables of lengths $|\vec{A}|$ and $|\vec{E}|$ where $S$ is of the form $\Pi : \Sigma * w \mapsto [\vec{A},\vec{E}]$ and $\Pi:\Sigma * w \mapsto [\vec{A},\vec{E}] \vdash x = w$ (this is the condition under which $\post(\texttt{c}, S)$ is defined, see Fig.~). Let $y'$ be a fresh variable, and define $\ol{S} = (y = z_i[y'/y]) * (\Pi : \Sigma * w \mapsto [\vec{a},\vec{z}])[y'/y]$. Note that $\ol{S} \vdash \qex{y'}{\ol{S}} \equiv \post(\texttt{c},S) \vdash I'$. We take $\interp{S}{\texttt{c}}{I'} = \qex{\vec{a},\vec{z}}{A[z_i/y,y/y'] * x \mapsto [\vec{a},\vec{z}]}$ where $A$ is obtained as a solution to $\ol{S} \vdash \qex{\vec{a},\vec{z}}{x[y'/y] \mapsto [\vec{a},\vec{z}] * [A]} \vdash I'$. \begin{example} Suppose that $S$ is $y = t : y \mapsto [1,\nil,x] * x \mapsto [5,\nil,\nil]$, \texttt{c} is \texttt{y := y->N$_1$}, and $I'$ is $y \neq \nil : \bt(t)$. Then $\ol{S}$ is \[ y = x \land y' = t : y' \mapsto [1,\nil,x] * x \mapsto [5,\nil,\nil] \] We can then solve the bounded abduction problem \[ \overline{S} \vdash \qex{a,z_0,z_1}{y' \mapsto [a,z_0,z_1] * [\ ]} \vdash I'\] A possible solution is $y \neq \nil \land y' = t : \bt(z_0) * \bt(z_1)$, yielding\\ \hspace*{0.5cm}$\interp{S}{\texttt{c}}{I'} = (\exists a,z_0,z_1. z_1 \neq \nil \land y = t :\bt(z_0) * \bt(z_1) * y \mapsto[a,z_0,z_1])\ .$\hspace*{0.5cm}\null \ensuremath{\lrcorner} \end{example} \paragraph{Assumptions} The interpolation rules defined up to this point cannot introduce recursive predicates, in the sense that if $I'$ is a \mbox{*-conjunction} of points-to predicates then so is $\interp{S}{\texttt{c}}{I'}$. contain recursive predicates, then $\interp{S}{\texttt{c}}{I'}$ may also.} A \mbox{*-conjunction} of points-to predicates is \emph{exact} in the sense that it gives the full layout of some part of the heap. The power of recursive predicates lies in their ability to be \emph{abstract} rather than exact, and describe only the shape of the heap rather than its exact layout. It is a special circumstance that $\hoare{P}{\texttt{c}}{I'}$ holds when $I'$ is exact in this sense and $P$ is not: intuitively, it means that by executing \texttt{c} we somehow gain information about the program state, which is precisely the case for \texttt{assume} commands. For an example of how spatial interpolation can introduce a recursive predicate at an \texttt{assume} command, consider the problem of computing an interpolant \[\interp{S}{\texttt{assume(x $\neq \nil$)}}{(\exists a,z.\; x\mapsto[a,z] * \true)}\] where $S \equiv x \mapsto [d,y] * y \mapsto [d',\nil]$: a desirable interpolant may be $\ls(x,\nil) * \true$. The disequality introduced by the assumption ensures that one of the $\mathit{cases}$ of the recursive predicate $\ls(x,\nil)$ (where the list from $x$ to $\nil$ is empty) is impossible, which implies that the other case (where $x$ is allocated) must hold. Towards this end, we now define an auxiliary function $\textsf{intro}$ which we will use to introduce recursive predicates for the \texttt{assume} interpolation rules. Let $P,Q$ be \sep formulas such that $P \vdash Q$, let $Z$ be a recursive predicate and $\vec{E}$ be a vector of heap terms. We define $\textsf{intro}(Z,\vec{E},P,Q)$ as follows: if $\abduce{P}{\emptyset}{Z(\vec{E})}{Q}$ has a solution and $A \nvdash Q$, define $\textsf{intro}(Z,\vec{E},P,Q) = Z(\vec{E}) * A$. Otherwise, define $\textsf{intro}(Z,\vec{E},P,Q) = Q$. Intuitively, the abduction problem has a solution when $P$ implies $Z(\vec{E})$ and $Z(\vec{E})$ can be \emph{excised} from $Q$. The condition $A \nvdash Q$ is used to ensure that the excision from $Q$ is non-trivial (i.e., the part of the heap that satisfies $Z(\vec{E})$ ``consumes'' some heaplet of $Q$). To define the interpolation rule for assumptions, suppose \texttt{c} is \texttt{assume(E $\neq$ F)} (the case of equality assumptions is similar). Letting $\{\tuple{Z_i,\vec{E}_i}\}_{i\leq n}$ be an enumeration of the (finitely many) possible choices of $Z$ and $\vec{E}$, we define a formula $M$ to be the result of applying $\textsf{intro}$ to $I'$ over all possible choices of $Z$ and $\vec{E}$: \shortenpar \[M = \textsf{intro}(Z_1,\vec{E}_1,S\land E \neq F,\textsf{intro}(Z_2,\vec{E}_2,S \land E \neq F,\dotsc))\] where the innermost occurrence of $\textsf{intro}$ in this definition is $\textsf{intro}(Z_n,\vec{E}_n,S\land E \neq F, I')$. Since \textsf{intro} preserves entailment (in the sense that if $P \vdash Q$ then $P \vdash \textsf{intro}(Z,\vec{E},P,Q)$), we have that $S \land E \neq F \vdash M$. From a proof of $S \land E \neq F \vdash M$, we can construct a formula $M'$ which is entailed by $S$ and differs from $M$ only in that it renames variables and exposes additional equalities and disequalities implied by $S$, and take $\interp{S}{\texttt{c}}{I'}$ to be this $M'$. The construction of $M'$ from $M$ is straightforward but tedious. \emph{The procedure is detailed in \iflong Appendix~ \else the extended version\fi; here, we will just give an example to give intuition on why it is necessary.} Suppose that $S$ is $x = w : y \mapsto z$ and $I'$ is $\ls(w,z)$, and \texttt{c} is \texttt{assume($x$ = $y$)}. Since there is no opportunity to introduce new recursive predicates in $I'$, $M$ is simply $\ls(w,z)$. However, $M$ is not a valid interpolant since $S \not\models M$, so we must expose the equality $x=w$ and rename $w$ to $y$ in the list segment in $M' \equiv x = w: \ls(y,z)$. In practice, it is undesirable to enumerate all possible choices of $Z$ and $\vec{E}$ when constructing $M$ (considering that if there are $k$ in-scope data terms, a recursive predicate of arity $n$ requires enumerating $k^n$ choices for $\vec{E}$). A reasonable heuristic is to let $\Pi$ be the strongest pure formula implied by $S$, and enumerate only those combinations of $Z$ and $\vec{E}$ such that there is some $\Pi' :\Sigma' \in \mathit{cases}(Z(\vec{R},\vec{x}))$ such that $\ul{\Pi'}[\vec{E}/\vec{x}] \land \Pi \land x \neq y$ is unsatisfiable. For example, for \texttt{assume(x $\neq$ y)}, this heuristic means that we enumerate only $\tuple{x,y}$ and $\tuple{y,x}$ (i.e, we attempt to introduce a list segment from $x$ to $y$ and from $y$ to $x$). \shortenpar We conclude this section with a theorem stating the correctness of our spatial interpolation procedure. \begin{theorem} Let $S$ and $I'$ be \sep formulas and let \texttt{c} be a command such that $\post(\texttt{c},S) \vdash I'$. Then $\interp{S}{\texttt{c}}{I'}$ is a spatial interpolant for $S$, \texttt{c}, and $I'$. \end{theorem} |
1501.04101 | Title: Quantization of the conformal arclength functional on space curves
Abstract: By a conformal string in Euclidean space is meant a closed critical curve
with non-constant conformal curvatures of the conformal arclength functional.
We prove that (1) the set of conformal classes of conformal strings is in 1-1
correspondence with the rational points of the complex domain $\{q\in
\mathbb{C} \,:\, 1/2 < \mathrm{Re}\, q < 1/\sqrt{2},\,\, \mathrm{Im}\, q >
0,\,\, |q| < 1/\sqrt{2}\}$ and (2) any conformal class has a model conformal
string, called symmetrical configuration, which is determined by three
phenomenological invariants: the order of its symmetry group and its linking
numbers with the two conformal circles representing the rotational axes of the
symmetry group. This amounts to the quantization of closed trajectories of the
contact dynamical system associated to the conformal arclength functional via
Griffiths' formalism of the calculus of variations.
Body: \pdfoutput=1 \title[Quantization of the conformal arclength functional]{Quantization of the conformal arclength functional on space curves} \author{Emilio Musso} \address{(E. Musso) Dipartimento di Scienze Matematiche, Politecnico di Torino, Corso Duca degli Abruzzi 24, I-10129 Torino, Italy} \email{emilio.musso@polito.it} \author{Lorenzo Nicolodi} \address{(L. Nicolodi) Di\-par\-ti\-men\-to di Scienze Ma\-te\-ma\-ti\-che, Fisiche e Informatiche, Uni\-ver\-si\-t\`a di Parma, Parco Area delle Scienze 53/A, I-43124 Parma, Italy} \email{lorenzo.nicolodi@unipr.it} \thanks{Authors partially supported by PRIN 2010-2011 ``Variet\`a reali e complesse: geometria, to\-po\-lo\-gia e analisi ar\-mo\-ni\-ca''; FIRB 2008 ``Geometria Differenziale Complessa e Dinamica Olomorfa''; and the GNSAGA of INDAM} \subjclass[2000]{53A30, 53A04, 53A55, 53D20, 58A17, 58-04} \keywords{M\"obius geometry of curves, closed trajectories, conformal arclength functional, conformal strings, quantization of trajectories, Griffiths' formalism, linking numbers} \begin{abstract} By a \textit{conformal string} in Euclidean space is meant a closed critical curve with non-constant conformal curvatures of the conformal arclength functional. We prove that (1) the set of conformal classes of conformal strings is in 1-1 correspondence with the rational points of the complex domain $\{q\in \mathbb{C} \,:\, {1}/{2} < \Re q< {1}/{\sqrt{2}},\,\, \Im q>0,\,\, |q| < {1}/{\sqrt{2}}\}$ and (2) any conformal class has a model conformal string, called \textit{symmetrical configuration}, which is determined by three phenomenological invariants: the order of its symmetry group and its linking numbers with the two conformal circles representing the rotational axes of the symmetry group. This amounts to the quantization of closed trajectories of the contact dynamical system associated to the conformal arclength functional via Griffiths' formalism of the calculus of variations. \end{abstract} \maketitle \section*{Introduction} The M\"obius geometry of space curves was mainly developed in the first half of the past century and later taken up starting from the early 1980's . Further developments of the subject as an instance of the conformal geometry of submanifolds can be found in and the literature therein. The subject has also received much attention for its many fields of application, including the theory of integrable systems , the topology and M\"obius energy of knots , and the geometric approach to shape analysis and medical imaging . Let $\gamma\subset\R^n$, $n\geq 3$, be a smooth curve parametrized by arclength $s$. The {\em conformal arclength} parameter $\zeta$ of $\gamma$ is defined (up to a constant) by \begin{equation} d\zeta = \left(\langle\dddot\gamma,\dddot\gamma\rangle -\langle\ddot\gamma, \ddot\gamma\rangle^2 \right)^{\frac{1}{4}}ds =: \eta_\gamma, \end{equation} where $\langle\,,\rangle$ is the standard scalar product on $\R^n$. The 1-form $\eta_\gamma$, the {\em infinitesimal conformal arclength} of $\gamma$, is conformally invariant. If ${\eta_\gamma}{\vert_s} \neq 0$, for each $s$, the curve is called {\em generic}.\footnote{Generic curves, either closed or not, constitute an open dense subset of all smooth curves in the $C^\infty$ topology (cf. ).} The conformal arclength $\zeta$ gives a conformally invariant parametrization of a generic curve. We consider the conformally invariant variational problem on generic curves defined by the {\em conformal arclength functional} \begin{equation} \mathcal{L}[\gamma] = \int_\gamma \eta_\gamma. \end{equation} This variational problem was studied in for $n=3$ and more recently in for higher dimensions.\footnote{We adhere to the standard terminology adopted for $\mathcal L$ in the literature. However, observe that $\left(\langle\dddot\gamma,\dddot\gamma\rangle -\langle\ddot\gamma, \ddot\gamma\rangle^2 \right)^{1/4}$ has the dimension $L^{-1}$, so that $\eta_\gamma$ is dimensionless.} Accordingly, the critical curves can be found by quadratures and explicit parametrizations are given in terms of elliptic functions and integrals. In this paper we address the question of existence and properties of closed critical curves for the functional $\mathcal L$. Actually, it suffices to consider the 3-dimensional case only, since from the results in we can see that any closed critical curve in $\R^n$ lies in some $\R^3 \subset \R^n$, up to a conformal transformation. It is known that a generic space curve is determined, up to conformal transformations, by the conformal arclength and two conformal curvatures (cf. Section ). As for a closed critical curve with constant conformal curvatures, one can see that it is conformally equivalent to a closed rhumb line (loxodrome) of a torus of revolution (cf. Example ). \vskip0.1cm The purpose of this paper is to study the class of closed critical curves with non-constant conformal curvatures, for brevity called {\it conformal strings}. We begin by describing our three main results. If $\gamma$, $\widetilde{\gamma}:\R \to \R^3$ are two curves and $[\gamma]$, $[\widetilde{\gamma}]$ denote their trajectories, then $\gamma$ and $\widetilde{\gamma}$ are said {\it M\"obius} ({\it conformally}) {\it equivalent} if there is an element $A$ of the M\"obius group $G$ of $\mathbb R^3$, such that $A\cdot [\gamma]=[\widetilde{\gamma}]$. By a {\it conformal symmetry} of a curve $\gamma$ is meant an element $A\in G$, such that $A\cdot [\gamma]=[\gamma]$. The set of all symmetries of $\gamma$ is a subgroup $G_{\gamma}$ of $G$. The symmetry group of a closed curve with non-constant conformal curvatures is finite and its cardinality is called the {\it symmetry index} of $\gamma$. Our first main result is the following. \begin{thmx} The M\"obius classes of conformal strings are in 1-1 correspondence with the rational points of the complex domain \[ \Omega= \left\{ q\in \mathbb{C} \,\,:\,\, \frac{1}{2} < \Re q<\frac{1}{\sqrt{2}},\,\, \Im q>0,\,\, |q|< \frac{1}{\sqrt{2}} \right\}. \] The rational points of $\Omega$ are called the {\em moduli} of conformal strings. \end{thmx} Using this theorem and other technical results, we will prove that any M\"obius class of strings is represented by a model string. This is our second main result. \begin{thmx} The conformal strings corresponding to a modulus $q\in \Omega$ are M\"obius equivalent to a model string $\gamma_{q}=(x(t),y(t),z(t)) : \R\to \R^3$, \begin{equation} \begin{cases} x(t)=\frac{\sqrt{2}}{r(t)}\mu\sqrt{k(t)^2-\upsilon^2}\cos \Theta_2(t) ,\\ y(t)=\frac{\sqrt{2}}{r(t)}\mu\sqrt{k(t)^2-\upsilon^2}\sin \Theta_2(t) ,\\ z(t)=\frac{\sqrt{2}}{r(t)}\upsilon\sqrt{\mu^2-k(t)^2}\sin \Theta_1(t),\\ \end{cases} \end{equation} called the {\em symmetrical configuration} of $q$. Here \[ \begin{aligned} k(t) &= \left\{\!\!\! \begin{array}{lr} \sqrt{a} \cn\big(\sqrt{a-b}\,t, \frac{a}{a-b}\big),& b<0,\\ \sqrt{a}\dn\big(\sqrt{a}\,t,\frac{a-b}{a}\big), & b>0, \end{array} \right.\\ r(t)&=\sqrt{\mu^2-\upsilon^2}k(t)+\upsilon\sqrt{\mu^2-k(t)^2} \cos \Theta_1(t),\\ \end{aligned} \] where $\mu=\frac{1}{\sqrt{2}}\sqrt{a+b+\sqrt{4+(a-b)^2}}$, $\upsilon=\frac{1}{\sqrt{2}}\sqrt{a+b-\sqrt{4+(a-b)^2}}$, \[ \Theta_1(t)=\int_0^t\frac{\mu}{\mu^2-k(u)^2}du,\quad \Theta_2(t)=\int_0^t\frac{\upsilon}{\upsilon^2-k(u)^2}du, \] and where $a$ and $b$ are real parameters, uniquely defined by $q$, such that $a>0$, $a>b$, $b\neq0$, and $ab>1$. \end{thmx} The symmetry group of a symmetrical configuration has a special structure, which is described by the following. \begin{thmx} Let $\gamma_{q} : \R\to \R^3$ be the symmetrical configuration corresponding to the modulus $q = q_1 +i q_2$, where $q_1=m_1/n_1$, $q_2=m_2/n_2$, and the pairs $(m_1,n_1)$, $(m_2,n_2)$ are coprime integers. Let $n$ be the least common multiple of $n_1$ and $n_2$, and consider the coprime integers $h_1=n/n_1$ and $h_2=n/n_2$. Then, \begin{enumerate} \item $n$ is the order of the symmetry group of $\gamma_{q}$; \item $m_1h_1$ and $m_2h_2$ are the linking numbers of $\gamma_{q}$ with the Clifford circle \[ \mathcal{C}=\left\{(x,y,0) \in\R^3 : x^2+y^2=2\right\} \] and the $z$-axis, respectively. \end{enumerate} \end{thmx} An important consequence of the previous results is that the conformal shape of a string is uniquely determined by three phenomenological invariants: the order of its symmetry group and its linking numbers with the two axes of the symmetry group. The explicit construction of a string from the phenomenological invariants requires the inversion of the period map of $\mathcal L$ (cf. Section ). In this respect, some numerical experiments carried out with the software {\sl Mathematica} suggest that conformal strings are simple curves (cf. Section ). However, a rigorous proof of this fact is still missing. Another interesting problem is to find an estimate for the asymptotic growth of $\varrho(n)$, the cardinality of the set of M\"obius classes of strings with symmetry order $n$. Numerical experiments suggest a quadratic growth of $\varrho(n)$. \vskip0.1cm The theorems above have a conceptual explanation within the general scheme of Griffiths' formalism of the calculus of variations . Using Griffiths' formalism, the {\it momentum space} of the variational problem can be identified with $Y=G_+\times \mathcal{A}$, where $G_+$ is the identity component of the M\"obius group of $\R^3$ and $\mathcal{A}$ is a 3-dimensional submanifold of $\mathfrak{g}^*$, the dual of the Lie algebra of $G_+$. Moreover, the restriction $\xi\in \Omega^1(Y)$ of the Liouville form of $T^*(G_+)$ defines an invariant contact structure on $Y$. By choosing a suitable set of coordinates on $\mathcal{A}$, say $p_1$, $p_2$ and $p_3$, the characteristic curves of the contact form are given by \[ t\in \R\mapsto (\mathrm{F}(t),p_1(t),p_2(t),p_3(t))\in Y, \] where $\mathrm{F}$ is the canonical lift of a stationary curve $\gamma$, $p_1$ and $p_2$ are the conformal curvatures, and $p_3=p_1'$. From a theoretical point of view, the study of the variational problem is equivalent to that of the dynamical system defined by the characteristic vector field of the contact form $\xi$. One can easily see that the contact momentum map is given by \[ J:(\mathrm{F},p)\in Y\mapsto \mathrm{Ad}^{*}(F)(\xi|_{(\mathrm{F},p)})\in \mathfrak{g}^*. \] (For the construction of the momentum map in contact geometry, see for instance .) Moreover, the action of $G_+$ on $Y$ is Hamiltonian and coisotropic, and the characteristic vector field is {\it collective completely integrable} .\footnote{Here, we adopt the terminology used in . In the literature, the term non-commutative completely integrable systems is also used.} Consequently, the flow can be linearized on the fibers of the momentum map and its trajectories can be found by quadratures (see for a general description of the integration procedure). Theorems and say that the contact dynamical system is quantizable, at least in the sense of ``the old quantum theory'' (Bohr's atom theory) , and that the quantum numbers of the closed (quantizable) trajectories have a precise geometric meaning. \vskip0.1cm The paper is organized as follows. Section collects some basic facts about the conformal geometry of space curves. Section recalls the Euler--Lagrange equations of the variational problem and discusses the example of closed critical curves with constant conformal curvatures. Then, after introducing the natural parameters of a critical curve with non-constant periodic conformal curvatures, the period map of the functional $\mathcal L$ is defined, and conformal strings are characterized in terms of the natural parameters, via the period map. Section deals with a technical result about the period map, from which Theorem follows directly. Section proves Theorem , while Section proves Theorem . Section discusses some examples. \vskip0.1cm Numerical and symbolic computations, as well as graphics, are made with the software {\sl Mathematica}. As basic references for the theory of elliptic functions and integrals we use the monographs (see also ). For the few notions of knot theory used in the paper we refer to . A general reference for M\"obius geometry is , to which we refer for an updated list of modern and classical references to the subject (see also ). The main results of the paper were previously announced in . \vskip0.1cm (Added in proof) Since acceptance of the manuscript, the paper has appeared, which studies the local and global conformal invariants of timelike curves in the (1+2)-Einstein universe and addresses the question of existence and properties of closed trajectories for the conformal strain functional. \subsubsection*{Acknowledgments} The authors would like to thank the referees for their valuable comments and suggestions. \section{Preliminaries} \subsection{The conformal group} Let $\R^{4,1}$ denote $\R^5$ with the Lorentz scalar product \begin{equation} (v,w)=-(v^0w^4+v^4w^0)+\sum_{j=1}^{3} v^jw^j = \sum_{a,b=0}^{4}g_{ab}v^aw^b,\quad g_{ab}=g_{ba}, \end{equation} where $v=(v^0,\dots,v^4)$, and with the space and time orientations defined, respectively, by the volume form $dv^0\wedge\cdots\wedge dv^4$ and the positive light cone \begin{equation} \mathcal{L}_+=\{v \in \R^{4,1} \,: \,(v,v)=0,\, v^0+v^4>0\}. \end{equation} The {\it M\"obius space} $\mathcal{M}_3$ is the projectivization of $\mathcal{L}_+$, endowed with the oriented conformal structure induced by the scalar product and the space and time orientations. If $(e_0,\dots,e_4)$ is the standard basis of $\R^{4,1}$, the map \begin{equation} \mathcal{J}:x=(x^1,x^2,x^3)\in \R^3\mapsto \left[\frac{\langle x, x\rangle}{2}e_0 +\sum_{j=1}^{3}x^je_j+e_4\right]\in \mathcal{M}_3. \end{equation} is an orientation-preserving conformal diffeomorphism of $\R^3$ onto the M\"obius space minus the point $P_{\infty}=[e_0]$. The inverse of $\mathcal{J}$ is the conformal projection \begin{equation} \mathcal{P}:\left[\sum_{a=0}^{4}v^ae_a\right]\in \mathcal{M}_3\setminus\{P_{\infty}\}\mapsto \frac{1}{v^4}(v^1,v^2,v^3)\in \R^3. \end{equation} The {\it M\"obius group} $G$ consists of all pseudo-orthogonal transformations preserving the volume form. It is a 10-dimensional Lie group with two connected components. The first component is the subgroup $G_+$ consisting of all $\mathrm{F}\in G$ preserving the positive light cone and the second one consists of all $\mathrm{F}\in G$ switching the positive light cone with the negative one. The group $G$ acts effectively and transitively on the left of $\mathcal{M}_3$ preserving the conformal structure. The classical Liouville theorem asserts that every conformal automorphism of $\mathcal{M}_3$ is induced by a unique element of $G$. Consequently, the M\"obius group can be viewed as the pseudo-group of all conformal transformations of Euclidean 3-space. The orientation-preserving conformal transformations are induced by the elements of $G_+$, while the conformal transformations induced the elements of $G_-$ are orientation-reversing. For each $\mathrm{F}\in G_+$, we denote by $\mathrm{F}_0,\dots,\mathrm{F}_4$ its column vectors. Then, $(\mathrm{F}_0,\dots,\mathrm{F}_4)$ is a {\it positive light cone} basis of $\R^{4,1}$, that is a positive-oriented basis such that \[ (\mathrm{F}_a,\mathrm{F}_b)=g_{ab},\quad \mathrm{F}_0,\mathrm{F}_4\in \mathcal{L}_+ , \quad a,b=0,\dots,4. \] Conversely, if $(\mathrm{F}_0,\dots,\mathrm{F}_4)$ is a positive light-cone basis, then the matrix $\mathrm{F}$ with column vectors $\mathrm{F}_0,\dots,\mathrm{F}_4$ i s an element of $G_+$. The Lie algebra of $G$ consists of all skew-adjoint matrices of the scalar product \eqref{1.1.1}, that is \[ \mathfrak{g}=\left\{\mathrm{X}\in \mathfrak{gl}(5,\R) \,:\, ^t\mathrm{X} g + g \mathrm{X}=0 \right\}, \] where $g=(g_{ab})$. The maximal compact abelian subgroups of $G$ are conjugate to the 2-dimensional torus \begin{equation} T =\left\{R(\phi_1,\phi_2)\,:\, \phi_1,\phi_2\in [0,2\pi)\right\}\cong SO(2)\times SO(2), \end{equation} where \begin{equation} R(\phi_1,\phi_2)=\left( \begin{smallmatrix} \frac{1+\cos \phi_2}{2} & 0 & 0 & -\frac{\sin\phi_2}{\sqrt{2}} & \frac{1-\cos\phi_2}{2} \\ 0 & \cos\phi_1 & -\sin\phi_1 & 0 & 0 \\ 0 & \sin\phi_1 & \cos\phi_1 & 0 & 0 \\ \frac{\sin\phi_2}{\sqrt{2}} & 0 & 0 & \cos\phi_2 & -\frac{\sin\phi_2}{\sqrt{2}} \\ \frac{1-\cos\phi_2}{2} & 0 & 0 & \frac{\sin\phi_2}{\sqrt{2}} & \frac{1+\cos\phi_2}{2} \\ \end{smallmatrix} \right). \end{equation} Note that $R(\phi_1,\phi_2)$ is the composition of the Euclidean rotation of angle $\phi_1$ around the $z$-axis with the {\it toroidal rotation} of angle $\phi_2$ around the {\it Clifford circle} $\mathcal{C}=\{(x,y,0) : x^2+y^2=2\}$. The $z$-axis and the Clifford circle are the {\it rotational axes} of $T$. The rotational axes of any other maximal torus $\mathrm{F} T \mathrm{F}^{-1}$ are the images under $\mathrm{F}$ of the axes of $T$. \subsection{M\"obius geometry of space curves} Let $\gamma : I\subset \R \to \R^3$ be a smooth curve parametrized by arclength $s$, $I$ an open interval. Points where the infinitesimal conformal arclength $\eta_\gamma$ (cf.~\eqref{inf-conf-par}) vanishes are called {\it vertices} of $\gamma$ (cf. ). Generic curves, i.e., without vertices, can be parametrized by the {\it conformal arc\-length} parameter $\zeta$, defined (up to a constant) by $d\zeta=\eta_\gamma$. If such a conformal parametrization is defined for every $\zeta\in \R$, the curve is said {\it complete}. A {\it frame field} along $\gamma$ is a smooth map $\mathrm{F} :I\to G_+$, such that $\mathcal{P}\circ \mathrm{F}_4=\gamma$. We have the following. \begin{prop}[cf. ] For any oriented generic curve $\gamma : I\to \R^3$, there is a unique frame field $\mathrm{F}:I\to G_+$ along $\gamma$, the {\em Vessiot frame}, such that \begin{equation} \mathrm{F}^{-1}d\mathrm{F}= \left( \begin{smallmatrix} 0 & 1 & 0 & 0 & 0 \\ k_2 & 0 & 0 & 0 & 1 \\ 1 & 0 & 0 & k_1 & 0 \\ 0 & 0 & -k_1 & 0 & 0 \\ 0 & k_2 & 1 & 0 & 0 \\ \end{smallmatrix} \right)\eta_\gamma, \end{equation} where $k_1$, $k_2$ are smooth functions, called the {\em conformal curvatures}. We call $\Gamma = \mathrm{F}_4:I\to \mathcal{L}_+$ the {\em canonical null lift} of $\gamma$. \end{prop} \begin{remark} If $\gamma$ is biregular and $T = \dot\gamma$, $N$, $B$ is its Frenet frame, with Frenet--Serret equations $\dot T = \kappa N$, $\dot N = -\kappa T + \tau B$, $\dot B = -\tau N$, we have ${\dddot\gamma}= -\kappa^2 T +\dot \kappa N + \kappa\tau B$ and then $\eta_\gamma = \sqrt[4]{\dot{\kappa}^2+\kappa^2\tau^2}\, ds$ (cf. ). The conformal curvatures take the form \begin{equation} k_1=r^5\left(\kappa^2\tau^3+\kappa \dot{\kappa}\dot{\tau} +\tau(2\dot{\kappa}^2-\kappa \ddot{\kappa}) \right),\quad k_2 =\frac{1}{2}\left(\dot{r}^2 -2r\ddot{r}-r^2\kappa^2 \right), \end{equation} where $r = (\dot{\kappa}^2+\kappa^2\tau^2)^{-1/4}$ (cf. ). For a geometric description of vertices and the conformal arclength via the osculating circles of $\gamma$ we refer to . \end{remark} \begin{remark} The construction of $\mathrm F$ is explicit and only involves derivatives and simplification of $\mathrm{F}^{-1}d\mathrm{F}$ by linear relations on entries. If $\gamma : \R\to \R^3$ is a periodic parametrization of a closed curve, $\mathrm F$ is a periodic $G_+$-valued map. The value of $\mathrm F$ at $t\in I$ depends on the fourth order jet of $\gamma$ at $t$. If $\eta_\gamma = d\zeta$, any solution of the linear system \eqref{1.2.3} is the Vessiot frame of the conformal parametrization $\gamma=\mathcal{P}\circ [\mathrm{F}_4]$ of a generic curve with curvatures $k_1$ and $k_2$. If $\mathrm{F}$, $\widetilde{\mathrm{F}}:\R\to G_+$ are two solutions of \eqref{1.2.3}, with initial conditions in $G_+$, there is a unique $A\in G_+$, such that $\widetilde{\mathrm{F}}=A \mathrm{F}$. This shows that the conformal curvatures determine the curve, up to an orientation-preserving conformal transformation. A curve $\gamma :\R\to \R^3$ is called {\it chiral} if its symmetry group $G_\gamma$ is contained in $G_+$. \end{remark} \begin{defn} A generic curve $\gamma: \R \to \R^3$ parametrized by conformal arclength is said {\it quasi-periodic} if its conformal curvatures are non constant, periodic, with a common minimal period $\omega >0$; $\omega$ is called the {\it conformal wavelength} of $\gamma$. The {\it monodromy} of a quasi-periodic curve is the element $M := \mathrm{F}(\omega)\mathrm{F}(0)^{-1}\in G_+$. By construction, $M$ satisfies \begin{equation} \mathrm{F}(t+p \omega)= M^p \mathrm{F}(t), \quad \forall t\in \R,\quad \forall p\in \mathbb{Z}. \end{equation} Therefore, $M$ generates a subgroup $\widehat{G}_{\gamma}\subset G_{\gamma}$, the {\it monodromy group} of $\gamma$. If $\gamma$ is closed, the integral of $\eta_\gamma$ along $\gamma$ is $n\omega$, where $n$ is the cardinality of $\widehat{G}_{\gamma}$. If $A$ is an orientation-reversing conformal transformation, the conformal curvatures of $A\cdot \gamma$ are $-k_1$ and $k_2$, respectively. Thus, generic curves whose first curvature is nowhere vanishing are chiral. In this case, we assume that $\gamma$ has {\it positive chirality} (i.e., $k_1>0$). If $\gamma$ is a real-analytic curve with positive chirality, then $\widehat{G}_{\gamma}=G_{\gamma}$ and $\gamma$ is closed if and only if $M$ has finite order. Note that the symmetry index of a real-analytic closed curve with positive chirality coincides with the order of the monodromy. \end{defn} \section{Critical curves} \subsection{The Euler-Lagrange equations} The critical curves of the functional $\mathcal{L}$ \eqref{var-pbm} are characterized by the Euler--Lagrange equations \begin{equation} k_1''=2k_1(C_1-k_1^2),\quad k_2=-\frac{3}{2}k_1^2+C_1, \end{equation} where $C_1$ is a constant of integration and $k_1''$ is the second order derivative of $k_1$ with respect to the conformal arclength $\zeta$. \begin{ex}[Critical curves with constant conformal curvatures] Let $\gamma$ be a critical curve with constant conformal curvatures. Then, either $k_1=0$ and $k_2\in \R$, or $k_2=-k_1^2/2$ and $k_1\in \R\setminus\{0\}$. In the second case, we may assume $k_1>0$. The class of curves with constant conformal curvatures was studied in : they are equivalent to the rhumb lines of either a torus of revolution, or a round cone, or else a circular cylinder. Since we deal with curves without vertices, the meridians and the parallels must be excluded from the discussion. The rhumb lines of a round cone, which possibly can degenerate into the punctured plane, are helices over logarithmic spirals, while the rhumb lines of a circular cylinder are circular helices. All of them are not closed. From the viewpoint of the conformal geometry, any rotationally invariant torus is equivalent to a torus $\mathcal{T}_r$ generated by rotating around the $z$-axis the circle in the $xz$-plane with radius $(2-r^2)/2r$ and center $((r^2+1)/2r,0,0)$, for some $r\in (0,\sqrt{2})$. The latter are the regular orbits of the action of the group $T$ (cf. \eqref{1.1.5}) on the Euclidean space. If \[ \begin{cases} x_r(\theta,\phi)=\frac{4r\cos\theta}{2+r^2+(2-r^2)\cos\phi},\\ y_r(\theta,\phi)=\frac{4r\sin\theta}{2+r^2+(2-r^2)\cos\phi},\\ z_r(\theta,\phi)=\frac{\sqrt{2}(r^2-2)\sin\phi}{2+r^2+(2-r^2)\cos\phi} \end{cases} \] are parametric equations of $\mathcal{T}_r$, its rhumb lines are \[ \alpha_{r,m,n}: t\in \R \mapsto (x_r(nt,mt),y_r(nt,mt),z_r(nt,mt))\in \R^3, \] where $m,n\in \R$ and $mn\neq 0,n\neq \pm m$. The conformal curvatures of $\alpha_{r,m,n}$ are \[ \begin{cases} k_1=-\frac{(2+r^2)mn}{\sqrt[4]{8m^2n^2(m^2-n^2)^2r^2(2-r^2)^2}},\\ k_2=\frac{(8n^4r^2+m^4(r^2-2)^2)\mid mn(m^2-n^2)\mid}{4\sqrt{2}rm^2n^2(m^2-n^2)(r^2-2)}. \end{cases} \] They verify the inequality $2k_1^2 k_2<-1$. Conversely, every curve with constant conformal curvatures that satisfies the above inequality is equivalent to a rhumb line of $\mathcal{T}_r$, for some $r\in (0,\sqrt{2})$. Possibly, $r$ can be computed in terms of the constants $k_1$ and $k_2$. From this it follows that, up to conformal transformations, the only closed critical curves with constant conformal curvatures are of the form $\alpha_{r(q),q,1}$, where $q$ is a positive rational number different from $1$ and \[ r(q)=\frac{\sqrt{2}}{q}\sqrt{2+q^2-2\sqrt{1+q^2}}. \] The trajectory of $\alpha_{r(q),q,1}$ is a torus knot of type $(m,n)$, where $m,n$ are coprime integers such that $m/n=q$, see Figure . \end{ex} From now on we consider critical curves with non-constant conformal curvatures, parametrized by the conformal arclength parameter. Then, \eqref{2.1.1} implies \begin{equation} (k_1')^2+k_1^4-2C_1k_1^2=C_2, \end{equation} where $C_2$ is another constant of integration. This equation has non-constant periodic solutions if and only if the polynomial $ P(t,C_1,C_2)=t^2-2C_1t-C_2 $ has two distinct real roots, denoted by $a,b$ and ordered in such a way that $a>b$, with $a>0$ and $b\neq 0$. Then \eqref{2.1.1} can be rewritten in the form \begin{equation} (k_1')^2+(k_1^2-a)(k_1^2-b)=0,\quad k_2 = -\frac{3}{2}k_1^2+\frac{a+b}{2}. \end{equation} The general solution of \eqref{2.1.3} is given by \begin{equation} \begin{aligned} k_1(t)=\epsilon_1 k_{(a,b)}(\epsilon_2 t+h),\quad k_2(t)=-\frac{3}{2}k_1(t)^2+\frac{a+b}{2}, \end{aligned} \end{equation} where $\epsilon_1,\epsilon_2 =\pm 1$, $h$ is any real constant, and $k_{(a,b)}$ is the Jacobi elliptic function \begin{equation} k_{(a,b)}(t)= \begin{cases} \sqrt{a}\cn\big(\sqrt{a-b} t,\frac{a}{a-b}\big),\quad b<0,\\ \sqrt{a}\dn\big(\sqrt{a} t,\frac{a-b}{a}\big), \quad\quad\,\,\, b>0. \end{cases} \end{equation} By possibly reversing the orientation along the curve, acting with an orientation-reversing conformal transformation and shifting the independent variable, we can assume that $\epsilon_1=\epsilon_2=1$ and $h=0$. We have proved the following. \begin{prop} The M\"obius classes of quasi-periodic critical curves are in 1-1 correspondence with the points $(a,b)$ of the planar region \begin{equation} S=\left\{(a,b) \,:\, a>0,\, a>b,\, b\neq 0 \right\}. \end{equation} We call $a$ and $b$ the {\em parameters} of a quasi-periodic critical curve. \end{prop} \begin{remark} The M\"obius class of a critical curve with parameters $a$ and $b$ is represented by any conformal parametrization $\gamma :\R\to \R^3$ with curvatures \begin{equation} k_1=k_{(a,b)},\quad k_2=-\frac{3}{2}k_{(a,b)}^2+\frac{a+b}{2}. \end{equation} Note that the minimal period of $k_{(a,b)}$ is given by \begin{equation} \omega_{(a,b)}= \begin{cases} {4K \left(a (a-b)^{-1}\right)}/{\sqrt{a-b}} \quad & b<0,\\ {2K\left(b^{-1}(a-b)\right)}/{\sqrt{a}} \quad & b>0, \end{cases} \end{equation} where \begin{equation} K(m)=\int_0^{\pi/2}\frac{1}{\sqrt{1-m\sin^2t}}dt \end{equation} is the complete elliptic integral of the first kind. \end{remark} \subsection{Closure conditions} Let $\Sigma$ denote the open subset of $S\subset \mathbb R^2$ defined by \begin{equation} \Sigma=\left\{(a,b)\in S \,:\, ab>1 \right\}. \end{equation} For every $(a,b)\in \Sigma$, let \begin{equation} \begin{aligned} \mu(a,b)&=\frac{1}{\sqrt{2}}\sqrt{a+b+\sqrt{4+(a-b)^2}},\\ \upsilon(a,b)& =\frac{1}{\sqrt{2}}\sqrt{a+b-\sqrt{4+(a-b)^2}}, \end{aligned} \end{equation} and define \begin{equation} \begin{aligned} \Phi_1(a,b)&:=\frac{1}{2\pi}\int_0^{\omega(a,b)}\frac{\mu(a,b)}{k_{(a,b)}(t)^2-\mu(a,b)^2}dt,\\ \Phi_2(a,b)&:=\frac{1}{2\pi}\int_0^{\omega(a,b)}\frac{\upsilon(a,b)}{k_{(a,b)}(t)^2-\upsilon(a,b)^2}dt. \end{aligned} \end{equation} These integrals can be evaluated in term of the complete integral of the third kind (cf., for example, ) \begin{equation} \Pi(n,m)=\int_0^{\pi/2}\frac{dt}{(1-n\sin^2 t)\sqrt{1-m\sin^2 t}},\quad -1<n,m<1. \end{equation} As a result, we have \begin{equation} \begin{aligned} \Phi_1(a,b)&=\frac{\mu(a,b)}{\pi \sqrt{a}(a-\mu(a,b)^2)}\Pi\left(\frac{a-b}{a-\mu(a,b)^2},\frac{a-b}{a}\right),\\ \Phi_2(a,b)&=\frac{\upsilon(a,b)}{\pi \sqrt{a}(a-\upsilon(a,b)^2)}\Pi\left(\frac{a-b}{a-\upsilon(a,b)^2},\frac{a-b}{a}\right). \end{aligned} \end{equation} \begin{defn} We call $\Phi = (\Phi_1,\Phi_2) : \Sigma \to \R^2$ the {\it period map} of the conformal arclength functional. By construction, $\Phi$ is non constant and real analytic. \end{defn} We are now in a position to state the following. \begin{thm} A quasi-periodic critical curve with parameters $a$ and $b$ is a conformal string if and only if $(a,b)\in \Sigma$ and $\Phi_1(a,b)$, $\Phi_2(a,b)\in \mathbb{Q}$. \end{thm} For every $(a,b)\in S$, let $\gamma:\R\to \R^3$ be the conformal parametrization of a quasi-periodic critical curve with parameters $(a,b)\in S$, such that $\mathrm{F}(0)=\mathrm{Id}$, where $\mathrm{F}$ is the Vessiot frame along $\gamma$. The claim is that $\gamma$ is periodic if and only if $(a,b)\in \Sigma$ and $\Phi_j(a,b)\in \mathbb{Q}$, $j=1,2$. For brevity, let $k$ denote the first conformal curvature $k_1$, let $\mu$ and $\upsilon$ denote the constants $\mu(a,b)$, $\upsilon(a,b)$, respectively, and let $\omega$ denote the minimal period $\omega(a,b)$. \begin{lemma}[cf. ] The canonical null lift $\Gamma :\R\to \mathcal{L}_+$ of $\gamma$ satisfies the linear system \begin{align} X'&=\left(\frac{k'}{k}\mathrm{Id}-\frac{k^2-a-b}{k'^2+1}\mathfrak{X} +\frac{k'(k^2-a-b)}{k(k'^2+1)}\mathfrak{X}^2 \right. \\ \notag &\qquad \left.+\frac{1}{k'^2+1}\mathfrak{X}^3-\frac{k'}{k(k'^2+1)}\mathfrak{X}^4\right) X, \end{align} with the initial condition $\Gamma(0)=e_4$, where \[ \mathfrak{X} =\left( \begin{smallmatrix} 0 & 0 & 1 & 0 & 0 \\ 1 & 0 & 0 & \sqrt{a} & 0 \\ -b/2 & 0 & 0 & 0 & 1 \\ 0 & -\sqrt{a} & 0 & 0 & 0 \\ 0 & 1 & -b/2 & 0 & 0\\ \end{smallmatrix} \right)\in \mathfrak{g}. \] \end{lemma} Using Lemma , an explicit integration of the critical curves was carried out in (see also ). The proof of Theorem is a direct consequence of this integration. According to the analysis carried out in , we are led to consider three cases, depending on whether $ab<1$, $ab=1$, or $ab>1$. It follows from that the conformal parametrizations of a quasi-periodic critical curve with parameters $a$, $b$ can only be periodic if $ab>1$, in which case the following holds. \begin{lemma} A quasi-periodic critical curve with $ab>1$ is closed if and only if $\Phi_1(a,b)$ and $\Phi_2(a,b)$ are rational numbers. \end{lemma} \begin{proof}[Proof of Lemma ] Assume that $a>0$, $a>b$, and $ab>1$. Then, the eigenvalues of $\mathfrak{X}$ are $\lambda_0=0$, $\lambda_{1,\pm}=\pm i\mu$, and $\lambda_{2,\pm}= \pm i \upsilon$. Choose $Y\in GL(5,\C)$, such that \[ \hat{\mathfrak{X}}=Y^{-1} \mathfrak{X} Y=\left( \begin{smallmatrix} 0 & 0 & 0 & 0 & 0 \\ 0 & i\mu & 0 & 0 & 0 \\ 0 & 0 & -i\mu & 0 & 0 \\ 0 & 0 & 0 & i\upsilon & 0 \\ 0 & 0 & 0 & 0 & -i\upsilon \\ \end{smallmatrix} \right) \] and consider the map $V=Y^{-1} \Gamma:\R\to \C^5$. From \eqref{2.2.4}, it follows that the components $v_0,\dots,v_4$ of $V$ satisfy \[ \begin{cases} v_0'=\frac{k'}{k}v_0,\\ v'_1=\left( \frac{k'}{k}-i\frac{k^2-a-b}{k'^2+1}\mu - \frac{k'(k^2-a-b)}{k(k'^2+1)}\mu^2 -i \frac{1}{k'^2+1}\mu^3+\frac{k'}{k(k'^2+1)}\mu^4\right)v_1,\\ v'_2=\left( \frac{k'}{k}+i\frac{k^2-a-b}{k'^2+1}\mu - \frac{k'(k^2-a-b)}{k(k'^2+1)}\mu^2 +i \frac{1}{k'^2+1}\mu^3+\frac{k'}{k(k'^2+1)}\mu^4\right)v_2,\\ v'_3=\left( \frac{k'}{k}-i\frac{k^2-a-b}{k'^2+1}\upsilon - \frac{k'(k^2-a-b)}{k(k'^2+1)}\upsilon^2 -i \frac{1}{k'^2+1}\upsilon^3+\frac{k'}{k(k'^2+1)}\upsilon^4\right)v_3,\\ v'_4=\left( \frac{k'}{k}+i\frac{k^2-a-b}{k'^2+1}\upsilon - \frac{k'(k^2-a-b)}{k(k'^2+1)}\upsilon^2 +i \frac{1}{k'^2+1}\upsilon^3+\frac{k'}{k(k'^2+1)}\upsilon^4\right)v_4. \end{cases} \] Using $k'^2+(k^2-a)(k^2-b)=0$, the above equations take the form \begin{align*} v'_1&=-\frac{kk'+i\mu}{\mu^2-k^2}v_1,\quad v_2'=\frac{-kk'+i\mu}{\mu^2-k^2}v_2,\\ v_3'&=-\frac{kk'+i\upsilon}{\upsilon^2-k^2}v_3,\quad v_4'=\frac{-kk'+i\upsilon}{\upsilon^2-k^2}v_4. \end{align*} Thus $v_a = p_a w_a$, where $p_0,\dots,p_4\in \C$ and $w_0,\dots,w_4$ are the complex-valued functions given by \[ \begin{aligned} w_0&=k, \quad w_1=\sqrt{\mu^2-k^2}e^{i\int \frac{\mu}{\mu^2-k^2}dt}, \quad w_2=\sqrt{\mu^2-k^2}e^{-i\int \frac{\mu}{\mu^2-k^2}dt},\\ w_3&=\sqrt{\upsilon^2-k^2}e^{i\int \frac{\upsilon}{\upsilon^2-k^2}dt},\quad w_4 =\sqrt{\upsilon^2-k^2}e^{-i\int \frac{\upsilon}{\upsilon^2-k^2}dt}. \end{aligned} \] Since $k$ is a non-constant periodic function with minimal period $\omega$, the functions $w_0,\dots,w_4$ are periodic if and only if \[ \Phi_1=\frac{1}{2\pi}\int_0^{\omega}\frac{\mu}{k(t)^2-\mu^2}dt\in \mathbb{Q}, \quad \Phi_2=\frac{1}{2\pi}\int_0^{\omega}\frac{\upsilon}{k(t)^2-\upsilon^2}dt\in \mathbb{Q}, \] which proves the lemma. \end{proof} This proves Theorem . As a corollary, we have the following. \begin{prop} The M\"obius classes of conformal strings are in 1-1 correspondence with the elements of the countable set \[ \Sigma_*=\left\{(a,b)\in \R^2 \,:\, a>0,\, a>b,\, ab>1,\, \Phi_1(a,b),\, \Phi_2(a,b)\in \mathbb{Q}\right\}. \] \end{prop} This Proposition will be used in the proof of Theorem . \section{The proof of Theorem } Theorem asserts that the M\"obius classes of conformal strings are in one-to-one correspondence with the rational points of the complex domain \[ \Omega = \left\{q\in \mathbb{C} \,:\, {1}/{2} < \Re q< {1}/{\sqrt{2}},\,\, \Im q>0,\,\, |q| < {1}/{\sqrt{2}}\right\}. \] The proof of Theorem will follow from Proposition and the next result about the period map $\Phi$ (cf. Definition ). \begin{thm} The period map $\Phi=(\Phi_1,\Phi_2) : \Sigma \to \R^2$ is a real-analytic diffeomorphism onto the domain \[ \widetilde{\Omega}=\left\{(x,y)\in \R^2 \,:\, -1/\sqrt{2}<x<-1/2,\, x^2+y^2<1/2,\, y>0 \right\}. \] \end{thm} The proof of Theorem is based on a detailed study of the analytic properties of $\Phi$. This study is split into several technical results which will take up the remaining part of the section. \subsection{Preparatory material} Note that $\Phi_1$, $\Phi_2 : \Sigma \to \mathbb R$, and hence the period map $\Phi =(\Phi_1,\Phi_2)$, extend analytically to the domain \[ \widetilde{\Sigma}=\left\{(a,b) \,:\, a>0,\, ab \ge 1 \right\}. \] Slightly abusing notation, we will retain the same letters for the extensions of $\Phi_1$ and $\Phi_2$ to $\widetilde{\Sigma}$. Let \begin{equation} E:m\in [0,1)\mapsto \int_0^{\pi/2}\sqrt{1-m\sin^2t} dt\in \R \end{equation} be the elliptic integral of the second kind. Then, $E(m)<K(m)$ and the ratio $E/K$ is a strictly decreasing function from $[0,1)$ onto $(0,1]$. The power series of $K$ and $E$ (cf. , p. 73, and ) are given by \begin{equation} \begin{cases} K(m)=\frac{\pi}{2}\left(1+\frac{1}{4}m+\left(\frac{1\cdot 3}{2\cdot 4}\right)^2m^2 +\left(\frac{1\cdot 3\cdot 5}{2 \cdot 4\cdot 6}\right)^2 m^3+\cdots \right),\\ E(m)=\frac{\pi}{2}\left(1-\frac{1}{4}m-\frac{1}{3}\left(\frac{1\cdot 3}{2\cdot 4}\right)^2m^2 -\frac{1}{5}\left(\frac{1\cdot 3\cdot 5}{2\cdot 4\cdot 6}\right)^2 m^3 +\cdots\right). \end{cases} \end{equation} We also recall the expressions of the derivatives of the complete elliptic integral of the third kind $\Pi$ given in \eqref{thirdkind} (cf. , \S 3.7-3.9, and ), \begin{equation} \begin{cases} \partial_n\Pi|_{(n,m)} = \frac{nE(m)+(m-n)K(m)+(n^2-m)\Pi(n,m)}{2(m-n)(n-1)n},\\ \partial_m\Pi|_{(n,m)} = \frac{E(m)}{2(m-1)(n-m)}+\frac{\Pi(n,m)}{2(n-m)}. \end{cases} \end{equation} \subsection{The derivatives of $\Phi_1$ and $\Phi_2$} Using \eqref{2.2.3} and \eqref{4.3}, we have \begin{equation} \begin{cases} (\mathrm i) & \partial_a\Phi_1|_{(a,b)}=\frac{X_{11}(a,b)E(\frac{a-b}{a}) +Y_{11}(a,b)K(\frac{a-b}{b})}{Z_{11}(a,b)},\\ (\mathrm {ii})&\partial_b\Phi_1|_{(a,b)}=\frac{X_{21}(a,b)E(\frac{a-b}{a}) +Y_{21}(a,b)K(\frac{a-b}{b})}{Z_{21}(a,b)},\\ (\mathrm {iii})&\partial_a\Phi_2|_{(a,b)}=\frac{X_{12}(a,b)E(\frac{a-b}{a}) +Y_{12}(a,b)K(\frac{a-b}{b})}{Z_{12}(a,b)},\\ (\mathrm {iv})&\partial_b\Phi_2|_{(a,b)}=\frac{X_{22}(a,b)E(\frac{a-b}{a}) +Y_{22}(a,b)K(\frac{a-b}{b})}{Z_{22}(a,b)},\\ \end{cases} \end{equation} where the coefficients $X_{ij}(a,b)$, $Y_{ij}(a,b)$, and $Z_{ij}(a,b)$ are given by \[ \begin{cases} X_{11}(a,b)=\sqrt{2}\left(2 z(a,b)-a^2b+a\left(4+ z(a,b)b\right)+b\left( 4+ z(a,b)b\right)\right),\\ Y_{11}(a,b)=-2\sqrt{2}\left(a+ z(a,b)-ab^2 +b\left(3+b( z(a,b)+b) \right) \right),\\ Z_{11}(a,b)=\pi\sqrt{a} z(a,b)(a-b)(a-b- z(a,b))\left(a+b+ z(a,b) \right)^{3/2},\\ X_{21}(a,b)=a(2b+ z(a,b)),\\ Y_{21}(a,b)=-b(a+b+ z(a,b)), \end{cases} \] and by \[ \begin{cases} Z_{21}(a,b)=\sqrt{2}\pi b(a-b) z(a,b)\sqrt{a(a+b+ z(a,b))},\\ X_{12}(a,b)=\sqrt{2}\left(2 z(a,b)+a^2b+a(-4+b z(a,b))-b(4-b z(a,b)+b^2)\right),\\ Y_{12}(a,b)=2\sqrt{2}\left( a- z(a,b)-ab^2+b(3-b z(a,b)+b^2)\right),\\ Z_{12}(a,b)=\pi\sqrt{a} z(a,b)(a-b)\left(a-b+ z(a,b)\right)\left( a+b- z(a,b)\right)^{3/2},\\ X_{22}(a,b)=a\left( z(a,b)-2b\right),\\ Y_{22}(a,b)=b\left(a+b- z(a,b)\right),\\ Z_{22}(a,b)=\sqrt{2}\pi(a-b)b z(a,b)\sqrt{a\left(a+b- z(a,b) \right)}. \end{cases} \] In the formulae above, $z(a,b)$ stands for $\sqrt{4+(a-b)^2}$. These formulae have been derived with the help of the software {\sl Mathematica}. We now prove the following. \begin{prop} The partial derivatives of $\Phi_1$ and $\Phi_2$ are strictly positive on \begin{equation} \Sigma'=\left\{(a,b)\,:\, a>1,\, ab>1,\, b\le a\right\}. \end{equation} \end{prop} \begin{proof} We begin by proving the following. \begin{lemma} $\partial_a\Phi_1|_{(a,b)}>0$, for every $(a,b)\in \Sigma'$. \end{lemma} \begin{proof}[Proof of Lemma ] Consider $a>1$ and $b\in (0,a]$. Let $b=(1-m)a$, with $m\in [0,1)$, and let $\tilde{z}(a,m):=\sqrt{4+m^2a^2}$. From \eqref{4.4}$(\mathrm {i})$, we have \[ \partial_a\Phi_1|_{(a,(1-m)a)}=\frac{\tilde{X}_{11}(a,m)E(m) +\tilde{Y}_{11}(a,m)K(m)}{\tilde{Z}_{11}(a,m)}, \] where \[ \begin{cases} \tilde{X}_{11}(a,m)=\sqrt{2}\left(-2\tilde{z}(a,m)+(2-m)a \left((1-m)a(ma-\tilde{z}(a,m))-4)\right) \right),\\ \tilde{Y}_{11}(a,m)=2\sqrt{2}\left((4\!-\!3m)a\!-\!(1\!-\!m)^2ma^3\!+\!\tilde{z}(a,m) \!+\!(1\!-\!m)^2a^2\tilde{z}(a,m) \right),\\ \tilde{Z}_{11}(a,m)=\pi ma\tilde{z}(a,m)(\tilde{z}(a,m)-am)\left((2-m)a+\tilde{z}(a,m)\right)^{3/2}. \end{cases} \] First, we prove that $\partial_a\Phi_1|_{(a,(1-m)a)}>0$, for every $m\in (0,1)$. If $m\in (0,1)$, then $\tilde{Z}_{11}(a,m)>0$. So, if we set $f_m(a)=\tilde{X}_{11}(a,m)/\tilde{Y}_{11}(a,m)$, it suffices to show that $f_m(a)+K(m)/E(m)>0$. The derivative of $f_m(a)$ with respect to $a$ is \[ f_m'(a)=- \frac{2m}{4(2\!-\!m)a\!+\!(2\!-\!m)m^2a^3\!+\!2\tilde{z}(a,m)\!+\!(2\!-\!(2\!-\!m)m)a^2\tilde{z}(a,m)}. \] Thus, $f_m$ is a strictly decreasing function such that $\lim_{a\to \infty} f_m(a) = -(2+m)/2$. This implies that \[ f_m(a)+\frac{K(m)}{E(m)}>\lim_{a\to \infty} f_m(a) + \frac{K(m)}{E(m)} = \frac{K(m)}{E(m)}-\frac{2+m}{2}. \] Deriving the function $\tilde{f}(m)=K(m) E(m)^{-1}-(2+m)/2$, we find \[ \tilde{f}'|_m = \frac{(E(m)-K(m))^2}{2mE(m)^2}+\frac{m E(m)^2}{2(1-m)E(m)^2}>0. \] Then, $\tilde{f}$ is strictly increasing and $\tilde{f}(m)>\lim_{m\to 0}\tilde{f}=0$, for every $m\in (0,1)$. This implies $\partial_a\Phi_1|_{(a,(1-m)a)}>0$, for every $m\in (0,1)$. \vskip0.1cm Next, we prove that $\partial_a\Phi_1|_{(a,a)}>0$. Using the power series expansions of $K(m)$ and $E(m)$, we find \[ \begin{cases} \tilde{X}_{11}(a,m)E(m)+\tilde{Y}_{11}(a,m)K(m)=\frac{\pi}{\sqrt{2}}(2a+ \tilde{z}(a,m))m+ o(m),\\ \tilde{Z}_{11}(a,m)=\pi a \sqrt{a}\tilde{z}(a,m)^2((2-m)a+\tilde{z}(a,m))^{3/2}m+o(m). \end{cases} \] From this, we have \[ \partial_a\Phi_1|_{(a,a)}=\lim_{m\to 0}\partial_a\Phi_1|_{(a,(1-m)a)} = \frac{1}{8a^{3/2}\sqrt{1+a}}>0. \] \end{proof} By similar arguments as in the proof of Lemma , one can prove that, for every $(a,b)\in \Sigma'$, $\partial_b\Phi_1|_{(a,b)}>0$, $\partial_a\Phi_2|_{(a,b)}>0$, and $\partial_b\Phi_2|_{(a,b)}>0$, which completes the proof of Proposition . \end{proof} \subsection{The image of $\Phi$} Let $\Sigma$ be as in \eqref{Sigma}. We will now prove the following. \begin{prop} The Jacobian of $\Phi$ is strictly positive on $\Sigma$. In particular, the image $\Phi(\Sigma)$ is a connected open set and $\Phi:\Sigma\to \Phi(\Sigma)$ is a local diffeomorphism. \end{prop} \begin{proof} Let $\Phi_*$ be the matrix of the differential of $\Phi$. Using \eqref{4.4}, the Jacobian $J\Phi(a,b)$ of $\Phi$ at $(a,b)$ is given by \[ J\Phi(a,b) =\mathrm{det}(\Phi_*)|_{(a,b)}=\frac{E(\frac{a-b}{b})\left(-2aE(\frac{b-a}{a}) +(a+b)K(\frac{a-b}{a}) \right)}{2\pi^2a(a-b)b \sqrt{(4+(a-b)^2)(ab-1)}}. \] Let $w(a,b)$ denote the numerator of the right hand side. If we let $b=(1-m)a$, where $m\in (0,1)$ and $a>1/\sqrt{1-m^2}$, then \[ w(a,(1-m)a)=a(-2E(m)+(2-m)K(m))>0,\quad \forall\, m\in (0,1), \] which implies the required result. \end{proof} We adopt the following conventions: \begin{itemize} \item $\Phi=(\Phi_1,\Phi_2)$ is considered as a function defined on $\Sigma$; \item $\widetilde{\Phi}$ denotes the analytic extension of $\Phi$ to $\widetilde{\Sigma}=\{(a,b) \,:\, a>0,\, ab \ge 1\}$; \item $\widehat{\Phi}$ denotes the restriction of $\tilde{\Phi}$ to the closure $\bar{\Sigma}$ of $\Sigma$. \end{itemize} Next, we prove the following. \begin{prop} The mapping $\Phi$ is a real-analytic local diffeomorphism onto $\widetilde{\Omega}$. \end{prop} \begin{proof} By Proposition , it suffices to prove that $\widetilde{\Omega}$ is the image of $\Phi$. The boundary of $\Sigma$ consists of the simple arcs \[ \partial_+\Sigma=\{(a,a) \,:\, a\ge 1\},\quad \partial_{-}\Sigma =\{(a,1/a) \,:\, a\ge 1\} \] which intersect at the point $V=(1,1)$, while the boundary of $\widetilde{\Omega}$ is made of the three simple arcs \[ \begin{aligned} & \partial_{-}\widetilde{\Omega}=\{(a,0) \,:\, a\in [-1/\sqrt{2},-1/2]\},\\ & \partial_0 \widetilde{\Omega} =\{(-1/2,b) \,:\, b\in [0,1/2]\},\\ &\partial_{+}\widetilde{\Omega} =\{(a,\sqrt{a^2-1/2})\,:\, a\in [-1/\sqrt{2},-1/2]\} \end{aligned} \] with vertices $Q_1=(-1/\sqrt{2},0)$, $Q_2=(-1/2,0)$ and $Q_3=(-1/2,-1/2)$, respectively. The restriction $ \Phi_{+}$ of $\widehat{\Phi}$ to $\partial_+\Sigma$ is given by \[ \Phi_{+}: (a,a) \in \partial_+\Sigma \mapsto \frac{1}{2} \left(-\sqrt{(a+1)a^{-1}}, \sqrt{(a-1)a^{-1}}\right) \in \partial_{+} \widetilde{\Omega}\setminus\{Q_3\}. \] It is then a diffeomorphism of $\partial_+\Sigma $ onto $\partial_{+}\widetilde{\Omega}\setminus\{Q_3\}$. Similarly, the restriction $\Phi_{-}$ of $\widehat{\Phi}$ to $\partial_{-}\Sigma$ is the diffeomorphism onto $\partial_{-}\widetilde{\Omega}\setminus\{Q_2\}$ given by \[ \Phi_{-}:(a,1/a)\in \partial_{-}\Sigma \mapsto \left(\!-\frac{1}{\pi}\sqrt{1\!+\!a^2}\Pi(1\!-\!a^2,(a^2\!-\!1)a^{-2}), 0\right)\!\in \partial_{-}\widetilde{\Omega}\setminus\{Q_2\}. \] Then $\widehat{\Phi}$ maps the boundary of $\Sigma$ to the boundary of $\widetilde{\Omega}$. Next, we show that the image of $\Phi$ is contained in $\widetilde{\Omega}$. For, fix $a>1$ and consider the curve defined by $\phi_a (b)=\Phi(a,b)$, for every $b\in (1/a,a)$. From Proposition , we know that the components $\phi^1_a$ and $\phi^2_a$ of $\phi_a$ are increasing functions. This implies \[ \begin{split} -\frac{1}{\sqrt{2}} <\phi_a^1(a^{-1})&=-\pi^{-1} \sqrt{1+a^2}\Pi(1-a^2,(a^2-1) a^{-2})\\ &<\phi_a^1(b)<\phi_a^1(a)= -\frac{1}{2}\sqrt{(a+1) a^{-1}}<-\frac{1}{2}, \end{split} \] and \[ 0=\phi_a^2(1/a)<\phi_a^2(b)<\phi_a^2(a)=\frac{1}{2}\sqrt{(a-1) a^{-1}}. \] Combining these two inequalities, we find \[ 0<\phi_a^1(b)^2+\phi_a^2(b)^2<\phi_a^1(a)^2+\phi_a^2(a)^2=1/2. \] This shows that the image of $\Phi$ is contained in $\widetilde{\Omega}$. To prove equality, we need the following technical lemma. \begin{lemma} Let $\{(a_n,b_n)\}_{n\in \mathbb{N}}\subset \Sigma$ be a sequence such that $a_n\to \infty$. Then, $\lim_{n\to \infty} \Phi_1(a_n,b_n)=-1/2$. \end{lemma} \begin{proof} The trajectory of $\phi_a$ is the graph of an increasing function \[ f_a:(\Phi_1(a,a^{-1}),\Phi_1(a,a))\to \R. \] Given the sequence $\{(a_n,b_n)\}_{n\in \mathbb{N}}$, we write $\Phi(a_n,b_n)=(t_n, f_{a_n}(t_n))$, where $t_n$ is an element of the open interval $(\Phi_1(a_n,1/a_n),\Phi_1(a_n,a_n))$. From the limits \[ \begin{aligned} \lim_{a\to \infty} \widehat{\Phi}_1(a,a^{-1}) &= - \lim_{a\to \infty} \frac{1}{\pi}\sqrt{1+a^2}\Pi(1-a^2,(a^2-1)a^{-2}) = -\frac{1}{2},\\ \lim_{a\to \infty} \widehat{\Phi}_1(a,a) &= - \lim_{a\to \infty} \frac{1}{2}\sqrt{(1+a)a^{-1}}=-\frac{1}{2}, \end{aligned} \] we have \[ -\frac{1}{2} = \lim_{n\to \infty} t_n = \lim_{n\to \infty}\Phi_1(a_n,b_n). \] \end{proof} By Proposition , $\Phi:\Sigma\to \widetilde{\Omega}$ is a real-analytic local diffeomorphism onto its image, and hence an open map. Consequently, $X=\widetilde{\Omega}\setminus\Phi(\Sigma)$ is a closed subset of $\widetilde{\Omega}$. The proof is complete if we show that $\widetilde{\Omega}\setminus\Phi(\Sigma)$ is also open. Suppose it is not open. Then, there is a point $Q\in X$, such that any open disk $D(Q,1/n)$ centered at $Q$ with radius $1/n$, $n\in \mathbb{N}$, intersects $\Phi(\Sigma)$. For every $n\in \mathbb{N}$, choose $Q_n\in D(p,1/n)\cap \Phi(\Sigma)$ and $P_n\in \Sigma$, such that $\Phi(P_n)=Q_n$. Two possibilities may occur: either $\{P_n\}$ is bounded, or else $\{P_n\}$ is unbounded. In the first case, we may assume that $\{P_n\}$ converges to a limit point $P$. By construction, $P$ belongs to $\bar{\Sigma}$. If $P$ is an element of $\Sigma$, then $Q=\Phi(P)$, which contradicts the fact that $Q\notin \Phi(\Sigma)$. So, $P$ is an element of the boundary $\partial \Sigma$. This implies that $Q=\widehat{\Phi}(P)$. On the other hand, $\widehat{\Phi}$ maps the boundary of $\Sigma$ onto the boundary of $\widetilde{\Omega}$. Consequently, $Q$ would be an element of $\partial \widetilde{\Omega}$, which is absurd, since $Q\in \widetilde{\Omega}$. If $\{P_n\}$ is unbounded, by Lemma , the sequence made up with the abscissae of the points $Q_n=\Phi(P_n)$ converges to $-1/2$. So, the first coordinate of $Q$ would be equal to $-1/2$ and hence $Q\notin \widetilde{\Omega}$, which is a contradiction. This concludes the proof of Proposition . \end{proof} \subsection{Injectivity of $\Phi$} By Proposition , we know that the first order partial derivatives of $\Phi_1$ and $\Phi_2$ are strictly positive on $\Sigma'$. Then, there is an open neighborhood $W$ of $\Sigma'$ such that the first order partial derivatives of $\Phi_1$ and $\Phi_2$ are positive on $W'=W\cap \mathrm{Int}(\widetilde{\Sigma})$. On this set, consider the nowhere vanishing vector fields (see Figure ) \begin{equation} U_1=\left(1,-\frac{\partial_a \Phi_1}{\partial_b\Phi_1}\right),\quad U_2=\left(1,-\frac{\partial_a\Phi_2}{\partial_b\Phi_2}\right). \end{equation} By construction, the trajectories of the integral curves of $U_1$ and $U_2$ are graphs of strictly decreasing functions, and hence they intersect $\partial_+ \Sigma$ in at most one point. In addition, for every $Q=(x,y)\in \widetilde{\Omega}$, we have \begin{itemize} \item the connected components of the level curve $\mathcal{V}_1(x)=\Phi_1^{-1}(x)\cap \Sigma$ are contained in the intersection of a trajectory of $U_1$ with $\Sigma$. \item the connected components of the level curve $\mathcal{V}_2(y)=\Phi_2^{-1}(y)\cap \Sigma$ are contained in the intersection of a trajectory of $U_2$ with $\Sigma$. \end{itemize} We now prove the following. \begin{lemma} The level curve $\mathcal{V}_1(x')$ is connected, for every $Q'=(x',y')\in \widetilde{\Omega}$. \end{lemma} \begin{proof} Consider the arcs $\partial_+\widetilde{\Omega}'=\partial_+\widetilde{\Omega}\setminus\{Q_1,Q_3\}$, $\partial_-\widetilde{\Omega}'=\partial_-\widetilde{\Omega}\setminus\{Q_1,Q_2\}$, $\partial_+ \Sigma'\!=\!\partial_+\Sigma \setminus\!\{V\}$ and $\partial_- \Sigma'\!=\!\partial_-\Sigma \setminus\!\{V\}$. From the proof of Proposition~, we know that the maps $\widehat{\Phi}|_{\partial_+ \Sigma'}:\partial_+ \Sigma'\to \partial_+\Omega'$ and $\widehat{\Phi}|_{\partial_- \Sigma'}:\partial_- \Sigma'\to \partial_-\Omega'$ are real-analytic diffeomorphisms. For every $Q'=(x',y')\in \widetilde{\Omega}$, let $Q'_{\pm}$ denote the intersections of $\partial_{\pm}\widetilde{\Omega}'$ with the line $x=x'$ and let $P'_+=(a'_+,a'_+)$ and $P'_- = (a'_-, 1/a'_{-})$ be the points of $\partial_{\pm} \Sigma'$, such that $\widehat{\Phi}(P'_{\pm})=Q'_{\pm}$. Note that $\widehat{\Phi}_1^{-1}(x')=\Phi_1^{-1}(x')\cup\{P'_+,P'_-\}$. \vskip0.1cm \noindent $\bullet$ First, we prove that $\widehat{\Phi}_1^{-1}(x')$ is compact. By continuity, $\widehat{\Phi}_1^{-1}(x')$ is closed. Suppose that $\widehat{\Phi}_1^{-1}(x')$ is unbounded. Then, there is a sequence $\{P_n\}=\{(a_n,b_n)\}\subset \Phi_1^{-1}(x')$, such that $\lim_{n\to \infty}a_n=+\infty$. By Lemma , $\{\Phi_1(P_n)\}$ converges to $-1/2$. On the other hand, the points $P_n$ belong to $\mathcal{V}_1(x')$, and hence $x'$, which is the limit of the sequence $\{\Phi_1(P_n)\}$, must be equal to $-1/2$. But this is absurd, since $(x',y')\in \widetilde{\Omega}$ and the abscissae of the points in $\widetilde{\Omega}$ are different from $-1/2$. This proves the compactness of $\widehat{\Phi}_1^{-1}(x')$. \vskip0.1cm \noindent $\bullet$ Next, we prove that the points $P'_{\pm}$ belong to the boundary of any connected component of the level curve $\mathcal{V}_1(x')$. Let $C$ be a connected component of $\mathcal{V}_1(x')$. We know that $C$ is the graph of a strictly decreasing function and that such a function is bounded, by the compactness of $\widehat{\Phi}_1^{-1}(x')$. Since $C$ is an embedded curve, its boundary consists of two distinct points. On the other hand, $C\subset \Phi_1^{-1}(x')\subset \widehat{\Phi}_1^{-1}(x')=\Phi_1^{-1}(x')\cup\{P'_+,P'_- \}$. Using the fact that $\Phi_1^{-1}(x')$ is an embedded curve, we have that $P'_+$ and $P'_-$ are the two boundary points of $C$. \vskip0.1cm \noindent $\bullet$ To complete the proof, we show that two connected components of the level curves intersect each other. The vector field $U_1$ does not vanishes at the point $P'_+$ and $P'_+$ belongs to the boundary of $C$. Therefore, $C$ is contained in the trajectory of $U_1$ passing through $P'_+$. Choose local coordinates $(u,v)$ defined on a cubical open neighborhood $A$ of the point $P'_+$, such that $U_1|_{A}=\partial_u$ and $v(P'_+)=0$. In these coordinates, the intersection $C\cap A$ consists of the points $P\in A$ such that $v(P)=0$ and $0<u(P)<\epsilon$, for some $\epsilon>0$. So, if $C$ and $C'$ are two connected components of $\mathcal{V}_1(x')$, there exist $\epsilon,\epsilon'>0$, such that $C\cap A = \{P\in A \,:\, v(P)=0,\, 0<u(P)<\epsilon\}$ and $C'\cap A = \{P\in A \,:\, v(P)=0,\, 0<u(P)<\epsilon'\}$. This implies that $C \cap C'\neq \emptyset$. \end{proof} A similar result also holds for the other family of level curves. \begin{lemma} The level curves $\mathcal{V}_2(y')$ are connected, for every $Q'=(x',y')\in \widetilde{\Omega}$. \end{lemma} \begin{proof} Let $C$ be any connected component of $\mathcal{V}_2(y')$. Fix $(a,b)\in C$ and take a strictly decreasing function $u : I \to \R$ whose graph coincides with $C$. Such a function is defined on an open interval $I$ of the form $(a-\epsilon_1,a+\epsilon_2)$, where $\epsilon_1,\epsilon_2>0$ and $a-\epsilon_1>1$. Let $\{a_n\}\subset (a-\epsilon_1,a)$ be a decreasing sequence, such that $\lim_{n\to \infty}a_n=a^*=a-\epsilon_1$. By construction, $\{u(a_n)\}$ is an increasing sequence and $(a_n,u(a_n))\in C\subset \Sigma$, for every $n$. In particular, $0<a_n^{-1}<u(a_n)<a_n<a$, which implies that $\{u(a_n)\}$ is bounded. By possibly taking a subsequence, we may assume that $\{u(a_n)\}$ converges to a limit point $b^*$. Since $\{(a_n,u(a_n))\}$ converges to $(a^*,b^*)$, the point $(a^*,b^*)$ belongs to $\partial C \cap \bar{\Sigma}$. \vskip0.1cm \noindent $\bullet$ We prove that $(a^*,b^*)$ is an element of $\partial\Sigma$. By contradiction, suppose that $(a^*,b^*)\in \Sigma$. Then, $C$ is contained in the trajectory, $C'$, of the vector field $U_2|_{\Sigma}$ passing through $(a^*,b^*)$. By definition, $(a^*,b^*)\in C'$ and $(a^*,b^*)\notin C$. Therefore, $C'$ is a connected subset of $\mathcal{V}_2(y')$ which contains properly $C$. This is absurd since $C$ is a connected component of $\mathcal{V}_2(y')$. \vskip0.1cm \noindent By construction, since $1/a^*<u(a_n)$ and ${u(a_n)}$ is an increasing sequence, $b^*\neq 1/a^*$. This shows that $(a^*,b^*)$ does not belong to $\partial_-\Sigma$. Therefore, $(a^*,b^*)$ is an element of $\partial_+\Sigma$ and hence $b^*=a^*$ and $a^*>1$. The point $(a^*,a^*)$ is independent of the choice of the connected component. Otherwise, the restriction of $\widetilde{\Phi}$ to $\partial_+\Sigma$ could not be injective. We are now in the same situation as in the last part of the proof of the previous lemma, i.e., the point $(a^*,a^*)$ belongs to the boundary of any connected component of the level curve $\mathcal{V}_2(y')$. By arguing as above, we deduce the result. \end{proof} We are now ready to prove the last ingredient for the proof of Theorem~. \begin{prop} The mapping $\Phi$ is injective. \end{prop} \begin{proof} It suffices to show that $\#\left(\mathcal{V}_1(x)\cap \mathcal{V}_2(y)\right)=1$, for every $(x,y)\in \widetilde{\Omega}$. By contradiction, suppose the existence of $(x,y)\in \widetilde{\Omega}$, such that $\sharp(\mathcal{V}_1(x)\cap \mathcal{V}_2(y))>1$. Let $(a,b)$ and $(a_1,b_1)$ be two distinct elements of $\Sigma$, such that $\Phi(a_1,b_1)$ $=$ $\Phi(a,b)$ $=$ $(x,y)$. From the above discussions, we know that the level curves $\mathcal{V}_1(x)$ and $\mathcal{V}_2(y)$ are connected and graphs of two strictly decreasing functions, denoted by $u$ and $v$, respectively. The domain of definition is an open interval $I\subset (1,+\infty)$, containing $a$ and $a_1$. By construction, $u(a)=v(a)=b$, $u(a_1)=v(a_1)=b_1$, with $a\neq a_1$. On the other hand, $\mathcal{V}_1(x)$ and $\mathcal{V}_2(y)$ are contained in the trajectories of the vector fields $U_1$ and $U_2$, respectively. From this, we have \[ u'(t)=-\frac{\partial_a\Phi_1}{\partial_b\Phi_1}\bigg|_{(t,u(t))},\quad v'(t)=-\frac{\partial_a\Phi_2}{\partial_b\Phi_2}\bigg|_{(t,v(t))},\quad \forall t\in I. \] Now, Proposition and Proposition imply \begin{equation} \frac{\partial_a\Phi_1}{\partial_b\Phi_1}\bigg|_{(\alpha,\beta)}- \frac{\partial_a\Phi_2}{\partial_b\Phi_2}\bigg|_{(\alpha,\beta)}>0, \quad \forall (\alpha,\beta)\in \Sigma. \end{equation} Then, the function $h=v-u$ satisfies $h(a)=h(a_1)=0$, $h'(a)>0$ and $h'(a_1)>0$. This implies the existence of $a_2\in I$, different from $a$ and $a_1$, such that $h(a_2)=0$ and $h'(a_2)\le 0$. Consequently, the Jacobian of $\Phi$ is non positive at $(a_2,b_2)$ $=$ $(a_2,u(a_2))=(a_2,v(a_2))\in \Sigma$, contrary to Proposition~. \end{proof} \section{The proof of Theorem } Theorem asserts that any M\"obius class of conformal strings is represented by a model conformal string. We will begin the proof by describing the model strings. \subsection{The symmetrical configuration of a string} We have just proved that the M\"obius classes of conformal strings are in one-to-one correspondence with the {\it moduli}, i.e., the elements of the countable set \begin{equation} \Omega_*=\left\{(q_1,q_2)\in \mathbb{Q}^2 \,:\, 1/2<q_1<1/\sqrt{2},\, q_2>0,\, q_1^2+q_2^2< 1/2\right\}. \end{equation} Thus, for every $q=(q_1,q_2)\in \Omega_*$, there is a unique $(a,b)\in \Sigma$ such that $\Phi_1(a,b)=-q_1$ and $\Phi_2(a,b)=q_2$. For every $(q_1,q_2)\in \Omega_*$, let \begin{equation} \Theta_1(t)=\int_0^t\frac{\mu}{\mu^2-k(u)^2}du,\quad \Theta_2(t)=\int_0^t\frac{\upsilon}{\upsilon^2-k(u)^2}du \end{equation} and \begin{equation} r(t)=\sqrt{\mu^2-\upsilon^2}k(t)+\upsilon\sqrt{\mu^2-k(t)^2} \cos\Theta_1(t), \end{equation} where $a,b$ are the parameters of $q$ and $k$, $\mu$, $\upsilon$ stand for $k_{a,b}$, $\mu(a,b)$ and $\upsilon(a,b)$, respectively. Note that $(q_1,q_2)$ and $(a,b)$ are related by \begin{equation} q_1= \frac{1}{2\pi}\int_0 ^{\omega} \Theta_1'(t) dt,\quad q_2=-\frac{1}{2\pi}\int_0 ^{\omega} \Theta_2'(t)dt, \end{equation} where $\omega=\omega(a,b)$ is the minimal period of $k$. \begin{defn} The {\it symmetrical configuration} of the conformal strings with modulus $q=(q_1,q_2)$ is the parametrized curve $\gamma_{q}=(x,y,z) : \R\to \R^3$, defined by \begin{equation} \begin{cases} x(t)=\frac{\sqrt{2}}{r(t)}\mu\sqrt{k(t)^2-\upsilon^2}\cos\Theta_2(t),\\ y(t)=\frac{\sqrt{2}}{r(t)}\mu\sqrt{k(t)^2-\upsilon^2}\sin\Theta_2(t),\\ z(t)=\frac{\sqrt{2}}{r(t)}\upsilon\sqrt{\mu^2-k(t)^2}\sin\Theta_1(t).\\ \end{cases} \end{equation} \end{defn} Theorem is now a direct consequence of the following. \begin{thm} Let $q=(q_1,q_2)$ be any element of $\Omega_*$ and $(a,b)\in \Sigma_*$ be the corresponding parameters. Any conformal string with parameters $(a,b)$ is conformally equivalent to the symmetrical configuration $\gamma_{q}$. \end{thm} \begin{proof} Consider the unique conformal parametrization $\gamma :\R\to \R^3$ of a conformal string with parameters $a$, $b$ whose Vessiot frame $\mathrm{F}$ satisfies the initial condition $\mathrm{F}(0)=\mathrm{Id}$. It suffices to prove that $\gamma$ is conformally equivalent to $\gamma_q$. By Lemma , the canonical lift $\Gamma:\R\to \mathcal{L}_+$ of $\gamma$ takes the form $\tilde{Y} W$, where $\tilde{Y}\in GL(5,\C)$ and $W={}^t\!(w_0, \dots, w_4)$ is the $\C^5$-valued map defined by \begin{equation} \begin{aligned} w_0&=k,\quad w_1=\sqrt{\mu^2-k^2}e^{i\Theta_1(t)},\quad w_2=\sqrt{\mu^2-k^2}e^{-i\Theta_1(t)},\\ w_3&=\sqrt{\upsilon^2-k^2}e^{i\Theta_2(t)},\quad w_4=\sqrt{\upsilon^2-k^2}e^{-i\Theta_2(t)}, \end{aligned} \end{equation} On the other hand, the curve $\widetilde{\Gamma}={}^t\!(\widetilde{\gamma}_0, \dots, \widetilde{\gamma}_4) :\R\to \mathcal{L}_+$, defined by \begin{equation} \begin{cases} \widetilde{\gamma}_0(t)=\frac{1}{\sqrt{2}}\left( \sqrt{\mu^2-\upsilon^2}k(t)-\upsilon\sqrt{\mu^2-k(t)^2}\cos\Theta_1(t)\right),\\ \widetilde{\gamma}_1(t)=\mu\sqrt{k(t)^2-\upsilon^2}\cos\Theta_2(t),\\ \widetilde{\gamma}_2(t)=\mu\sqrt{k(t)^2-\upsilon^2}\sin\Theta_2(t),\\ \widetilde{\gamma}_3(t)=\upsilon\sqrt{\mu^2-k(t)^2}\sin\Theta_1(t),\\ \widetilde{\gamma}_4(t)=\frac{1}{\sqrt{2}}\left( \sqrt{\mu^2-\upsilon^2}k(t)+\upsilon\sqrt{\mu^2-k(t)^2}\cos\Theta_1(t)\right), \end{cases} \end{equation} is a null lift of $\gamma_{q}$. From \eqref{5.5.a} and \eqref{5.5}, it follows that $W=\mathrm{Z} \widetilde{\Gamma}$, for some $\mathrm{Z}\in GL(5,\C)$. Consequently, $\Gamma=\mathrm{L} \widetilde{\Gamma}$, for a suitable $\mathrm{L}\in GL(5,\C)$, which yields $\mathrm{F}=\mathrm{L} \mathrm{F}_{q}$, where $\mathrm{F}_{q}$ is the Vessiot frame along $\gamma_{q}$. Thus $\mathrm{L}\in G_+$, which implies that $\gamma$ and $\gamma_{q}$ are equivalent to each other. \end{proof} \begin{remark} The map $\Phi$ can be inverted by numerical methods. Once we know the parameters $a$ and $b$ which correspond to the modulus $q$, we can use~\eqref{5.4} to find the explicit parametrization of the symmetrical configuration $\gamma_{q}$. \end{remark} \section{The proof of Theorem } Let $\Omega_*$ be the countable set introduced in \eqref{Omega*}. For every $q=(q_1,q_2)\in \Omega_*$, let $q_1=m_1/n_1$ and $q_2=m_2/n_2$, with $(m_1,n_1)=(m_2,n_2)=1$. Let $n$ denote the least common multiple of $n_1$ and $n_2$. Then, $n=h_1n_1$ and $n=h_2n_2$, where $h_1$ and $h_2$ are two positive coprime integers. Let $\gamma_q$ be the symmetrical configuration corresponding to $q$ (cf. Definition ). We are now ready to prove Theorem asserting that (1) $n$ is the order of the symmetry group of $\gamma_q$ and (2) $m_1h_1$ and $m_2h_2$ are, respectively, the linking numbers of $\gamma_q$ with the Clifford circle $\mathcal{C}=\left\{(x,y,0) \in\R^3 : x^2+y^2=2\right\}$ and with the $z$-axis. \begin{proof}[Proof of Theorem ] The curve $\gamma_q$ is a real-analytic closed curve with positive chirality (cf. Section , Definition ). Therefore, the symmetry group of $\gamma_q$ is generated by the monodromy of $\gamma_q$. By construction, the canonical null lift $\Gamma:\R\to \mathcal{L}_+$ is as in \eqref{5.5} and the first conformal curvature is a strictly positive periodic function, with minimal period $\omega$. Using~\eqref{5.5}, we find $\Gamma_{q}(t+\omega)=R(2\pi q_2,2\pi q_1) \Gamma_{q}(t)$, where $R(2\pi q_2,2\pi q_1)$ is as in \eqref{1.1.6}. Therefore, $R(2\pi q_2,2\pi q_1)$ is the monodromy of $\gamma_q$. This implies that the symmetry group of $\gamma_q$ is generated by $R(2\pi q_2,2\pi q_1)$. Let $[z]$ denote the $z$-axis with the downward orientation induced by the parametrization $\alpha(s)=(0,0,-s)$, and let $[\mathcal{C}]$ be the Clifford circle equipped with the orientation induced by the rational parametrization \[ \beta : t\in \R \mapsto (-\sqrt{2}(t^2-2)/(t^2+2), 4t/(2+t^2), 0)\in \R^3. \] Now we compute the Gauss linking integrals $\mathrm{lk}(\gamma_q,[z])$ and $\mathrm{lk}(\gamma_q,[\mathcal{C}])$. If $\gamma_q(t) = (x(t),y(t),z(t))$, the Gauss linking integral $\mathrm{lk}(\gamma_q,[z])$ is given by \[ \begin{split} \mathrm{lk}(\gamma_q,[z])&=\frac{1}{4\pi} \int_{[\gamma_q]}\int_{z} \left\langle \frac{\gamma_q-\alpha}{\|\gamma_q-\alpha\|^3}, d\gamma_q\times d\alpha\right \rangle\\ &=-\frac{1}{4\pi} \int_{0}^{n\omega}\left(\int_{-\infty}^{+\infty}\left(\frac{x(t)y'(t)-x'(t)y(t)}{\left[x(t)^2 +y(t)^2+(z(t)+ s)^2\right]^{3/2}}\right)ds\right)dt\\ &=-\frac{1}{2\pi} \int_0^{n\omega}\frac{x(t)y'(t)-x'(t)y(t)}{x(t)^2+y(t)^2}dt\\ &= -\frac{n}{2\pi}\int_0^{\omega}\Theta_2'(t)dt = n\Phi_2= n q_2 = h_2m_2. \end{split} \] To compute the linking integral of $\gamma_q$ with the Clifford circle we consider the o\-rienta\-tion-\-pre\-serving conformal involution \begin{align} \Psi:(x,y,z)\in \R^3&\mapsto \frac{1}{x^2+y^2+z^2+2\sqrt{2}x+2}\\ \notag &\quad\times\left(x^2+y^2+z^2+2\sqrt{2},4z,4y\right)\in \R^3. \end{align} This map takes $(-\sqrt{2},0,0)$ to the point at infinity and exchanges the roles of the two axes of symmetry, i.e., $\Phi\circ \beta=-\alpha$. A direct computation shows that the parametric equations of $\gamma_* :=\Phi\circ \gamma_q$ are \[ \begin{cases} x_*(t)=\frac{\sqrt{2}}{r^*(t)}\upsilon\sqrt{\mu^2-k(t)^2}\cos \Theta_1(t),\\ y_*(t)=\frac{\sqrt{2}}{r^*(t)}\upsilon\sqrt{\mu^2-k(t)^2}\sin \Theta_1(t),\\ z_*(t)=\frac{\sqrt{2}}{r^*(t)}\mu\sqrt{k(t)^2-\upsilon^2}\sin \Theta_2(t), \end{cases} \] where $r^*(t)=\sqrt{\mu^2-\upsilon^2}k(t)+\mu\sqrt{k(t)^2-\upsilon}\cos \Theta_2(t) $. We then have \[ \begin{split} \mathrm{lk}(\gamma_q,[\mathcal{C}])&=-\mathrm{lk}(\gamma_*,[z])= -\frac{1}{4\pi} \int_{[\gamma_*]}\int_{z} \left\langle \frac{\gamma_*-\alpha}{\|\gamma_*-\alpha\|^3}, d\gamma_*\times d\alpha\right\rangle\\ &=\frac{1}{4\pi} \int_{0}^{n\omega}\left(\int_{-\infty}^{+\infty} \left(\frac{x_*(t)y_*'(t)-x_*'(t)y_*(t)}{\left[x_*(t)^2+y_*(t)^2+ (z_*(t)+ s)^2\right]^{3/2}}\right)ds\right)dt\\ &=\frac{1}{2\pi} \int_0^{n\omega}\frac{x_*(t)y_* '(t)- x_* '(t)y_*(t)}{x_*(t)^2+y_*(t)^2}dt\\ &= \frac{n}{2\pi}\int_0^{\omega}\Theta_1'(t)dt = -n\Phi_1= nq_1 = h_1m_1, \end{split} \] which yields the required result. \end{proof} \section{Final remarks and examples} In this section we discuss some examples and comment on some numerical experiments carried out with the software {\sl Mathematica}. \subsection{Euclidean and Clifford symmetries} Let $\gamma$ be the symmetrical configuration with modulus $q=(q_1,q_2)$. Then, $E(q_1,q_2)=R(2\pi q_2,2\pi q_1)^{n_1}$ is the Euclidean rotation of angle $2\pi m_2 n_1/n_2$ around the $z$-axis. The subgroup of order $h_1$ generated by $E(q_1,q_2)$ is the {\it Euclidean symmetry group} of the symmetrical configuration. Similarly, $C(q_1,q_2)=R(2\pi q_2,2\pi q_1)^{n_2}$ is the toroidal rotation of angle $2\pi m_1 n_2/n_1$ around the Clifford circle. The cyclic subgroup of order $h_2$ generated by $C(q_1,q_2)$ is the {\it Clifford symmetry group} of the symmetrical configuration. Denoting by $\omega$ the conformal wavelength of $\gamma$, the arc $[\gamma]_0=\gamma([0,\omega))\subset [\gamma]$ is the {\it fundamental domain} of the string and the trajectory $[\gamma]$ is obtained from $[\gamma]_{0}$ via $G_{\gamma}$, i.e., $ [\gamma]=\bigcup_{A\in G_{\gamma}} A\cdot [\gamma]_0 $. \subsection{Quantum numbers} The symmetrical configurations are labeled by three quantum numbers: the order of the symmetry group and the linking numbers with the Clifford circle and the $z$-axis. We use the notation $|n,\ell_1,\ell_2>$ for the symmetrical configuration with a symmetry group of order $n$ and linking numbers $\mathrm{lk}(\gamma,[\mathcal{C}])=\ell_1$ and $\mathrm{lk}(\gamma,[z])=\ell_2$, respectively. Let $\varrho(n)$ be the cardinality of the set of the symmetrical configurations with index of symmetry $n$. There are no symmetrical configurations with $n=1,2,3,4$. Figure reproduces the graph of the function $4\sqrt{\varrho}$, for $5\leq n \leq 120$. This suggest that $\varrho$ has an asymptotic quadratic growth. \begin{ex} In this final example, we consider the symmetrical configurations with $n=9$. This set consists of five elements: $|9,5,1>$, $|9,5,2>$, $|9,5,3>$, $|9,6,1>$, and $|9,6,2>$. The moduli of these strings are $(5/9,1/9)$, $(5/9,2/9)$, $(5/9,1/3)$, $(2/3,1/9)$, and $(2/3,2/9)$, respectively. The Euclidean and the Clifford symmetry subgroups of the first two strings are trivial. The Euclidean symmetry group of the third string is trivial, while the Clifford symmetry group has order 3 (see Figure ). The fourth and fifth strings have an Euclidean symmetry group of order 3 and have no non-trivial Clifford symmetries (see Figure ). The invariants $a$, $b$ and the conformal wavelength $\omega$ can be tabulated against the quantum numbers $n$, $\ell_1$ and $\ell_2$. For instance, the values of the natural parameters $a$, $b$ and those of the conformal wavelength $\omega$ of the stings with $n=9$ are \[ \begin{aligned} <a|9,5,1>&\approx 18.6403, \quad <b|9,5,1>\approx 0.06069,\quad <\omega|9,5,1> \approx 1.96996,\\ <a|9,5,2>&\approx 16.9699 \quad <b|9,5,2>\approx 0.09982,\quad <\omega|9,5,2> \approx 1.92188,\\ <a|9,5,3>&\approx 13.3269, \quad <b|9,5,3>\approx 0.29125 ,\quad <\omega|9,5,3> \approx 1.81376,\\ <a|9,6,1>&\approx 2.6203, \quad <b|9,6,1>\approx 0.42577,\quad <\omega|9,6,1>\approx 2.90618,\\ <a|9,6,2>&\approx 1.7209, \quad <b|9,6,2>\approx 0.90777,\quad <\omega|9,6,2>\approx 2.79219. \end{aligned} \] The shape of the strings becomes more complicated when $n$, $\ell_1$ and $\ell_2$ increase, see for instance Figure , where the symmetrical configuration with $n=63$, $\ell_1=37$ and $\ell_2=24$ is reproduced. This string has no non-trivial Euclidean symmetries and the subgroup of its Clifford symmetries has order~$3$. \end{ex} \begin{thebibliography}{99} \bibitem{BFHW} {S. Bryson, M. H. Freedman, Z.-X. He, and Z. Wang}, {\em M\"obius invariance of knot energy}, {Bull. Amer. Math. Soc. (N.S.)} {\bf 28} (1993), no.~1, 99--103. \bibitem{BC} {F. E. Burstall and D. M. J. Calderbank}, {\em Conformal submanifold geometry I--III}, {\tt arXiv:1006.5700 [math.DG]}. \bibitem{By} {P. F. Byrd and M. D. Friedman}, \textit{Handbook of elliptic integrals for engineers and physicists}, Springer-Verlag, Berlin, G\"ottingen, Heidelberg, 1954. \bibitem{CSW} {G. Cairns, R. W. Sharpe, and L. Webb}, {\em Conformal invariants for curves and surfaces in three dimensional space forms}, {Rocky Mountain J.~Math.} \textbf{24} (1994), 933--954. \bibitem{CQ} {K.-S. Chou and C.-Z. Qu}, {\em Integrable equations arising from motions of plane curves. II}, {J.~Nonlinear Sci.} \textbf{13} (2003), 487--517. \bibitem{DNF} {B. A. Dubrovin, A. T. Fomenko, and S. P. Novikov}, \textit{Modern geometry--methods and applications. Part I,} 2nd edn., GTM {\bf 93}, Springer-Verlag, New York, 1992. \bibitem{DMNna} {A. Dzhalilov, E. Musso, and L. Nicolodi}, {\em Conformal geometry of timelike curves in the (1+2)-Einstein universe}, {Nonlinear Anal.} \textbf{143} (2016), 224--255. \bibitem{EB} {M. Eastwood and G. Mar\'i-Beffa}, {\em Geometric Poisson brackets on Grassmannians and conformal spheres}, {Proc. Roy. Soc. Edinburgh Sect. A} {\bf 142} (2012), no.~3, 525--561. \bibitem{F} {A. Fialkow}, {\em The conformal theory of curves}, {Trans. Amer. Math. Soc.} {\bf 51} (1942), 435--501. \bibitem{FT} {A. T. Fomenko and V. V. Trofimov}, {\em Geometric and algebraic mechanisms of the integrability of Hamiltonian systems on homogeneous spaces and Lie algebras}, {Dynamical Systems}, Vol.~7, Springer-Verlag, 1994. \bibitem{FHW} {M. H. Freedman, Z.-X. He, and Z. Wang}, {\em M\"obius energy of knots and unknots}, {Ann. of Math. (2)} {\bf 139} (1994), no.~1, 1--50. \enlargethispage{1em} \bibitem{GM} {J.~D. Grant and E. Musso}, {\em Coisotropic variational problems}, {J.~Geom. Phys.} {\bf 50} (2004), 303--338. \bibitem{Gr} {P. A. Griffiths}, {\em Exterior differential systems and the calculus of variations}, Progress in Mathematics {\bf 25}, Birkh\"auser, Boston, 1982. \bibitem{GS2} {V. Guillemin and S. Sternberg}, \textit{Symplectic techniques in Physics}, Cambridge University Press, Cambridge, 1990. \bibitem{H} {U. Hertrich-Jeromin}, \textit{Introduction to M\"obius differential geometry}, London Mathematical Society Lecture Note Series {\bf 300}, Cambridge University Press, 2003. \bibitem{JMNbook} {G. R. Jensen, E. Musso, and L. Nicolodi}, \textit{Surfaces in classical geometries. A treatment by moving frames}, Universitext, Springer, Cham, 2016. \bibitem{J} {B. Jovanovic}, {\em Noncommutative integrability and action-angle variables in contact geometry}, {J. Symplectic Geom.} \textbf{10} (2012), no.~4, 535--561. \bibitem{LO} {R. Langevin and J. O'Hara}, {\em Conformally invariant energies of knots}, {J.~Inst. Math. Jussieu} \textbf{4} (2005), no.~2, 219--280. \bibitem{LO2010} {R. Langevin and J. O'Hara}, {\em Conformal arc-length as $\frac{1}{2}$-dimensional length of the set of osculating circles}, {Comment. Math. Helv.} \textbf{85} (2010), no.~2, 273--312. \bibitem{La} {D. F. Lawden}, \textit{Elliptic functions and applications}, Series in Applied Mathematical Science \textbf{80}, Springer-Verlag, New York, 1989. \bibitem{L} {H. Liebmann}, {\em Beitr\"age zur Inversionsgeometrie der Kurven}, M\"unchener Berichte, 1923. \bibitem{MMR} {M. Magliaro, L. Mari, and M. Rigoli}, {\em On the geometry of curves and conformal geodesics in the M\"obius space}, {Ann. Global Anal. Geom.} \textbf{40} (2011), 133--165. \bibitem{Ma} {V. O. Manturov}, \textit{Knot theory}, Chapman \& Hall/CRC Press, Boca Raton, 2004. \bibitem{Messiah} {A. Messiah}, \textit{Quantum mechanics}, Dover Publications, Inc., Mineola, NY, 2003. \bibitem{Mon} {A. Montesinos Amilibia, M. C. Romero Fuster, and E. Sanabria}, {\em Conformal curvatures of curves in $\R^{n+1}$}, {Indag. Math. (N.S.)} \textbf{12} (2001), 369--382. \bibitem{M1} {E. Musso}, {\em The conformal arclength functional}, {Math. Nachr.} \textbf{165} (1994), 107--131. \bibitem{M2} {E. Musso}, {\em Closed trajectories of the conformal arclength functional}, {Journal of Physics: Conference Series} \textbf{410} (2013), 012031. \bibitem{MNforum} {E. Musso and L. Nicolodi}, {\em Reduction for the projective arclength functional}, {Forum Math.} \textbf{17} (2005), 569--590. \bibitem{MN-CCG} {E. Musso and L. Nicolodi}, {\em Closed trajectories of a particle model on null curves in anti-de Sitter $3$-space}, {Class. Quantum Grav.} \textbf{24} (2007), 5401--5411. \bibitem{MN-SICON} {E. Musso and L. Nicolodi}, {\em Reduction for constrained variational problems on $3$-dimensional null curves}, {SIAM J. Control Optim.} \textbf{47} (2008), 1399--1414. \bibitem{MN-JMIV} {E. Musso and L. Nicolodi}, {\em Invariant signatures of closed planar curves}, {J. Math. Imaging Vision} \textbf{35} (2009), 68--85. \bibitem{MN} {E. Musso and L. Nicolodi}, {\em Hamiltonian flows on null curves}, {Nonlinearity} \textbf{23} (2010), 2117--2129. \bibitem{OR} {P. Ortega and T. Ratiu}, \textit{Moment maps and hamiltonian reductions}, Progress in Mathematics, {\bf 222}, Birkh\"auser, Boston, 2004. \bibitem{OS-Med-im} S. Ourselin and M. A. Styner (eds.), {\em Medical Imaging 2014: Image Processing}, Proceedings of SPIE, vol.~9034; {\tt DOI:10.1117/12.2052780}. \bibitem{SS} {C. Schiemangk and R. Sulanke}, {\em Submanifolds of the M\"obius space}, {Math. Nachr.} \textbf{96} (1980), 165--183. \bibitem{S} {R. Sulanke}, {\em Submanifolds of the M\"obius space II, Frenet formula and curves of constant curvatures}, {Math. Nachr.} \textbf{100} (1981), 235--257. \bibitem{T} {T. Takasu}, \textit{Differentialgeometrien in den Kugelr\"aumen. Band I. Konforme Differentialkugelgeometrie von Liouville und M\"obius}, Maruzen, Tokyo, 1938. \bibitem{V} {E. Vessiot}, {\em Contribution \`a la g\'eom\'etrie conforme. Enveloppes de sph\`eres et courbes gauches}, {J.~\'Ecole Polytechnique} \textbf{25} (1925), 43--91. \bibitem{W} {E. W. Weisstein}, {\em Elliptic Integral of the Third Kind}, from: {MathWorld} --- A Wolfram Web Resource. {http://mathworld.wolfram.com/EllipticIntegraloftheThirdKind.html}. \end{thebibliography} |
1501.04103 | Title: Usefulness of effective field theory for boosted Higgs production
Abstract: The Higgs + jet channel at the LHC is sensitive to the effects of new physics
both in the total rate and in the transverse momentum distribution at high p_T.
We examine the production process using an effective field theory (EFT)
language and discuss the possibility of determining the nature of the
underlying high scale physics from boosted Higgs production. The effects of
heavy color triplet scalars and top partner fermions with TeV scale masses are
considered as examples and Higgs-gluon couplings of dimension-5 and dimension-7
are included in the EFT. As a by-product of our study, we examine the region of
validity of the EFT. Dimension-7 contributions in realistic new physics models
give effects in the high p_T tail of the Higgs signal which are so tiny that
they are likely to be unobservable.
Body: \title{Usefulness of effective field theory for boosted Higgs production} \vspace*{1cm} \author{S.~Dawson} \affiliation{Department of Physics, Brookhaven National Laboratory, Upton, N.Y., 11973, U.S.A.} \author{I.~M.~Lewis} \affiliation{SLAC National Accelerator Laboratory, Menlo Park, CA, 94025, U.S.A.} \author{Mao Zeng} \affiliation{C.N. Yang Institute for Theoretical Physics, Stony Brook University, Stony Brook, N.Y., 11794, U.S.A.} \begin{abstract} \vspace*{0.5cm} The Higgs + jet channel at the LHC is sensitive to the effects of new physics both in the total rate and in the transverse momentum distribution at high $p_T$. We examine the production process using an effective field theory (EFT) language and discuss the possibility of determining the nature of the underlying high scale physics from boosted Higgs production. The effects of heavy color triplet scalars and top partner fermions with TeV scale masses are considered as examples and Higgs-gluon couplings of dimension-5 and dimension-7 are included in the EFT. As a by-product of our study, we examine the region of validity of the EFT. Dimension-7 contributions in realistic new physics models give effects in the high $p_T$ tail of the Higgs signal which are so tiny that they are likely to be unobservable. \end{abstract} \maketitle \section{Introduction} \makeatletter{}The recently discovered Higgs boson has all the generic characteristics of a Standard Model (SM) Higgs boson and measurements of the production and decay rates agree to the $\sim 20\ predictions . Precision measurements of Higgs couplings are essential for understanding whether there exist small deviations from the Standard Model predictions which could be indications of undiscovered high scale physics. If there are no weak scale particles beyond those of the SM, then effective field theory (EFT) techniques can be used to probe the Beyond the Standard Model (BSM) physics . The EFT is the most general description of low energy processes and new physics manifests itself as small deviations from the SM predictions. In the electroweak sector, this approach has been extensively studied . The effects of BSM operators affecting Higgs production in the strong sector have been less studied . The largest contribution to Standard Model Higgs boson production at the LHC comes from gluon fusion through a top quark loop and we examine new physics effects in this channel, along with the related Higgs + jet channel. We consider an effective Lagrangian containing the SM fermions and gauge bosons, along with a single Higgs boson, $h$. At dimension-4, the fermion- Higgs couplings can be altered from the SM couplings by a simple rescaling, \begin{equation} -\mathcal L_f={\kappa_f}\biggl({m_f\over v}\biggr) {\overline f} f h +H.c.\, , \end{equation} where $\kappa_f=1$ in the SM. In models with new physics, the gluon fusion rate can also be altered by new heavy particles interacting with the Higgs boson at one-loop, which contribute to an effective dimension-5 operator \begin{equation} \mathcal L_5=C_1 G^{A,\mu\nu}G_{\mu\nu}^A h\, , \end{equation} where $C_1=\alpha_s / (12 \pi v) $ for an infinitely heavy fermion with $\kappa_f=1$. For convenience, we define $\kappa_g$ to be the ratio of $C_1$ to this reference value, \begin{equation} \kappa_g \equiv C_1 / \left( \frac{\alpha_s}{12\pi v} \right). \end{equation} We compute the top quark contribution to scattering processes exactly using Eq. , (${\it {i.e.}}$, not in the infinite top quark mass limit), and consider $C_1$ to be only the contribution from new physics. The measurement of gluon fusion by itself can determine a combination of $\kappa_g$ and the top quark Yukawa coupling, $\kappa_t$, but cannot distinguish between the two for $m_t \gg m_h$ . Including the dimension-5 operator of Eq. \eqref{l5def}, the cross section is generically, \begin{equation} \mu_{ggh}\equiv {\sigma(gg\rightarrow h)\over \sigma(gg\rightarrow h)_{SM}}\sim \mid \kappa_t + \kappa_g \mid^2+{\cal O}\biggl({m_h^2\over m_t^2}\biggr)\, . \end{equation} The requirement that $\mid \mu_{ggh}-1\mid < 10\ is shown in Fig. , where top quark mass effects are included exactly. The SM corresponds to the point $\kappa_g=0, \kappa_t=1$. The contribution from $b-$ quarks is small and has been neglected. The boosted production of the Higgs boson through the process $pp\rightarrow h$+jet is sensitive to the Higgs- gluon effective coupling and offers the possibility of disentangling new physics effects and hence breaking the degeneracy between $\kappa_t$ and $\kappa_g$. An effective Lagrangian approach is useful for studying this high $p_T$ BSM physics and the Higgs-parton interactions can be described as a sum of higher dimension operators, \begin{equation} \mathcal L_{EFT}\sim \mathcal L_4+ \mathcal L_5+\mathcal L_6+ \mathcal L_7+ \dots\, , \end{equation} where $\mathcal L_n$ includes all dimension-$n$ operators. At dimension-5 and assuming CP conservation, there is only the single operator of Eq. \eqref{l5def} modifying the Higgs-gluon interactions. The dimension-5 operator has been broadly used to obtain higher order QCD corrections to Higgs rates . Dimension-7 operators affecting Higgs- gluon interactions from QCD interactions have received less attention . Because their contributions are proportional to the strong coupling, $g_s$, these operators can have numerically significant effects. In a previous work , we considered the effects of dimension-7 operators affecting Higgs- gluon interactions and demonstrated the importance of including these operators along with the NLO QCD corrections in order to obtain realistic predictions of boosted Higgs spectra. The largest contribution to Higgs + jet production is from the $O_1$ operators in the $gg$ initial channel. The NLO QCD corrections to this channel are relatively flat in $p_T$ and lead to an enhancement of roughly a factor of $2$ in the rate at the $14$ TeV LHC. The contributions from $O_3$ to Higgs + jet production are suppressed at lowest order QCD (LO) for large $p_T$, since they vanish in the soft Higgs limit. These contributions receive large NLO corrections, but remain numerically small and are never important. The contributions from the interference of the $O_1$ and $O_5$ operators can be important for large $p_T\sim 300~GeV$ and receive NLO QCD corrections which are again fairly $p_T$ independent and increase the rate by a factor of $\sim 1.2$. In this paper, we examine the expected size of the coefficients of the Higgs-gluon EFT dimension-5 and dimension-7 operators in several representative UV models with heavy colored scalars and fermions. We are particularly interested in the question of whether the measurement of the boosted Higgs $p_T$ distribution can distinguish the nature of the underlying UV physics, should there be any deviation from the SM. We then demonstrate how the inclusion of the dimension-7 operators affects fit to EFT Higgs parameters from gluon fusion. We work at LO QCD. In Section , we review the EFT. The heavy colored scalar and fermion models which we study are introduced in Section and the matching coefficients of the EFT presented. Phenomenological results at the LHC are given in Section IV and some conclusions about the usefulness of the EFT in this channel presented in Section . \section{Effective Lagrangian} \makeatletter{}In this section, we review the effective Lagrangian relevant for Higgs + jet production containing non-SM Higgs-gluon interactions. We consider a CP conserving Lagrangian, with no new Higgs particles, \begin{equation} \mathcal L = \mathcal L_{SM} + (\kappa_t - 1) (-1) \bar t t h + \mathcal L_5 + \mathcal L_7 + \dots, \end{equation} where \begin{equation} \mathcal L_5+ \mathcal L_7\equiv \hat C_1 O_1+\Sigma_{i=2,3,4,5} \hat C_i O_i \, , \end{equation} Note that there are no relevant dimension-$6$ operators of the type we are considering. At dimension-$5$, the unique operator is \begin{equation} O_1=G_{\mu\nu}^A G^{\mu\nu,A}h\, , \end{equation} where $G^A_{\mu\nu}$ is the gluon field strength tensor. The dimension-$7$ operators needed for the gluon fusion production of Higgs are , \begin{eqnarray} O_2&=& D_\sigma G^A_{\mu\nu}D^\sigma G^{A,\mu\nu}h \\ O_3&=&f_{ABC}G_\nu^{A,\mu} G_\sigma^{B,\nu}G_\mu^{C,\sigma} h \\ O_4&=&g_s^2h \Sigma_{i,j=1}^{n_{lf}} {\overline \psi}_i\gamma_\mu T^A \psi_i \, {\overline \psi}_j\gamma^\mu T^A \psi_j \\ O_5&=& g_s h\Sigma_{i=1}^{n_{lf}} G_{\mu\nu}^A D^\mu\, {\overline \psi}_i\gamma^\nu T^A\psi_i\, , \end{eqnarray} where our convention for the covariant derivative is $D^\sigma=\partial^\sigma -ig_s T^A G^{A,\sigma} $, $Tr(T^AT^B)={1\over 2}\delta_{AB}$ and $n_{lf}=5$ is the number of light fermions. Including light quarks, $O_4$ and $O_5$ are needed, which are related by the equations of motion (eom) to gluon-Higgs operators, \begin{eqnarray} O_4\mid_{eom}&\rightarrow & D^\sigma G^A_{\sigma\nu}D_\rho G^{A,\rho\nu} h \equiv O_4' \nonumber \\ O_5\mid_{eom} &\rightarrow & G^A_{\sigma\nu}D^\nu D^\rho G_\rho^{A,\sigma} h \equiv O_5' \, . \end{eqnarray} Since $O_4$ involves 4 light fermions, the operator only contributes to Higgs + jet production starting at NLO. A different dimension- $7$ operator is useful, \begin{equation} O_6 = -D^{\rho} D_{\rho} \left( G_{\mu\nu}^A G^{\mu\nu,A} \right) h = -\partial^{\rho} \partial _{\rho} \left( G_{\mu\nu}^A G^{\mu\nu,A} \right) h = m_h^2 O_1, \end{equation} where the last equal sign is only valid for on-shell Higgs production. Using the Jacobi identities, \begin{equation} O_6 = m_h^2 O_1 = -2 O_2 + 4 g_s O_3 + 4 O_5. \end{equation} Therefore, we can choose $O_6 = m_h^2 O_1$, $O_3$, $O_4$, and $O_5$ as a complete basis for the dimension- $7$ Higgs-gluon-light quark operators. We rewrite Eq. \eqref{ldef} as \begin{equation} \mathcal L_{\rm eff} = C_1 O_1 + \left( C_3 O_3 + C_4 O_4 + C_5 O_5 \right). \end{equation} The lowest order amplitudes for Higgs + jet production including all fermion mass dependence (bottom and top) are given in Refs. . A study of Higgs + jet production at LO QCD in the EFT approximation involves only $C_1,C_3$ and $C_5$ . At the lowest order in $\alpha_s$, $O_3$ is the only dimension-7 operator which contributes to the $gg \rightarrow gh$ channel, while $O_5$ is the only dimension-$7$ operator which contributes to channels with initial state quarks. The lowest order amplitudes in the EFT for Higgs + jet production can be found in Ref. , along with the NLO results including the effects of dimension-$7$ operators. For Higgs + jet production at NLO in BSM models, the EFT description also needs to include the higher-dimensional $3-$gluon effective vertex generated at one-loop , which could affect dijet and top quark rates . \section{UV Physics AND THE EFT} \makeatletter{}In this section, we discuss several prototype BSM physics models which have heavy particles contributing to Higgs + jet production and we compute the matching coefficients for the EFT in these models. This will allow us to estimate the size of BSM contributions. \subsection{Heavy Colored Scalars} We consider the addition of either real or complex $SU(3)$ scalars, $\phi_i$ . Our numerical results are all derived for a complex scalar triplet. The scalar portion of the Lagrangian involving a new complex scalar, $\phi_i$, and the SM-like Higgs doublet, $H$, is , \begin{eqnarray} V_{\rm complex}&=&V_{SM}(H)+m_i^2\phi_i^\dagger \phi_i +{C_h\over v}\phi_i^\dagger \phi_i (H^\dagger H) -\lambda_4(\phi_i^\dagger\phi_i)^2 \, , \end{eqnarray} where $V_{SM}$ is the SM Higgs potential. For a real scalar, \begin{eqnarray} V_{\rm real}&=& V_{SM} +{m_i^2\over 2} (\phi_i)^2 +{C_h\over 2v}(\phi_i)^2 (H^\dagger H) -\lambda_4(\phi_i)^4 \, . \end{eqnarray} In unitary gauge, $H\rightarrow (0, (h+v)/\sqrt{2})$. \subsection{Top Partner Model} Many BSM contain a charge - ${2\over 3} $ partner of the top quark. We consider a general case with a vector-like $SU(2)_L$ singlet fermion which is allowed to mix with the Standard Model like top quark \cite{ Lavoura:1992qd,AguilarSaavedra:2002kr,Aguilar-Saavedra:2013qpa, Popovic:2000dx,Dawson:2012di}. The mass eigenstates are defined to be $t$ and $T$ with masses $m_t$ and $M_T$ and are derived from the gauge eigenstates using bi-unitary transformations involving two mixing angles $\theta_L$ and $\theta_R$. Without loss of generality, $\theta_R$ can be removed by a redefinition of the top partner gauge eigenstate and the Higgs couplings are then modified from those of the SM : \begin{eqnarray} L_h^{top~partner}&=&-\biggl\{ \cos^2\theta_L{m_t\over v} {\overline t}_L t_Rh +\sin^2\theta_L{M_T\over v} {\overline T}_L T_R h\nonumber \\ && +{M_T\over 2v}\sin(2\theta_L) {\overline t}_LT_Rh+{m_t\over2 v} \sin(2\theta_L){\overline T}_L t_R h +H.c.\biggr\}\, . \end{eqnarray} Precision electroweak fits to the oblique parameters, as well as $M_W$, place stringent restrictions on the product $\sin^2\theta_L M_T^2$ and for $M_T\sim 1~\TeV$, $\sin \theta_L < .17$ . Higgs production has been investigated at NNLO for top partner models in Ref. and the rate determined to be within a few $\ channel require values of $\sin\theta_L$ that are excluded by precision measurements. ATLAS and CMS have searched for top singlet partners and excluded $M_T$ below $655$~\GeV and $687$~\GeV, respectively. Similar limits on top partner masses and mixing can be obtained for different choices of top partner $SU(2)_L$ properties . \subsection{Predictions for Coefficients} The exact results for the contributions from high scale fermion and scalar loops to the rates for $q {\overline q} \rightarrow g h$ and $gg\rightarrow gh$ are well known. Matching to the EFT expressions, the coefficient functions can be extracted. The EFT amplitude for $q{\overline q} \rightarrow gh$ from virtual heavy particles with mass, $m$, is \begin{eqnarray} \mid A(q {\overline q}\rightarrow gh)\mid^2 &=& 64 g_s^2 \biggl( {{\hat t}^2+{\hat u}^2\over {\hat s}} \biggr) \biggl[ C_1^2 +{{\hat s} C_1 C_5\over 2}\biggr]\nonumber \\ &=& \lim_{ m\rightarrow\infty} \biggl( {4\alpha_s^3\over \pi}\biggr) \biggl({{\hat u}^2+{\hat t} ^2\over {\hat s}v^2}\biggr) \mid {\cal{A}}_5({\hat s},{\hat t},{\hat u},m^2)\mid^2\, , \end{eqnarray} while the EFT amplitude for $gg\rightarrow gh$ from virtual heavy particles with mass, $m$, is \begin{eqnarray} \mid A(gg\rightarrow gh)\mid^2 &=& g_s^2\biggl[384 C_1^2 \biggl[ {m_h^8+{\hat s}^4+{\hat t}^4+{\hat u}^4\over {\hat s}{\hat t}{\hat u}}\biggr] +1152 C_1 C_3 m_h^4 \biggr]\nonumber \\ &=& \lim_{ m\rightarrow\infty} \biggl({96\alpha_s^3\over \pi}{m_h^8\over {\hat s}{\hat t}{\hat u}v^2}\biggr) \biggl\{ \mid {\cal{A}}_2({\hat s},{\hat t},{\hat u},m^2)\mid^2+ \mid {\cal{A}}_2({\hat u},{\hat s},{\hat t},m^2)\mid^2 \nonumber \\ &&+\mid {\cal{A}}_2({\hat t},{\hat u},{\hat s},m^2)\mid^2+ \mid {\cal{A}}_4({\hat s},{\hat t},{\hat u},m^2)\mid^2\biggr\}\, , \end{eqnarray} where ${\hat s},{\hat t},$ and ${\hat u}$ are the usual Mandelstam variables. The coefficient functions ${\cal{A}}_2({\hat s},{\hat t},{\hat u},m^2)$, ${\cal{A}}_4({\hat s},{\hat t},{\hat u},m^2)$ and ${\cal {A}}_5({\hat s},{\hat t},{\hat u},m^2)$ are given in Ref. for fermion loops and in Ref. for scalar loops. The $C_1, C_3$ and $C_5$ coefficients of Eqs. and depend in general on the parameters of the underlying UV completion of the model. By matching the EFT predictions with the heavy fermion expansions, we obtain the EFT coefficients given in Table . At LO, the dimension -$7$ term contributing to the $gg\rightarrow gh$ amplitude does not contain any dependence on the kinematic variables. For $\TeV$ scale masses, it is clear that the coefficients are quite small. For the top partner model, the coefficient functions for the heavier Dirac fermion contributions need to be multiplied by the factor $\sin^2 \theta_L$ appearing in Eq. \eqref{eq:TopPartner}, while the SM top quark contribution is included exactly without using the EFT. \begin{table}[tp] \centering \begin{tabular}{|c|c|c|c|} \hline & Dirac Fermion & $SU(3)$ Triplet Scalar & $SU(3)$ Octet Scalar \\ \hline $C_1(\Lambda)$ & ${\alpha_s \kappa_F\over 12\pi v}\biggl[1+{7m_h^2\over 120 m_F^2}\biggr]$ &$-{\alpha_s\over 96 \pi M_S^2 } C_h \biggl[1+{2m_h^2\over 15 M_S^2}\biggr]$ & $-{\alpha_s\over 16 \pi M_S^2 } C_h \biggl[1+{2m_h^2\over 15 M_S^2}\biggr]$ \\ $C_3(\Lambda)$ & $-{g_s \alpha_s \kappa_F\over 360 \pi v m_F^2}$ & $-{g_s\alpha_s\over 1440M_S^4}C_h$ &$-{g_s\alpha_s\over 240M_S^4}C_h$ \\ $C_5(\Lambda)$ & ${11 \kappa_F\alpha_s \over 360 \pi v m_F^2}$ & $-{\alpha_s\over 360 \pi M_S^4}C_h$ &$-{\alpha_s\over 60 \pi M_S^4}C_h$ \\ \hline \end{tabular} \vspace{-1ex} \caption{The effective Lagrangian coefficient functions for heavy Dirac fermions and heavy scalars with mass, $m_F$ and $M_S$, respectively. The coefficient functions, along with $g_s$ and $\alpha_s$, are evaluated at the scale $\Lambda=m_{F}, M_S$.} \end{table} The matching of the EFT and the underlying UV theory are done at the high scale $\Lambda$. Using the anomalous dimensions found in Ref. , the coefficients can be evolved to a low scale, $\mu_R\sim m_h$, \begin{align} \frac {d} {d \ln \mu_R} \ln \left( \frac {C_1(\mu_R)}{g_s^2(\mu_R)} \right) &= \mathcal O(\alpha_s^2(\mu_R)), \\ \frac {d} {d \ln \mu_R} \ln \left( \frac {C_3(\mu_R)}{g_s^3(\mu_R)} \right) &= \frac{\alpha_s(\mu_R)}{\pi}\, 3 C_A, \\ \frac {d} {d \ln \mu_R} \ln \left( \frac {C_5(\mu_R)}{g_s^2(\mu_R)} \right) &= \frac{\alpha_s(\mu_R)}{\pi} \left( \frac {11}{6}C_A + \frac{4}{3} C_F \right)\, , \end{align} where $C_A=3$ and $C_F={4\over 3}$. The one-loop electroweak RG running of $C_1 / g_s^2$ is non-zero, and its effect on the Higgs $p_T$ distribution in the TeV range is found to be at the percent level . The leading-logarithmic solutions to the renormalization group running equations Eq. \eqref{c1run}-\eqref{c5run} are \begin{align} C_1(\mu_R) / g_s^2 (\mu_R) &= C_1(\mu_0) / g_s^2 (\mu_0), \\ C_3(\mu_R) / g_s^3 (\mu_R) &= \left( \frac{\alpha_s(\mu_R)}{\alpha_s(\mu_0)} \right)^{-\frac {3 C_A}{2b_0}} \cdot C_3(\mu_0) / g_s^3 (\mu_0), \\ C_5(\mu_R) / g_s^2 (\mu_R) &= \left( \frac{\alpha_s(\mu_R)}{\alpha_s(\mu_0)} \right)^{-\frac {1}{2b_0} \left( \frac{11}{6} C_A + \frac 4 3 C_F \right)} \cdot C_5(\mu_0) / g_s^2 (\mu_0)\, , \end{align} where $b_0={1\over 12}(11C_A-2n_{lf})$ and $\mu_0\sim\Lambda$. The evolution of the coefficient functions is shown in Fig. . $C_1$ is increased by $\sim$ a factor of $2$ when evolving from $\Lambda\sim 5~\TeV$ to the weak scale, while $C_3$ and $C_5$ are reduced by a similar factor. \section{Phenomenology} \makeatletter{} We will eventually be interested in whether measurements of the $p_T$ spectrum can distinguish between the effects of the dimension-5 and dimension-7 operators resulting from scalars and from fermions; that is, \emph{``Is the EFT a useful tool for disentangling the source of high scale physics?"} Throughout this paper, diagrams involving the SM top quark are evaluated with exact $m_t$ dependence without using the Higgs-gluon EFT, while the contributions from heavy BSM particles, such as a color triplet scalar or a fermionic top partner, are considered both exactly and in the EFT approximation. \subsection{Heavy Colored Scalars} We begin by considering the effect of heavy color triplet scalars on Higgs + jet production. (The case of a light colored scalar has been considered in .) We use CJ12 NLO PDFs and $\mu_R=\mu_F=\sqrt{m_h^2+p_T^2}$ for all curves, with $m_h=125~\GeV$, $m_t=173~\GeV$, and $m_b=4.5~\GeV$. All plots refer to Higgs + jet production at lowest order and with $\sqrt{s}=14~$\TeV. When using the EFT, the effects of heavy scalars are included using the coefficients of Table I. Since the effects are suppressed by $1/M_S^2$ in $C_1$ and $1/M_S^4$ in the other $C_i$, we expect relatively small effects unless the coefficient function $C_h$ is large. We expect $C_h$ to be of order the electroweak scale in a realistic model and in our plots, we take $C_h=3M_Z$.\footnote{If $\phi_i$ corresponds to the left-handed top squark of the MSSM, then in the alignment limit ($\sin\beta=\cos\alpha$), $C_h\sim 3 M_Z$, which motivates our choice. This numerical value is not important for our conclusions, as long as $C_h/M_Z$ is not a large number.} Numerically, the effects are linear in $C_h$ for modest values of $C_h / M_Z$ and our results can be trivially rescaled. The exact one-loop contribution of the heavy scalars relative to the SM rate are shown in Fig. and as expected, they cause only a few percent deviation from the SM rate at low $p_T$. We define the ratio, ``BSM/SM'' to be the differential (or integrated) rate in the theory with the SM top quark and scalar included exactly normalized to the SM rate minus $1$, i.e. it is the incremental contribution from the addition of a scalar. At large $p_T$, the deviation becomes significant, approaching $\sim 15\ scalar. We note that the effects of a color octet scalar are a factor of $C_A / T_F = 6$ larger than those of a color triplet scalar. The integrated cross sections with a $p_{T_{cut}}$ are shown in Fig. , and a significant contribution from the scalars to the boosted Higgs signal is apparent for $p_{T_{cut}} \sim M_S$ for $M_S=500$~\GeV. For the heavier scalar, $M_S\sim $~\TeV, the effects are only a few $\ Since the lowest order contribution from scalars is known exactly, we can explore the range of validity of the EFT. Fig. shows the deviation of the EFT calculation from the exact 1-loop result when color triplet scalars are included. For a $500~\GeV$ scalar, the EFT is accurate to within a few $\ has large deviations above $500~\GeV$ when only the dimension-$5$ ($\sim 1/ M_S^2$) contributions are included. Including the dimension-7 contributions improves the accuracy of the EFT. Interestingly, for $M_S=1~\TeV$, the EFT becomes less accurate at large $p_T$ when the dimension-$7$ effects are included. The EFT expansion clearly breaks down at a scale $p_T\sim M_S$. Fig. demonstrates the accuracy of the EFT in the $p_T$ integrated cross section and we observe the same behavior. (The cross section is integrated to $p_T=1~\TeV$, where the EFT is breaking down. Since the partonic results are integrated with a falling PDF spectrum, we expect the results to be reasonably reliable.) The contributions from the $gg$ and $qg$ initial states are shown separately in Figs. and . \subsection{Heavy Fermion Top Partners} In this section we consider the effect of a top partner model on the shape of the Higgs $p_T$ distribution. We take the top partner mass $M_T=500~\GeV$ and the mixing angle $\cos\theta_L= 0.966$. Fig. shows the ratio of the inclusive cross section in the top partner model to that in the SM, minus 1, evaluated with the exact dependence on the masses $m_t$ and $M_T$, along with the same quantity integrated with a $P_{T_{cut}}$. We note that the results of Ref. demonstrate large effects at high $p_T\sim1$~\TeV\ when $\sin\theta_L=0.4$. Regretably, such large mixing angles are excluded by precision electroweak data. (We agree with the results of Ref. for small $\sin\theta_L$.) Fig. shows the accuracy of the EFT predictions for differential and integrated $p_T$ distributions, relative to the results with exact $m_t$ and $M_T$ dependence. We close this section by summarizing our results for top partners and scalars in Fig. , which dramatically demonstrates the difficulty of extracting information about the underlying UV physics. \subsection{EFT Fits} In this subsection, we consider the effects of a general rescaling of the EFT coefficients. As in Eq. \eqref{eq:kappaF} and Eq. \eqref{eq:defKg}, we consider the SM top quark contribution rescaled by $\kappa_t$, and the $C_1$ coefficients rescaled by $\kappa_g$ relative to an infinitely heavy Dirac fermion whose mass comes entirely from the Higgs, i.e. $C_1 = \kappa_g \cdot \alpha_s / (12\pi v)$. For the dimension-7 operators, we vary the matching coefficients $C_i = \kappa_i C_i(M_S=500~{\rm GeV}, C_h =3m_Z)$ for $i=3,5$, where the reference values, scaled by $\kappa_i$, are $C_3(M_S, C_h)=-g_s \alpha_s C_h / (1440 M_S^4)$ and $C_5(M_S, C_h) = -\alpha_s C_h / (360 \pi M_S^4)$ corresponding to the EFT coefficients from Table I for a $500$~\GeV \ scalar. The total cross section for single Higgs production is roughly unchanged from the SM, if we fix $\kappa_t+\kappa_g$ to be 1, according to Eq. \eqref{eq:SMrel}. Fig. demonstrates that excessively large values of $\kappa_5$ are required for a large effect from $O_5$. Fig. shows that the inclusion of $O_3$ has very little effect even for huge values of $\kappa_3$, as expected from the helicity arguments in . On the other hand, the effect of rescaling $\kappa_t$ and $\kappa_g$ separately can have a relatively large effect. \section{Conclusion} \makeatletter{}The process Higgs + jet has been proposed as a useful channel for studying BSM physics and for disentangling the effects of a modification of the dimension-$4$ $t {\overline t}h$ Yukawa coupling from a non-SM dimension-$5$ Higgs-gluon effective vertex. We further include dimension-$7$ effective Higgs-gluon operators and compute the EFT coefficient functions in two representative models with heavy colored scalars and fermions. The coefficient functions are suppressed by inverse powers of the heavy mass scales, $m$, and are therefore quite small. At lowest order, the effects of colored scalars and fermions can be computed exactly and the accuracy of the EFT determined. Typically, better accuracy is obtained in the $gg$ channel than in the $qg$ channel, and the EFT is accurate to a few percent for $p_T < m$. Our results illustrate the dilemma of the EFT approach: large effects are only obtained at high $p_T$ and the contribution from the dimension -$7$ operators is small for $p_T < m$. On the other hand, Fig. demonstrates a modest sensitivity to $C_1$, independent of $\kappa_t$. If any deviation is found in the Higgs transverse momentum distribution up to $1~\TeV$, the deviation is unlikely to provide information about the UV physics beyond the single parameter $C_1$. Our results support the validity of an approach using only the dimension-$5$ Higgs-gluon operator. Inclusion of the NLO QCD corrections is unlikely to change this conclusion, since the NLO corrections to the $C_1^2$ contribution do not have a large $p_T$ dependence in the region where the EFT is valid. \begin{acknowledgments} S.D. thanks A. Ismail and I. Low for discussions about the effects of virtual scalar particles. The work of SD and IL is supported the U.S. Department of Energy under grants No.~DE-SC0012704 and DE-AC02-76SF00515. The work of MZ is supported by NSF grant PHY-1316617. \end{acknowledgments} \clearpage \bibliography{hjet} |
1501.04111 | Title: A Precise Determination of $\alpha_s$ from the C-parameter Distribution
Abstract: We present a global fit for $\alpha_s(m_Z)$, analyzing the available
C-parameter data measured at center-of-mass energies between $Q=35$ and $207$
GeV. The experimental data is compared to a N$^3$LL$^\prime$ +
$\mathcal{O}(\alpha_s^3)$ + $\Omega_1$ theoretical prediction (up to the
missing 4-loop cusp anomalous dimension), which includes power corrections
coming from a field theoretical nonperturbative soft function. The dominant
hadronic parameter is its first moment $\Omega_1$, which is defined in a scheme
which eliminates the $\mathcal{O}(\Lambda_{\rm QCD})$ renormalon ambiguity. The
resummation region plays a dominant role in the C-parameter spectrum, and in
this region a fit for $\alpha_s(m_Z)$ and $\Omega_1$ is sufficient. We find
$\alpha_s(m_Z)=0.1123\pm 0.0015$ and $\Omega_1=0.421\pm 0.063\,{\rm GeV}$ with
$\chi^2/\rm{dof}=0.988$ for $404$ bins of data. These results agree with the
prediction of universality for $\Omega_1$ between thrust and C-parameter within
1-$\sigma$.
Body: \preprint{\vbox{\hbox{UWTHPH 2015-1}\hbox{MIT--CTP 4630}\hbox{LPN14-128} }} \title{A Precise Determination of $\alpha_s$ from the C-parameter Distribution} \author{Andr\'e H.~Hoang} \affiliation{University of Vienna, Faculty of Physics, Boltzmanngasse 5, A-1090 Wien, Austria} \affiliation{Erwin Schr\"odinger International Institute for Mathematical Physics, University of Vienna, Boltzmanngasse 9, A-1090 Vienna, Austria} \author{Daniel W.~Kolodrubetz} \affiliation{Center for Theoretical Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA} \author{Vicent Mateu} \affiliation{University of Vienna, Faculty of Physics, Boltzmanngasse 5, A-1090 Wien, Austria} \author{Iain W.~Stewart} \affiliation{Center for Theoretical Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA} \begin{abstract} We present a global fit for $\alpha_s(m_Z)$, analyzing the available C-parameter data measured at center-of-mass energies between $Q=35$ and $207$\,GeV. The experimental data is compared to a \mbox{N$^3$LL$^\prime$ + $\mathcal{O}(\alpha_s^3)$ + $\Omega_1$} theoretical prediction (up to the missing four-loop cusp anomalous dimension), which includes power corrections coming from a field theoretical nonperturbative soft function. The dominant hadronic parameter is its first moment $\Omega_1$, which is defined in a scheme which eliminates the $\ord{\Lambda_{\rm QCD}}$ renormalon ambiguity. The resummation region plays a dominant role in the C-parameter spectrum, and in this region a fit for $\alpha_s(m_Z)$ and $\Omega_1$ is sufficient. We find $\alpha_s(m_Z)=0.1123\pm 0.0015$ and $\Omega_1=0.421\pm 0.063\,{\rm GeV}$ with $\chi^2/\rm{dof}=0.988$ for $404$ bins of data. These results agree with the prediction of universality for $\Omega_1$ between thrust and C-parameter within 1-$\sigma$. \end{abstract} \maketitle \section{Introduction} In order to study Quantum Chromodynamics (QCD) accurately in the high-energy regime, it is useful to exploit the wealth of data from previous $e^+\, e^-$ colliders such as LEP. Here the final states coming from the underlying partons created in the collisions appear as boosted and collimated groups of hadrons known as jets. Event shapes have proven to be very successful to study these collisions quantitatively. They combine the energy and momenta of all of the measured hadrons into an infrared- and collinear-safe parameter which describes the geometric properties of the whole event by a single variable distribution. Due to their global nature event shapes have nice theoretical properties, making it possible to obtain very accurate theoretical predictions using QCD. Most $e^+e^-$ event shape variables quantify how well the event resembles the situation of two narrow back-to-back jets, called dijets, by vanishing in this limit. Because the dijet limit involves restrictions that only allow collinear and soft degrees of freedom for the final-state radiation, such QCD predictions involve a number of theoretical aspects that go beyond the calculation of higher-order perturbative loop corrections. These include factorization, to systematically account for perturbative and nonperturbative contributions, and the resummation of large logarithmic corrections by renormalization group evolution. Comparisons of predictions for event shapes with experimental data thus provide non-trivial tests of the dynamics of QCD. Due to the high sensitivity of event shapes to jets induced by gluon radiation they are an excellent tool to measure the strong coupling $\alpha_s$. For more inclusive hadronic cross sections (like $e^+e^-\to $ hadrons) the $\alpha_s$ dependence is subleading because it only occurs in corrections to a leading order term, while for event shapes the $\alpha_s$ dependence is a leading-order effect. For this reason, the study of event shapes for determining $\alpha_s$ has a long history in the literature (see the review~ and the workshop proceedings~), including recent analyses which include higher-order resummation and corrections up to ${\cal O}(\alpha_s^3)$~. Several previous high-precision studies which determine $\alpha_s(m_Z)$~\cite{Davison:2008vx,Becher:2008cf,Abbate:2010xh, Abbate:2012jh,Gehrmann:2012sc} focus on the event shape called thrust~, \begin{equation} \tau \,=\, 1-T \,=\, \min_{\vec{n}}\! \left( 1 -\, \frac{ \sum_i|\vec{n} \cdot \vec{p}_i|}{\sum_j |\vec{p}_j|} \right), \end{equation} where $\vec{n}$ is called the thrust axis and it follows from the above equation that $0 \le \tau \le 1/2$. Another event shape, known as C-parameter~, can be written as: \begin{equation} C=\frac{3}{2} \, \frac{\sum_{i,j} | \vec{p}_i | | \vec{p}_j | \sin^2 \theta_{ij}} {\left( \sum_i | \vec{p}_i | \right)^2}\,, \end{equation} where $\theta_{ij}$ gives the angle between particles $i$ and $j$. It is straightforward to show that $0\le C\le 1$. In a previous paper~ we computed the C-parameter distribution with a resummation of large logarithms at N$^3$LL$^\prime$ accuracy, including fixed-order terms up to ${\cal O}(\alpha_s^3)$ and hadronization effects using a field-theoretic nonperturbative soft function. These results were achieved by using the Soft Collinear Effective Theory (SCET)~. Our results for $C$ are valid in all three of the peak, tail, and far-tail regions of the distribution, and are the most accurate predictions available in the literature, having a perturbative uncertainty of $\simeq 3\ Since both $\tau$ and $C$ vanish in the dijet limit, it is worthwhile to contrast them in order to anticipate differences that will appear in the analysis. Differences between $C$ and $\tau$ include the following: \begin{enumerate} \item [a)] calculating $\tau$ requires identifying the thrust axis with a minimization procedure, while $C$ does not involve a minimization; \item [b)] $\tau$ has a single sum over particles while $C$ has a double sum; \item [c)] the size of the nonperturbative region, where the entire shape function is important, is larger for $C$ compared to $\tau$ due to an enhancement by a factor of $3\pi/2$; \item [d)] the resummation region for $C$ is larger than that for thrust since the logarithms appear as $\ln(C/6)$ compared to $\ln(\tau)$, which increases the range of $C$ values that are useful for $\alpha_s$ fits; \item [e)] fixed-order predictions for the thrust cross section are smooth across the threshold where non-planar events first contribute, $\tau=1/3$, while the fixed-order C-parameter cross section has an integrable singularity at this threshold, $C_\text{shoulder}=3/4$. The singularity for $C$ comes from the fact that the leading-order distribution is not continuous at $C=C_\text{shoulder}$. (The $C$-distribution can be made smooth here using LL resummation~.) \end{enumerate} A key similarity is that in the dijet limit ($C,\,\tau \ll 1$) the partonic cross sections for thrust and C-parameter are related up to NLL by using \mbox{$\tau_{\text{NLL}}=C_{\text{NLL}}/6$}~. This relation quantifies several qualitative relations between $C$ and $\tau$ in the dijet region. Recent higher-order event-shape analyses~ have found values of $\alpha_s(m_Z)$ significantly lower than the world average of $\alpha_s(m_Z)=0.1185\pm 0.0006$~ which is dominated by the lattice QCD determination~. This includes the determination carried out for thrust at N$^3$LL$^\prime\,$+$\,{\cal O}(\alpha_s^3)$ in Ref.\, cusp anomalous, but conservative estimates show that this has a negligible impact on the perturbative uncertainties. Results at N$^3$LL$^\prime$ also technically require the unknown 3-loop non-logarithmic constants for the jet and soft functions which are also varied when determining our uncertainties, but these parameters turn out to only impact the peak region which is outside the range of our $\alpha_s(m_Z)$ fits in the resummation region.}, which is also consistent with analyses at N$^2$LL\,+\,${\cal O}(\alpha_s^3)$~ which consider the resummation of logs at one lower order. In Ref.~ a framework for a numerical code with N$^2$LL precision for many $e^+e^-$ event shapes was found, which could also be utilized for $\alpha_s$ fits in the future. In Ref.~ it was pointed out that including a proper fit to power corrections for thrust causes a significant negative shift to the value obtained for $\alpha_s(m_Z)$, and this was also confirmed by subsequent analyses~. Recent results for $\alpha_s(m_Z)$ from $\tau$ decays~, DIS data~, the static potential for quarks~, as well as global PDF fits~ also find values below the world average. With the new analysis we present here, we provide another event-shape determination of $\alpha_s(m_Z)$ with a high level of precision. We will also simultaneously examine the leading power correction to the distribution, which should be universal between thrust and \mbox{C-parameter}. This paper is organized as follows. In Sec.~ we review the theoretical calculation of the C-parameter cross section, presented in more detail in Ref.~. The details on the experimental data and fit procedure used in our analysis are given in Secs.~ and~. In Sec.~ we present the results of our fit for $\alpha_s(m_Z)$ and the first moment of the nonperturbative soft function, $\Omega_1^C$. Fits which include hadron-mass effects are discussed in Sec.~. In Sec.~ we make predictions for the peak and far-tail regions of the distribution, which are not used in our fit, and compare those regions to experimental data. The universality of $\Omega_1$ is discussed in Sec~, where we compare our results with the previous fit done using thrust in Ref.~. Finally, Sec.~ contains our conclusions. We also include three appendices. The first, Appendix~, contains the formulae needed to calculate the profile functions, the second Appendix~ contains results that support our choice for the definition of the renormalon free $\Omega_1$ parameter to use for $C$; and the third compares fit results for thrust with our profiles and those from Ref.~. \section{Theory Review} Until a few years ago, the theoretical uncertainties related to perturbative contributions as well as hadronic power corrections were still larger than the experimental uncertainties. The situation on the theory side has dramatically changed with the calculation of ${\cal O}(\alpha_s^3)$ corrections~, and the pioneering use of SCET to obtain higher-order perturbative corrections in Ref.~, and to obtain accurate predictions for the full spectrum and a precision description of power corrections in Ref.~. In this section we review the theoretical work behind our calculation of $\df \sigma/\df C$, presented in Ref.\,. SCET separates the physics occurring at the different energy and momentum scales relevant to the underlying jets whose properties are characterized by an event shape. For the C-parameter distribution the relevant scales are: (i) the hard scale $\mu_H$, which is related to the production of partons at short distances and is of the order of the center-of-mass energy $Q$; (ii) the jet scale $\mu_J\sim Q\sqrt{C/6}$, which governs the formation and evolution of the jets; and (iii) the soft scale $\mu_S\sim QC/6$, which is the scale of large-angle soft radiation. All three scales are widely separated in the dijet region $C\ll 1$, where most of the events occur and where the distribution is maximal. There is one function associated to each one of these scales in the factorization theorem that describes the dominant contribution of the C-parameter distribution in the dijet limit: (i) the hard function $H$ (the modulus squared of the QCD-to-SCET matching coefficient), which is common to all dijet event-shape factorization theorems; (ii) the jet function $J_\tau$ (given by matrix elements of quark fields with collinear Wilson lines), which is common for C-parameter~, thrust~ and Heavy-Jet-Mass ($\rho$)~; and (iii) the soft function $S_{\widetilde C}$ (defined by a vacuum matrix element of purely soft Wilson lines), which in general depends on the specific form of the event shape. Whereas the former two are perturbative ($\mu_H, \mu_J \gg \Lambda_{\rm QCD}$), permitting the calculation of the hard and jet functions as an expansion in powers of $\alpha_s$, the soft function has perturbative corrections ($\mu_S\gg \Lambda_{\rm QCD}$) as well as nonperturbative contributions that need to be accounted for ($\mu_S\gtrsim \Lambda_{\rm QCD}$). Renormalization evolution between the three scales $\mu_H$, $\mu_J$, and $\mu_S$ sums up large logarithms to all orders in perturbation theory. It turns out that the soft function anomalous dimension for $C$ and $\tau$ are identical~, providing a connection between these two event shapes at every order of perturbation theory. Only corrections related to non-logarithmic terms in their soft functions, and the associated towers of logarithms, differ between $C$ and $\tau$. The soft function can be further factorized into a partonic soft function $\hat S_{\!\widetilde C}$, calculable in perturbation theory, and a nonperturbative shape function $F_{\!C}$, which has to be obtained from fits to data. In the strict $\overline {\rm MS}$ scheme this factorization was achieved in Refs.~. (Analytic power corrections for the C-parameter distributions have also been studied in other schemes and frameworks, see e.g.\ Refs.~.) The treatment of hadronic power corrections greatly simplifies in the tail of the distribution, defined by \mbox{$QC\gg 3\pi \Lambda_{\rm QCD}$}, where the shape function can be expanded in an OPE. Here the leading power correction is parametrized by $\Omega_1^C$, the first moment of the shape function. The main effect of this leading power correction is a shift of the cross section, \mbox{${\df\hat\sigma} (C) \to {\df\hat\sigma}(C \,-\, \Omega_1^C/Q)$}. Interestingly, this matrix element is related to the corresponding one in thrust by \begin{align} \frac{\Omega_1^\tau}{2} = \frac{\Omega_1^C}{3\pi}\equiv \Omega_1 \,. \end{align} This relation was first derived using dispersive models with a single soft-gluon approximation in Ref.~. The equality can actually be derived to all orders in QCD just using quantum field theory~, but ignoring hadron-mass effects~. These hadron-mass effects can also be formulated purely with quantum field theory operators~. Although they may in general give large corrections, hadron-mass effects turn out to violate \Eq{eq:O1univ} at only the $2\ \subsection{Factorized Cross Section Formula} In order to understand the perturbative components of the C-parameter cross section we make use of the \mbox{C-parameter} factorization formula. To make the connection to thrust simpler we will often use functions defined with the variable $\widetilde C = C/6$. For the perturbative cross section we find~: \begin{align} \!\!\!\frac{1}{\sigma_0}\frac{\df \hat\sigma_{\rm s}}{\df C}=\frac{Q}{6} H(Q,\mu)\! \int\! \df s\, J_\tau(s,\mu) S_{\widetilde C}\bigg(\frac{Q C}{6}- \frac{s}{Q},\mu\bigg). \end{align} Here $J_\tau$ is the thrust jet function, obtained by the convolution of the two hemisphere jet functions, $J_\tau = J \otimes J$. $J_\tau$ describes the collinear radiation in the direction of the two jets. Its definition and expression up to $\mathcal{O}(\alpha_s^3)$\, can be found in Refs.~. The three-loop non-logarithmic coefficient of this jet function, $j_3$, is not known, and we vary it in our scans. The anomalous dimension of $J_\tau$ is known at three loops, and can be obtained from Ref.~. The hard factor $H$ contains short-distance QCD effects and is obtained from the Wilson coefficient of the QCD-SCET matching of the vector and axial-vector currents. The hard function is the same for all event shapes for massless quarks, and its expression up to $\mathcal{O}(\alpha_s^3)$ , can be found in Ref.~. The full anomalous dimension of $H$ is known at three loops, ${\cal O}(\alpha_s^3)$~. The soft function $S_{\widetilde C}$ describes wide-angle soft radiation between the two jets. It is defined as \begin{align} S_{\widetilde C}(\ell,\mu) = \frac{1}{N_C}\langle \, 0\,| \tr\, \overline{Y}_{\bar n}^T Y_n \delta\bigg(\!\ell- \frac{Q\widehat C}{6}\bigg) Y_n^\dagger \overline{Y}^*_{\bar n}\, |\, 0\,\rangle\,, \end{align} where $\widehat C$ is an operator whose eigenvalues on physical states correspond to the value of \mbox{C-parameter} for that state: $\widehat C |X\rangle = C(X) |X\rangle$. Since the hard and jet functions are the same as for thrust, the anomalous dimension of the soft function has to coincide with the anomalous dimension of the thrust soft function. Hence one only needs to determine the non-logarithmic terms of the \mbox{C-parameter} soft function. In Ref.~ we computed it analytically at one loop, $s_1^{\widetilde C} = -\,\pi^2 C_F/6$, and used EVENT2 to numerically determine the two-loop non-logarithmic coefficient $s_2^{\widetilde C}$, with the result \begin{align} s_2^{\widetilde C}=-\,43.2\,\pm\,1.0 \,. \end{align} The three-loop non-logarithmic coefficient of the \mbox{C-parameter} soft function, $s_3^{\widetilde C}$, is currently not known, and we estimate it with a Pad\'e approximation, assigning a very conservative uncertainty. We vary this constant in our scan analysis. The precise definitions of $j_3$ and $s_2^{\widetilde C}$ as well as $s_3^{\widetilde C}$ can be found in Eqs.~(A10) and (A12) of Ref.~. In Eq.~() the hard, jet and soft functions are evaluated at a common scale $\mu$. If fixed-order expressions are used for these functions, then there is no scale choice that simultaneously minimizes the logarithms for these three functions. One can instead renormalization-group evolve from $\mu$ to the respective scales $\mu_H\sim Q$, $\mu_J\sim Q \sqrt{C/6}$ and $\mu_S\sim Q C/6$ at which the logs in each of $H$, $J_\tau$, and $S_{\widetilde C}$ are minimized, and only use fixed-order expressions for these functions at these scales. In this way, large logs of ratios of the scales are summed up in the renormalization group evolution kernels $U_H$, $U_J^\tau$, and $U_S^\tau$: \begin{align} \frac{1}{\sigma_0}\frac{\df \hat\sigma_{\rm s}}{\df C} & \,=\, \frac{Q}{6} H(Q,\mu_H)\,U_H(Q,\mu_H,\mu)\,\times \\\nonumber &\int\! \df s\, \df s^\prime\, J_\tau(s,\mu_J)\,U_J^\tau(s-s^\prime,\mu,\mu_J) \\\nonumber &\int\! \df k\,U_S^\tau(k,\mu,\mu_S)\, S_{\widetilde C}\bigg(\!\frac{Q C}{6}- \frac{s}{Q}-k,\mu_S\!\bigg). \end{align} The evolution kernels $U_H$, $U_J^\tau$ and $U_S^\tau$ resum the large logarithms, $\ln(C/6)$, and explicit expressions can be found in Ref.~. The only unknown piece in our resummation of logarithms at N$^3$LL order is the small contribution from the four-loop cusp anomalous dimension, $\Gamma_{3}^\text{cusp}$, which we estimate using a Pad\'e approximation and conservatively vary in our analysis. While \Eq{eq:singular-resummation} gives the part of the cross section that is singular and non-integrable as $C \rightarrow 0$, we also need to include the integrable or nonsingular contribution. This can be written as \begin{align} \frac{1}{\sigma_0} \frac{\df \hat{\sigma}_{\rm ns}}{\df C} & \,=\, \frac{\alpha_s(\mu_{\rm ns})}{2\pi}\, f_1(C)\, \\ &+\left(\frac{\alpha_s(\mu_{\rm ns})}{2\pi}\right)^{\!\!2}\! \bigg[f_2(C) + \beta_0 \ln\!\Big(\frac{\mu_{\rm ns}}{Q}\Big) f_1(C)\bigg]\nn\\ &+\,\left(\frac{\alpha_s(\mu_{\rm ns})}{2\pi}\right)^{\!\!3}\! \bigg\{f_3(C) + 2\, \beta_0 \ln\!\Big(\frac{\mu_{\rm ns}}{Q}\Big) f_2(C) \nn \\ & \quad + \bigg[\frac{\beta_1}{2} \ln\!\Big(\frac{\mu_{\rm ns}}{Q}\Big) + \beta_0^2 \ln^2\!\Big(\frac{\mu_{\rm ns}}{Q}\Big)\bigg] f_1(C) \bigg\} \nn\\ &+\, {\cal O}(\alpha_s^4) \,.\nonumber \end{align} The functions $f_1(C)$, $f_2(C)$, and $f_3(C)$ were determined in Ref.~ using the fixed-order results at ${\cal O}(\alpha_s^{1,2,3})$~. The nonsingular cross section $\df\hat{\sigma}_{\rm ns}/\df C$ is independent of the renormalization scale $\mu$ order by order, and therefore we evaluate these pieces at the nonsingular scale $\mu_{\rm ns}$, and vary this scale to estimate higher-order perturbative nonsingular corrections. The scale variation of $\mu_{\rm ns}$ will be discussed further in Section . The full partonic cross section is then given by \begin{align} \frac{1}{\sigma_0} \frac{\df \hat{\sigma}}{\df C} &= \frac{1}{\sigma_0} \frac{\df \hat{\sigma}_{\rm s}}{\df C} + \frac{1}{\sigma_0} \frac{\df \hat{\sigma}_{\rm ns}}{\df C}\,. \end{align} Nonperturbative effects are included by convolving Eq.~() with a shape function: \begin{align} \frac{1}{\sigma_0}\frac{\df \sigma}{\df C} = \!\int \!\df p \,\frac{1}{\sigma_0}\frac{\df \hat\sigma}{\df C}\Big(C-\frac{p}{Q}\Big)F_C(p)\,. \end{align} One important property of this shape function is that its first moment encodes the leading power correction to our cross section. In the $\overline{\rm MS}$ scheme this moment is given by \begin{align} \overline \Omega_1^{\,C} &\equiv \!\int \df k\, k\, F_C(k)\,. \end{align} Up to the normalization factors shown in Eq.~() we expect approximate universality between the $\Omega_1$ for \mbox{C-parameter} and thrust. For the calculation of the cross section, the shape function is expanded in a complete basis of functions obtained by an appropriate infinite-range mapping of the Legendre polynomials~, with the coefficients chosen to maintain the first moment. For further details on the implementation of the shape function for C-parameter see Ref.~. We remove an $\ord{\Lambda_\text{QCD}}$ renormalon by using the Rgap scheme , which introduces a subtraction scale $R$ into our formula, as well as the gap parameter $\bar{\Delta}$ and the perturbative scheme-change gap parameter $\delta(R,\mu_S)$. Here, $\delta(R,\mu_S)$ is given by a perturbative series in $\alpha_s(\mu_S)$ whose mass dimension is set by an overall factor of $R$, and which also contains $\ln(\mu_S/R)$ factors. The convolution with the shape function now becomes, \begin{align} \frac{\df \sigma}{\df C} = \!\int \!\df p \,e^{-3 \pi \frac{\delta(R,\mu_{s})}{Q}\frac{\partial}{\partial C}}\frac{\df \hat\sigma}{\df C}\Big(C-\frac{p}{Q}\Big) &\\ \times F_C\big(p - 3 \pi \bar{\Delta} (R,\mu_S)\big)&\,.\nn \end{align} The final component of our cross section is properly accounting for hadron-mass effects following Ref.~. Hadron-mass effects induce an additional series of large perturbative logarithms which start at NLL, $\alpha_s^k \ln^k(\mu_S/\Lambda_{\rm QCD})$, and also break the exact universality between $\Omega_1^C$ and $\Omega_1^\tau$ given in Eq.~(). These effects are accounted for by including dependence on the transverse velocity, $r \equiv \frac{p_\perp}{\sqrt{p_\perp^2 + m_H^2}}$, in the nonperturbative matrix elements (here, $m_H$ gives the non-zero hadron mass). In particular, in the Rgap scheme the first moment of the shape function is actually given by \begin{align} \int\! \df k \, k\, F_C\big(k-3 \pi \bar{\Delta}(R,\mu_S)\big)\, = \,\Omega_1^C(R,\mu_S) &\\ =3 \pi \!\int_0^1 \! \df r\, g_C(r) \, \Omega_1(R,\mu_S,r)&\,.\nn \end{align} In the $\overline{\rm MS}$ scheme the definition accounting for hadron-mass effects is the same as \Eq{eq:firstmomentwithhadron}, but one sets \mbox{$\bar\Delta=0$} and removes $R$ as an argument for these parameters. Accounting for both the the $\overline{\rm MS}$ running due to hadron masses and the R-evolution running in the Rgap scheme, the evolution of the integrand on the right-hand side of \Eq{eq:firstmomentwithhadron} is given by, \begin{align} & g_C(r)\,\Omega_1(R,\mu,r) = g_C(r)\!\left[ \frac{\alpha_s(\mu)}{\alpha_s(\mu_\Delta)} \right]^{\hat{\gamma}_1(r)}\! \Omega_1(R_\Delta, \mu_\Delta,r) \nn\\ &+\Delta^{\rm diff}(R_\Delta,R,\mu_\Delta,\mu,r)\,, \end{align} where $R_\Delta$ and $\mu_\Delta$ give the initial scales where the function $\Omega_1(R_\Delta,\mu_\Delta,r)$ is defined. The perturbative evolution kernel $\Delta^{\rm diff}$ gives the $R$ and $\mu$ running for each value of $r$. The function $g_C$ encodes the event-shape dependence of the hadron-mass effects and $\hat{\gamma}_1$ gives the solution to the one-loop RGE for $\Omega_1$ with hadron masses derived in Ref.~. Since the two- and three-loop $r$-dependent anomalous dimensions are unknown, we do not treat the logs generated by hadron-mass effects to the same level of precision. When hadron-mass effects are accounted for we always sum the associated logarithms at NLL. An analogous formula to \Eq{eq:hadron-masses} also holds for the thrust parameter $\Omega_1^\tau$. Combining all of these elements gives the complete cross section. Note that we can resum to any order up to N${}^3$LL$^\prime$ and can choose to include or leave out the shape function, renormalon subtraction and hadron-mass effects. This flexibility allows us to see how the analysis changes when we take into account each of these additional physical considerations and enables us to test how robust the fits are to various changes in the theoretical treatment. \subsection{Profile Functions} In order to smoothly transition between the nonperturbative, resummation, and fixed-order regions we make use of profile functions $\mu_i(C)$ for the renormalization scales $\mu_H$, $\mu_J(C)$, $\mu_S(C)$, $R(C)$, and $\mu_{\rm ns}(C)$. In the three $C$ regions, the requirements on the scales which properly deal with large logarithms, nonperturbative effects, and the cancellations between singular and nonsingular contributions in the fixed-order region are \begin{align} & \text{1) nonperturbative:~} C \lesssim 3\pi\,\Lambda_\text{QCD}\nn \\ & \qquad \mu_H \sim Q,\ \mu_J \sim \sqrt{\Lambda_\text{QCD} Q},\ \mu_S \!\sim\! R \!\sim\! \Lambda_\text{QCD} \,, \nn \\[5pt] & \text{2) resummation:~} 3\pi\,\Lambda_\text{QCD} \ll C < 0.75 \\ & \qquad \mu_H \sim Q,\ \mu_J \sim Q \sqrt{\frac{C}{6}},\ \mu_S \! \sim \! R \!\sim\! \frac{QC}{6} \gg \Lambda_{\rm QCD} \,, \nn \\ & \text{3) fixed-order:~} C > 0.75 \nn \\ & \qquad \mu_{\rm ns} = \mu_H = \mu_J = \mu_S = R \sim Q\gg \Lambda_{\rm QCD}\nn \,.\end{align} In addition we take the fixed-order nonsingular scale $\mu_{\rm ns}\sim \mu_H$ in the nonperturbative and resummation regions. Our profile functions $\mu_i(C)$ satisfy these constraints, and provide continuous and smooth transitions between these $C$ regions. The resummed perturbative cross section is independent of ${\cal O}(1)$ variations in all renormalization scales order by order in the logarithmic resummation. Therefore the dependence on parameters appearing in the profile functions gets systematically smaller as we go to higher orders, and their variation provides us with a method of assessing perturbative uncertainties. For the hard renormalization scale we use the \mbox{$C$-independent} formula \begin{align} \mu_H \,=\, e_H \, Q\,, \end{align} where $e_H$ is a parameter that we vary from $0.5$ to $2$ in order to account for theory uncertainties. The soft scale has different functional dependence in the three regions of \Eq{eq:profile-constraints}, and hence depends on the following parameters: \begin{align} \mu_S = \mu_S(C, \mu_0,r_s,\mu_H,t_0,t_1,t_2,t_s). \end{align} Here, $\mu_0$ controls the intercept of the soft scale at $C=0$, $t_0$ is near the boundary of the purely nonperturbative region and $t_1$ controls the end of this transition, where the resummation region begins. The transition from nonperturbative to perturbative is $Q$ dependent, so we use the $Q$-independent parameters $n_0 \equiv t_0\, Q/(1$\,{\rm GeV}) and $n_1 \equiv t_1\,Q/(1$\,{\rm GeV}). In the resummation region the parameter $r_s$ determines the slope of the soft scale relative to the canonical resummation region scaling, with $\mu_S = r_s \mu_H C/ 6 $. The parameter $t_2$ controls where the transition occurs between the resummation and fixed-order regions and $t_s$ sets the value of $C$ where the renormalization scales all become equal as required in the fixed-order region. For the jet scale we have the dependence \begin{equation} \mu_J = \mu_J(\mu_H, \mu_S(C),e_J)\,, \end{equation} where $e_J$ is a parameter that is varied in our theory scans to slightly modify the natural relation between the scales. The exact functional form for $\mu_S$ and $\mu_J$ in \Eqs{eq:muSprofile2}{eq:muJprofileshort} is given in Appendix~. To avoid large logarithms in the soft function subtractions $\delta$, the scale $R(C)$ is chosen to be the same as $\mu_S(C)$ in the resummation and fixed-order regions. In the nonperturbative region we need $R(C)<\mu_S(C)$ to obtain an ${\cal O}(\alpha_s)$ subtraction that stabilizes the soft function in this region (removing unphysical negative dips that appear in the $\overline{\rm MS}$ scheme). This introduces an additional parameter $R_0 \equiv R(C=0)$. Therefore we have \begin{align} R = R(\mu_S(C), R_0). \end{align} The exact functional form for $R$ is also given in Appendix~. For the nonsingular scale $\mu_{\rm ns}$, we use the variations \begin{equation} \mu_{\rm ns}(C) = \left\{\! \begin{array}{ll} \frac{1}{2} \big[\,\mu_H(C) + \mu_J (C)\,\big]\,, &~ n_s \,= ~~\,1 \\ \mu_H\,, & ~n_s \,= ~~\,0 \\ \frac{1}{2} \big[\,3\,\mu_H(C) - \mu_J (C)\,\big]\,, &~ n_s \,= -\,1 \end{array} \right.\!\!. \end{equation} Here the three choices vary $\mu_{\rm ns}$ in a manner that allows it to have some independence from $\mu_H$ in the resummation and nonperturbative regions, while still being equal $\mu_H$ in the fixed-order region (where $\mu_J=\mu_H$). These variations of $\mu_{\rm ns}$ probe the higher-order fixed-order uncertainty in the nonsingular cross section contribution. In the fixed-order region the variation of $\mu_H$ alone precisely reproduces the standard fixed-order scale variation. \begin{table}[tbh!] \begin{tabular}{ccc} parameter\ & \ default value\ & \ range of values \ \\ \hline $\mu_0$ & $1.1$\,GeV & - \\ $R_0$ & $0.7$\,GeV & - \\ $n_0$ & $12$ & $10$ to $16$\\ $n_1$ & $25$ & $22$ to $28$\\ $t_2$ & $0.67$ & $0.64$ to $0.7$\\ $t_s$ & $0.83$ & $0.8$ to $0.86$\\ $r_s$ & $2$ & $1.78$ to $2.26$\\ $e_J$ & $0$ & $-\,0.5$ to $0.5$\\ $e_H$ & $1$ & $0.5$ to $2.0$\\ $n_s$ & $0$ & $-\,1$, $0$, $1$\\ \hline $\Gamma^{\rm cusp}_3$ & $1553.06$ & $-\,1553.06$ to $+\,4659.18$ \\ $s_2^{\widetilde C}$ & $-\,43.2$ & $-\,44.2$ to $-\,42.2$ \\ $j_3$ & $0$ & $-\,3000$ to $+\,3000$ \\ $s_3^{\widetilde C}$ & $0$ & $-\,500$ to $+\,500$ \\ \hline $\epsilon^\text{low}_{2}$ & $0$ & $-\,1$, $0$, $1$ \\ $\epsilon^\text{high}_{2}$ & $0$ & $-\,1$, $0$, $1$ \\ $\epsilon^\text{low}_{3}$ & $0$ & $-\,1$, $0$, $1$ \\ $\epsilon^\text{high}_{3}$ & $0$ & $-\,1$, $0$, $1$ \\ \end{tabular} \caption{C-parameter theory parameters relevant for estimating the theory uncertainty, their default values and range of values used for the theory scan during the fit procedure. The last four parameters control the statistical errors induced by fit functions used in the non-singular terms at $\mathcal{O}(\alpha_s^2)$ ($\epsilon^\text{low}_{2}$ and $\epsilon^\text{high}_{2}$) and $\mathcal{O}(\alpha_s^3)$ ($\epsilon^\text{low}_{3}$ and $\epsilon^\text{high}_{3}$) in the region below ($\epsilon^\text{low}_{2}$ and $\epsilon^\text{low}_{3}$) and above ($\epsilon^\text{high}_{2}$ and $\epsilon^\text{high}_{3}$) the shoulder; see Sec. V of Ref.~.} \end{table} \begin{table}[tbh!] \begin{tabular}{ccc} parameter\ & \ default value\ & \ range of values \ \\ \hline $\mu_0$ & $1.1$\,GeV & -\\ $R_0$ & $0.7$\,GeV & -\\ $n_0$ & $2$ & $1.5$ to $2.5$ \\ $n_1$ & $10$ & $8.5$ to $11.5$\\ $t_2$ & $0.25$ & $0.225$ to $0.275$\\ $t_s$ & $0.4$ & $0.375$ to $0.425$\\ $r_s$ & $2$ & $1.77$ to $2.26$\\ $e_J$ & $0$ & $-\,1.5$ to $1.5$\\ $e_H$ & $1$ & $0.5$ to $2.0$\\ $n_s$ & $0$ & $-\,1$, $0$, $1$\\ \hline $j_3$ & $0$ & $-\,3000$ to $+\,3000$ \\ $s_3^{\tau}$ & $0$ & $-\,500$ to $+\,500$ \\ \hline $\epsilon_{2}$ & $0$ & $-\,1$, $0$, $1$ \\ $\epsilon_{3}$ & $0$ & $-\,1$, $0$, $1$ \\ \end{tabular} \caption{Thrust theory parameters relevant for estimating the theory uncertainty, their default values and range of values used for the theory scan during the fit procedure. The last two parameters control the statistical errors induced by fit functions used in the non-singular terms at $\mathcal{O}(\alpha_s^2)$ ($\epsilon_2$) and $\mathcal{O}(\alpha_s^3)$ ($\epsilon_3$); see Sec. E of Ref.~.} \end{table} The details of the variations of the profile function parameters used to assess uncertainties are given in Tab.~. The plot in Fig.~ shows how the scales vary with the changes to our C-parameter profile parameters. The vertical arrow on the hard scale indicates the overall up/down variation, which causes a variation to all the scales. Also shown (as gray dashed lines) are plots of the canonical soft scale $QC/6$ and canonical jet scale $Q\sqrt{C/6}$. In the resummation region, these correspond fairly well with the profile functions, indicating that in this region our analysis will avoid large logarithms. As discussed in detail in Ref.~, to improve the convergence of the C-parameter cross sections we take $r_s=2$ as our default slope parameter, explaining why our soft and jet scales are larger than the canonical values in Fig.~ (by a factor that does not induce further large logs). In the analysis of Sec.~, we will see how varying each of these profile parameters affects our final fit results. For the numerical analyses carried out in this work we have created within our collaboration two completely independent codes. One code within Mathematica~ implements the theoretical expressions exactly as given in Ref.~, and another code is based on theoretical formulae in Fourier space and realized as a fast Fortran~ code suitable for parallelized runs on computer clusters. These two codes agree for the C-parameter distribution at the level of $10^{-6}$. We will also repeat the thrust fits of Ref.~, implementing the same type of profile functions used here. These profiles have several advantages over those in Ref.~, including a free variable for the slope, a flat nonperturbative region, and parameters whose impact is much more confined to one of the three regions in \Eq{eq:profile-constraints}. For the thrust profiles we redefine \mbox{$r_s \to 6 \,r_s$}, which eliminates the factors of $6$ in \Eqs{eq:muSprofile}{eq:muRprofile}. This way, the canonical choice of slope is $r_s = 1$ for both C-parameter and thrust. We use $r_s=2$ as our default for thrust as well, again to improve the perturbative convergence, as discussed in Ref.~. The profile parameters for thrust and their variations are summarized in \tab{theoryerrthrust}. These choices create profiles that are very similar to those used in Ref.~. The new fit results for thrust are fully compatible with those of Ref.~ in the resummation region used for the $\alpha_s$ fits. Additionally, they give a better description in the nonperturbative region which is outside of our fit range. \subsection{Hadron-mass effects} In Ref.~ it was shown that hadron-mass effects, apart from breaking the universality properties of the leading power correction for various event shapes, also induce a nontrivial running. Since these are single logarithms, they start at NLL order. In Ref.~ the corresponding leading anomalous dimension was determined, which yields the NLL resummation of larger logs between the scales $\mu_S$ and $\Lambda_{\rm QCD}$ for a large set of event shapes. The pieces necessary for a higher-order resummation have not yet been computed. One might be worried that accounting for only the NLL running for $\Omega_1$ in an expression as accurate as N$^3$LL in cross-section logs could be inadequate, or that it could leave significant perturbative uncertainties. However, one should recall that the hadronic parameter $\Omega_1$ itself is a correction, and hence it is valid to account for the related resummation with less precision. In this section we show that the $\Omega_1$ evolution at NLL order suffices for the precision of our N$^3$LL$^\prime+{\cal O}(\alpha_s^3)$ analysis. Indeed, it turns out that the effect of the hadron mass running on the fit outcome is very small as compared to other uncertainties, and therefore can be safely neglected. For our C-parameter analysis the implementation of hadron mass running effects has been explained at length in Ref.~, and we only summarize here the most relevant aspects needed to understand the fit results. In the $\overline{\rm MS}$ scheme the leading power correction can be written as an integral of a universal hadron function, $\overline\Omega_1(\mu,r)$, common to all event shapes \begin{equation} \overline\Omega_1^{\,e}(\mu) \,=\, c_e\! \int_0^1\! \df r \, g_e(r)\, \overline\Omega_1(\mu,r)\,, \end{equation} where $r$ is the transverse velocity, $e$ denotes a specific event shape, $c_e$ is a calculable normalization factor, and $g_e(r)$ is an event-shape-dependent function encoding the hadron-mass effects. The functions $g_e(r)$ are known analytically, and specific examples can be found in Ref.~. For C-parameter $c_C=3\pi$, while for thrust $c_\tau =2$. For the simple case of the $\overline{\rm MS}$ scheme the running between the initial reference scale $\mu_0$ where the universal hadron function is specified, and the soft scale $\mu_S$, is given at leading order by \begin{equation} \overline \Omega_1(\mu_S,r) = \overline \Omega_1(\mu_0,r)\! \left[ \frac{\alpha_s(\mu_S)}{\alpha_s(\mu_0)} \right]^{\hat\gamma_1(r)}, \end{equation} with $\hat\gamma_1(r) = \frac{2 C_{\!A}}{\beta_0} \ln (1-r^2)$. The corresponding evolution formula for the Rgap scheme is considerably more complex, as shown by the form displayed in \Eq{eq:hadron-masses} above. Ensuring that the renormalon is not reintroduced by the renormalization group evolution requires an additional evolution in the scale $R$, so $\Delta^{\rm diff}(R_\Delta,R,\mu_\Delta,\mu,r)$ contains evolution in both the $\mu$ and $R$ scales. Also here we have two reference scales $\mu_\Delta$ and $R_\Delta$ to specify the initial parameter $\Omega_1(R_\Delta,\mu_\Delta,r)$. The full formula for $\Delta^{\rm diff}$ is given in Eq.~(67) of Ref.~. Note that the renormalization group evolution is a function of $r$ and takes place independently for each $r$, but the required result for C-parameter or thrust requires an integral over $r$. Due to this integration the functional form that we assume for the initial condition $\overline\Omega_1(r,\mu_0)$ or $\Omega_1(R_\Delta,\mu_\Delta,r)$ gets entangled with the perturbative resummation. With current constraints on the $r$ dependence, and with the lack of even more precise experimental data to probe this issue, a model-independent formulation (like a complete functional basis for the $r$ dependence at the reference scales, $R_\Delta$ and $\mu_\Delta$) is not feasible. To implement this running we have therefore tested several models for the $r$ dependence in Ref.~, and found that generically the experimental data is sensitive to the normalization which specifies $\Omega_1$, but not to the detailed form used for the $r$ dependence as long as it satisfies several reasonable constraints. Therefore for the fits performed here we simply adopt the default form from Ref.~, \begin{align} \Omega_1(R_\Delta,\mu_\Delta,r) &= [\, a(R_\Delta,\mu_\Delta) f_a(r) + b(R_\Delta,\mu_\Delta) f_b(r) \,]^2, \nn \\ f_a(r) &= 3.510 \, e^{-\frac{r^2}{1-r^2}}, \\ f_b(r) &= 13.585 \, e^{-\frac{2\,r^2}{1-r^2}} - 21.687 \,\, e^{-\frac{4\,r^2}{1-r^2}}. \nn \end{align} This model ensures that $\Omega_1(R_\Delta,\mu_\Delta,r)$ is always positive definite and smoothly goes to zero at the $r=1$ endpoint (where the ratio of the hadron mass to $p_T$ goes to zero). The functions $f_{a,b}$ form an orthonormal basis upon integration with $g_C(r)$, which yields the following relation: \begin{align} \Omega_1^C(R_\Delta,\mu_\Delta) & = 3\pi\, [\,a(R_\Delta,\mu_\Delta)^2 + b(R_\Delta,\mu_\Delta)^2\,]\,, \end{align} which determines the normalization. We also define \begin{align} \theta(R_\Delta,\mu_\Delta) & \equiv \arctan\left(\frac{b(R_\Delta,\mu_\Delta)}{a(R_\Delta,\mu_\Delta)}\right), \end{align} which was chosen to have an effect orthogonal to the more relevant parameter $\Omega_1^C(R_\Delta,\mu_\Delta)$. By examining our ability to simultaneously measure $\Omega_1^C(R_\Delta,\mu_\Delta)$ and $\theta(R_\Delta,\mu_\Delta)$ we have a means to test for the impact that the initial $r$ dependence has on our fits. As we can see in \Fig{fig:Omega1-different-thetas}, our model captures different behavior for the $r$ dependence of $\Omega_1(R_\Delta,\mu_\Delta,r)$ by choosing different values of $\theta(R_\Delta,\mu_\Delta)$. Over the interval $r\in [0,1]$, all the curves in this figure are normalized so that they integrate to $1$. \section{Experimental Data} Data on the C-parameter cross section are given by several experiments for a range of $Q$ from $35$ to $207$ GeV. We use data from ALEPH\,\footnote{The ALEPH dataset with $Q=91.2\,$GeV has two systematic uncertainties for each bin. The second of these uncertainties is treated as correlated while the first one is treated as an uncorrelated uncertainty and simply added in quadrature to the statistical uncertainty.} with $Q = \{91.2$, $133$, $161$, $172$, $183$, $189$, $200$, $206\}$ GeV , DELPHI with $Q = \{45$, $66$, $76$, $89.5$, $91.2$, $93$, $133$, $161$, $172$, $183$, $189$, $192$, $196$, $200$, $202$, $205$, $207\}$ GeV \cite{Abdallah:2004xe, Abreu:1996mk, Abreu:1999rc, Wicke:1999zz}, JADE with $Q=\{35$, $44\}$ GeV , L3 with $Q=\{91.2$, $130.1$, $136.1$, $161.3$, $172.3$, $182.8$, $188.6$, $194.4$, $200.2$, $206.2\}$ GeV , OPAL with $Q=\{91$, $133$, $177$, $197\}$ GeV , and SLD with $Q=91.2$ GeV . As each of these datasets is given in binned form, our cross section in Eq.~() is integrated over each bin before being compared to the data. The default range on $C$ used in fitting the data is $25\,\text{ GeV}/Q \le C \le 0.7$. A lower limit of $25\text{ GeV}/Q$ eliminates the peak region where higher nonperturbative moments $\Omega_{n>1}^C$ become important. The upper limit is chosen to be 0.7 in order to avoid the far-tail region as well as the Sudakov shoulder at $C=0.75$. Any bin that contains one of the end points of our range ($C= 25\, \text{GeV}/Q$ or 0.7) is included if more than half of that bin lies within the range. Using the default range and datasets gives a total of $404$ bins. As a further test of the stability of our analysis, both this C-parameter range and the selection of datasets is varied in the numerical analysis contained in Sec.~. In our fitting procedure, we consider both the statistical and systematic experimental uncertainties. The statistical uncertainties can be treated as independent between bins. The systematic experimental uncertainties come from various sources and full documentation of their correlations are not available, so dealing with them in our $\chi^2$ analysis is more complicated, and we have to use a correlation model. For this purpose we follow the LEP QCD working group~ and use the minimal overlap model. Within one C-parameter dataset, which consists of various C-parameter bins at one $Q$ value for one experiment, we take for the bin $i$, bin $j$ off-diagonal entry of the experimental covariance matrix $[{\rm min}(\Delta_i^{\rm sys},\Delta_j^{\rm sys})]^2$. Here $\Delta_{i,j}^{\rm sys}$ are the quoted systematic uncertainties of the bins $i$ and $j$. Within each dataset, this model implies a positive correlation of systematic uncertainties. In addition to this default model choice, we also do the fits assuming uncorrelated systematic uncertainties, in order to test whether the minimal overlap model introduces any bias. See Sec.~ for more details on the correlation matrix. \section{Fit Procedure} In order to accurately determine both $\alpha_s(m_Z)$ and the leading power correction in the same fit, it is important to perform a global analysis, that is, simultaneously fitting $C$-spectra for a wide range of center-of-mass energies $Q$. For each $Q$, effects on the cross sections induced by changes in $\alpha_s(m_Z)$ can be partly compensated by changes in $\Omega_1$, resulting in a fairly strong degeneracy. This is resolved by the global fit, just as in the thrust analysis of Ref.~. Fig.~ shows the difference between the theoretical prediction for the cross section at three different values of $Q$, when $\alpha_s(m_Z)$ or $\Omega_1$ are varied by $\pm\,0.001$ and $\pm\,0.05$ GeV, respectively. It is clear that the potential degeneracy in these parameters is broken by having data at multiple $Q$ values. In Fig.~ we also vary the higher-order power correction parameter $\Omega^C_2$, which clearly has a much smaller effect than the dominant power correction parameter $\Omega_1$. To carry out a fit to the experimental data we fix the profile and theory parameters to the values shown in \tab{theoryerr}. The default values are used for our primary theory cross section. We integrate the resulting theoretical distribution over the same C-parameter bins as those available experimentally, and construct a $\chi^2$ function with the uncorrelated statistical experimental uncertainties and correlated systematic uncertainties. This $\chi^2$ is a function of $\alpha_s(m_Z)$ and $\Omega_1$, and is very accurately described by a quadratic near its global minimum, which therefore determines the central values and experimental uncertainties. The value of $\Omega_1$ and its associated uncertainties encode the dominant hadronization effect as well as the dominant residual uncertainty from hadronization. To obtain the perturbative theoretical uncertainty we consider the range of values shown for the theory parameters in \tab{theoryerr}. Treating each of these as a flat distribution, we randomly generate values for each of these parameters and then repeat the fit described above with the new $\chi^2$ function. This random sampling and fit is then repeated 500 times. We then construct the minimum ellipse that fully contains all 500 of the central-fit values by first creating the convex envelope that contains all of these points within it. Then, we find the equation for the ellipse that best fits the points on the envelope, with the additional restrictions that all values lie within the ellipse and its center is the average of the maximum and minimum values in each direction. This ellipse determines the perturbative theoretical uncertainty, which turns out to be the dominant uncertainty in our fit results. In our final results the perturbative and experimental uncertainties are added in quadrature. This procedure is similar to that discussed in the Appendix of Ref.~. \section{Results} In this section we discuss the results from our global analysis. We split the presentation into several subsections. In Sec.~ we discuss the impact that resummation and the inclusion of power corrections have on the fit results. In Sec.~ we present the analysis which yields the perturbative uncertainty in detail, cross-checking our method by analyzing the order-by-order convergence. We also analyze the impact of removing the renormalon. In Sec.~ we discuss the experimental uncertainties obtained from the fit. Section~ discusses the impact that varying the theory parameters one by one has on the best-fit points, allowing us to determine which parameters dominate the theoretical uncertainty. The impact of hadron-mass resummation is discussed in detail in \Sec{sec:hadmassresum}. We examine the effects of changing the default dataset in \Sec{sec:dataset}. The final fit results are collected in Sec.~. When indicating the perturbative precision, and whether or not the power correction $\Omega_1$ is included and at what level of precision, we use the following notation: \begin{align} & {\cal O}(\alpha_s^k) & \phantom{x} & \text{fixed order up to ${\cal O}(\alpha_s^k)$} \nn\\ & \text{N}^k\text{LL}^\prime \!+\! {\cal O}(\alpha_s^k) & \phantom{x} & \text{perturbative resummation} \nn\\ & \text{N}^k\text{LL}^\prime \!+\! {\cal O}(\alpha_s^k) \!+\! {\overline \Omega}_1 & \phantom{x} & \text{$\overline{\rm MS}$ scheme for $\Omega_1$} \nn\\ & \text{N}^k\text{LL}^\prime \!+\! {\cal O}(\alpha_s^k) \!+\! {\Omega}_1(R,\mu) & \phantom{x} & \text{Rgap scheme for $\Omega_1$} \nn\\ & \text{N}^k\text{LL}^\prime \!+\! {\cal O}(\alpha_s^k) \!+\! {\Omega}_1(R,\mu,r) & \phantom{x} & \text{Rgap scheme with } \nn\\ & & \phantom{x} & \ \text{hadron masses for $\Omega_1$} \,. \nn \end{align} \begin{figure*}[t!] \begin{center} \caption{The evolution of the value of $\alpha_s(m_Z)$ adding components of the calculation. An additional $\sim 8\ \end{center} \end{figure*} \subsection{Impact of Resummation and Power Corrections} In Fig.~ we show $\alpha_s(m_Z)$ extracted from fits to the tail of the C-parameter distribution including sequential improvements to the treatment of perturbative and nonperturbative components of our code, using the highest perturbative accuracy at each stage. The sequence from left to right shows the fit results using: ${\cal O}(\alpha_s^3)$ fixed-order results only, adding N$^3$LL resummation, adding the $\overline\Omega_1$ power correction, adding renormalon subtractions and using the Rgap power correction parameter $\Omega_1(R_\Delta,\mu_\Delta)$, and adding hadron-mass effects. These same results together with the corresponding $\chi^2/{\rm dof}$ are also collected in Tab.~. The fit with only fixed-order ${\cal O}(\alpha_s^3)$ results has a relatively large $\chi^2/{\rm dof}$ and also its central value has the largest value of $\alpha_s(m_Z)$. Including the resummation of large logarithms decreases the central $\alpha_s(m_Z)$ by 8\ Note that the error bars of the first two purely perturbative determinations, shown at the left-hand side of the vertical thick dashed line in Fig.~ and in the last two entries in \tab{lesser}, do not include the $\sim 8\ The remaining corrections we consider are the use of the R-scheme for $\Omega_1$ which includes the renormalon subtractions, and the inclusion of the log-resummation effects associated to the hadron-mass effects. Both of these corrections have a fairly small impact on the determination of $\alpha_s(m_Z)$, shifting the central value by $+0.5\ The values of $\Omega_1$ obtained from the fits discussed above can be directly compared to the $\Omega_1$ power correction obtained from the thrust distribution. Values for $\Omega_1$ from the C-parameter fits are given below in \Secs{sec:random}{sec:up-down} and the comparison with thrust is considered in \Sec{sec:universality}. \begin{table}[t!] \begin{tabular}{l|c c} & $\alpha_s(m_Z)$ & $\chi^2/{\rm dof}$ \\ \hline N$^3$LL$^\prime$ + hadron & ~~$0.1119(13)(06)$ & $0.991$\\ N$^3$LL$^\prime$ with $\Omega_1(R,\mu)$ & ~~$0.1123(14)(06)$ & $0.988$\\ N$^3$LL$^\prime$ with $\overline\Omega_1$ & ~~$0.1117(16)(06)$ & $1.004$\\ N$^3$LL$^\prime$ no power corr.\,\, & ~~$0.1219(28)(02)$& $2.091$\\ \!\!\parbox{20ex}{${\cal O}(\alpha_s^3)$ fixed order\\[2pt] no power corr.} & ~~$0.1317(52)(03)$ & $1.486$ \end{tabular} \caption{ Comparison of C-parameter tail fit results for analyses when we add various components of the theoretical result (from the bottom to top). The first parentheses gives the theory uncertainty, and the second is the experimental and hadronic uncertainties added in quadrature for the first three rows, and experimental uncertainty for the last two rows. } \end{table} \subsection{Perturbative Uncertainty from the Scan} \begin{figure*}[t!] \subfigure[]{ } \subfigure[]{ } \subfigure[]{ } \subfigure[]{ } \vspace{-0.2cm} \caption{The first two panels show the distribution of best-fit points in the $\alpha_s(m_Z)$-$2\Omega_1$ and $\alpha_s(m_Z)$-$2\overline\Omega_1$ planes. Panel~(a) shows results including perturbation theory, resummation of large logs, the soft nonperturbative function and $\Omega_1$ defined in the Rgap scheme with renormalon subtractions. Panel~(b) shows the results as in panel~(a), but with $\overline\Omega_1$ defined in the $\msbar$ scheme without renormalon subtractions. In both panels the dashed lines corresponds to an ellipse fit to the contour of the best-fit points to determine the theoretical uncertainty. The respective total (experimental\,+\,theoretical) 39\ ellipses are displayed (solid lines), which correspond to $1$-$\sigma$ (68\ The big points represent the central values in the random scan for $\alpha_s(m_Z)$ and $2\,\Omega_1$. Likewise, the two panels at the bottom show the distribution of best-fit points in the $\alpha_s(m_Z)$-$\chi^2/{\rm dof}$ plane. Panel~(c) shows the $\chi^2/{\rm dof}$ values of the points given in panel~(a), whereas panel~(b) shows the $\chi^2/{\rm dof}$ values of the points given in panel~(b). } \end{figure*} To examine the robustness of our method of determining the perturbative uncertainty by the random scan, we consider the convergence and overlap of the results at different perturbative orders. Figure~ shows the spread of best-fit values at NLL$^\prime$, N$^2$LL$^\prime$ and N$^3$LL$^\prime$. The upper left panel, Fig.~, shows results from fits performed in the Rgap scheme, which implements a renormalon subtraction for $\Omega_1$, and the upper right-panel, Fig.~, shows results in the $\msbar$ scheme without renormalon subtractions. Each point in the plot represents the outcome of a single fit, and different colors correspond to different orders in perturbation theory. Not unexpectedly, fits in the Rgap scheme show generally smaller theory uncertainties. In order to estimate correlations induced by theoretical uncertainties, each ellipse in the $\alpha_s$-$2\Omega_1$ plane is constructed following the procedure discussed in \Sec{sec:fitprocedure}. Each theory ellipse constructed in this manner is interpreted as an estimate for the \mbox{1-$\sigma$} theoretical uncertainty ellipse for each individual parameter (39\ The solid lines represent the combined (theoretical plus experimental) standard uncertainty ellipses at 39\ ellipses, where the experimental ellipse corresponds to $\Delta \chi^2 = 1$. Figure~ clearly shows a substantial reduction of the perturbative uncertainties when increasing the resummation accuracy, and given that they are 39\ The results for $\alpha_s(m_Z)$ and $\Omega_1$ from the theory scan at different perturbative orders are collected in Tabs~ and . Central values here are determined from the average of the maximal and minimal values of the theory scan, and are very close to the central values obtained when running with our default parameters. The quoted perturbative uncertainties are one-parameter uncertainties. \begin{table}[t!] \begin{tabular}{ccc} order &$\alpha_s(m_Z)$ (with $\overline\Omega_1$) & $\alpha_s(m_Z)$ (with $\Omega_1(R_\Delta,\mu_\Delta)$)\\ \hline NLL$^\prime$ & $0.1071(60)(05)$ & $0.1059(62)(05)$ \\ N$^2$LL$^\prime$ & $0.1102(32)(06)$ & $0.1100(33)(06)$ \\ N$^3$LL$^\prime$ (full) & $0.1117(16)(06)$ & $\mathbf{0.1123(14)(06)}$ \end{tabular} \caption{Central values for $\alpha_s(m_Z)$ at various orders with theory uncertainties from the parameter scan (first value in parentheses), and experimental and hadronic uncertainty added in quadrature (second value in parentheses). The bold N$^3$LL$^\prime$ value is our final result.} \end{table} \begin{table}[t!] \begin{tabular}{ccc} order & \hspace{5mm}$\overline\Omega_1$ [GeV]\hspace{4mm} & \hspace{1mm}$\Omega_1(R_\Delta,\mu_\Delta)$ [GeV]\hspace{1mm} \\ \hline NLL$^\prime$ & $0.533(154)(18)$ & $0.582(134)(16)$ \\ N$^2$LL$^\prime$ & $0.443(119)(19)$ & $0.457(83)(19)$ \\ N$^3$LL$^\prime$ (full) & $0.384(91)(20)$ & $\mathbf{0.421(60)(20)}$ \\ \end{tabular} \caption{Central values for $\Omega_1$ at the reference scales $R_\Delta=\mu_\Delta=2$\,GeV and for $\overline\Omega_1$ and at various orders. The parentheses show first the theory uncertainties from the parameter scan, and second the experimental plus the uncertainty due to the imprecise determination of $\alpha_s$ (added in quadrature). The bold N$^3$LL$^\prime$ value is our final result.} \end{table} In Tab.~ above we also present $\alpha_s(m_Z)$ results with no power corrections and either using resummation or fixed-order perturbative results. Without power corrections there is no fit for $\Omega_1$, so we take the central value to be the average of the maximum and minimum value of $\alpha_s(m_Z)$ that comes from our parameter scan. Our estimate of the uncertainty is given by the difference between our result and the maximum fit value. For the fixed-order case, since there is only one renormalization scale, we know that the uncertainties from our parameter variation for $e_H$, $s_2^{\widetilde C}$, $\epsilon_2^{\rm low}$ and $\epsilon_3^{\rm low}$ are uncorrelated. So, we take the fit value for $\alpha_s(m_Z)$ with the default parameters as our result and add the uncertainties from variations of these parameter in quadrature to give the total uncertainty. An additional attractive result of our fits is that the experimental data is better described when increasing the order of the resummation and fixed-order terms. This can be seen by looking at the minimal $\chi^2/$dof values for the best-fit points, which are shown in Fig.~. In Figs.~ and we show the distribution of $\chi^2_{\rm min}/{\rm dof}$ values for the various $\alpha_s(m_Z)$ best-fit points. Figure~ displays the results in the Rgap scheme, whereas Fig.~ shows the results in the $\msbar$ scheme. In both cases we find that the $\chi^2_{\rm min}$ values systematically decrease with increasing perturbative order. The highest-order analysis in the $\msbar$ scheme leads to $\chi^2_{\rm min}/{\rm dof}$ values around unity and thus provides an adequate description of the whole dataset, however one also observes that accounting for the renormalon subtraction in the Rgap scheme leads to a substantially improved theoretical description having $\chi^2_{\rm min}/{\rm dof}$ values below unity essentially for all points in the random scan. Computing the average of the $\chi^2_{\rm min}$ values we find at N$^3$LL$^\prime$ order for the Rgap and $\msbar$ schemes $0.988$ and $1.004$, respectively (where the spread of values is smaller in the Rgap scheme). Likewise for N$^2$LL$^\prime$ we find $1.00$ and $1.02$, and for NLL$^\prime$ we find $1.09$ and $1.14$. These results show the excellent description of the experimental data for various center-of-mass energies. They also validate the smaller theoretical uncertainties obtained for $\alpha_s$ and $\Omega_1$ at N$^2$LL$^\prime$ and N$^3$LL$^\prime$ orders in the Rgap scheme. \subsection{Experimental Fit Uncertainty} Next we discuss in more detail the experimental uncertainty in $\alpha_s(m_Z)$ and the hadronization parameter $\Omega_1$ as well as the combination with the perturbative uncertainty done to obtain the total uncertainty. Results are depicted in Fig.~ for our highest order fit including resummation, power corrections and renormalon subtractions. The inner green dotted ellipse, blue dashed ellipse, and solid red ellipse represent the $\Delta \chi^2=1$ uncertainty ellipses for the experimental, theoretical, and combined theoretical and experimental uncertainties respectively. These ellipses correspond to the \mbox{one-dimensional} projection of the uncertainties onto either $\alpha_s(m_Z)$ or $\Omega_1$ (39\ theory, and total uncertainty ellipses are (for $i,j=\alpha_s, 2\,\Omega_1$), \begin{align} V_{ij}& =\, \left( \begin{array}{cc} \sigma_{\alpha_s}^2 & \,\, 2 \sigma_{\alpha_s} \sigma_{\Omega_1}\rho_{\alpha\Omega}\\ 2\sigma_{\alpha_s} \sigma_{\Omega_1}\rho_{\alpha\Omega} & \,\,4 \sigma_{\Omega_1}^2 \end{array}\right) , \\ V^{\rm exp}_{ij}& = \left( \begin{array}{cr} 4.18(52)\cdot 10^{-7} & \,\, -\,0.24(5)\cdot 10^{-4}\,\mbox{GeV}\\ -\,0.24(5)\cdot 10^{-4}\,\mbox{GeV} & \,\, 1.60(47)\cdot 10^{-3}\,\mbox{GeV}^2 \end{array}\right) \! , \nn\\ V^{\rm theo}_{ij}& = \left( \begin{array}{cr} 1.93\cdot 10^{-6} & \,\, -\,0.27\cdot 10^{-4}~\mbox{GeV}\\ -\,0.27\cdot 10^{-4}~\mbox{GeV} & \,\, 1.45\cdot 10^{-2}~\mbox{GeV}^2 \end{array}\right), \nn\\ V^{\rm tot}_{ij}& = \left( \begin{array}{cr} 2.35(5)\cdot 10^{-6} & \,\, -\,0.51(5)\cdot 10^{-4}\,\mbox{GeV}\\ -\,0.51(5)\cdot 10^{-4}\,\mbox{GeV} & \,\, 1.61(5)\cdot 10^{-2}\,\mbox{GeV}^2 \end{array}\right)\! . \nn \end{align} Note that the theoretical uncertainties dominate by a significant amount. The experimental correlation coefficient is significant and reads \begin{align} \rho^{\rm exp}_{\alpha\Omega}\,=\,-\,0.93(15) \,. \end{align} The theory correlation coefficient is small, $\rho^{\rm theo}_{\alpha\Omega}\,=\,-\,0.16$, and since these uncertainties dominate it reduces the correlation coefficient for the total uncertainty to \begin{align} \rho_{\alpha\Omega}^{\rm total} \,=\, -\,0.26(2)\,. \end{align} In both \eqs{rhoaO}{rhoaOtot} the numbers in parentheses indicate a $\pm$ range that captures all values obtained from the theory scan. The correlation exhibited by the green dotted experimental uncertainty ellipse in Fig.~ is given by the line describing the semimajor axis \begin{align} \frac{\Omega_1}{30.84\,{\rm GeV}} = 0.1257 - \alpha_s(m_Z) \,. \end{align} Note that extrapolating this correlation to the extreme case where we neglect the nonperturbative corrections ($\Omega_1=0$) gives $\alpha_s(m_Z)\to 0.1257$ which is consistent with the $0.1219 \pm 0.0028$ result of our fit without power corrections in \tab{lesser}. \begin{figure*}[t!] \begin{center} ~~~ \caption{Variations of the best-fit values for $\alpha_s(m_Z)$ and $\Omega_1$ from up (dark blue) and down (light green) variations for the theory parameters according to Tab.~. We do not display those parameters which do not affect the fit region ($\epsilon_2^{\rm high}$, $\epsilon_3^{\rm high}$, $\mu_0$, $R_0$, $n_0$).} \end{center} \end{figure*} From $V_{ij}^{\rm exp}$ in Eq.~() it is possible to extract the experimental uncertainty for $\alpha_s(m_Z)$ and the uncertainty due to the imprecise determination of $\Omega_1$, \begin{align} \sigma_{\alpha_s}^{\rm exp} & = \,\sigma_{\alpha_s}\sqrt{1-\rho^2_{\alpha\Omega}} = \,0.0002 \,, \nonumber\\ \sigma_{\alpha_s}^{\rm \Omega_1} & = \,\sigma_{\alpha_s}\, |\rho_{\alpha\Omega}|\, = \,0.0006 \,, \end{align} and to extract the experimental uncertainty for $\Omega_1$ and its uncertainty due to the imprecise determination of $\alpha_s(m_Z)$, \begin{align} \sigma_{\Omega_1}^{\rm exp} & = \,\sigma_{\Omega_1}\sqrt{1-\rho^2_{\alpha\Omega}} = \,0.014~\mbox{GeV} \,, \nn \\ \sigma_{\Omega_1}^{\rm \alpha_s} & = \,\sigma_{\Omega_1}\, |\rho_{\alpha\Omega}|\, = \,0.037~\mbox{GeV} \,. \end{align} The projections of the outer solid ellipse in Fig.~\, show the total uncertainty in our final one-parameter results obtained from $V_{ij}^{\rm tot}$, which are quoted below in \eq{asfinal}. \subsection{Individual Theory Scan Errors} To gain further insight into our theoretical precision and in order to estimate the dominant source for theory uncertainty from missing higher-order terms, we look at the size of the theory uncertainties caused by the individual variation of each one of the theory parameters included in our random scan. In Fig.~ two bar charts are shown with these results for $\alpha_s(m_Z)$ (left panel) and $\Omega_1(R_\Delta,\mu_\Delta)$ (right panel) for fits corresponding to our best theoretical setup (with N$^3$LL$^\prime$ accuracy and in the Rgap scheme). The dark blue bars correspond to the result of the fit with an upward variation of the given parameter from \tab{theoryerr}, while the light green bars correspond to the fit result from the downward variation in \tab{theoryerr}. Here we vary a single parameter keeping the rest fixed at their default values. We do not show parameters that have a negligibly small impact in the fit region, e.g.\ $\epsilon_2^{\rm high}$ and $\epsilon_3^{\rm high}$, which only have an effect on the cross section to the right of the shoulder, or $n_0$, which only affects the cross section in the nonperturbative region. We see that the dominant theory uncertainties are related to variations of the profile functions ($e_H, r_s, e_J, t_2$), where $e_H$ is the largest source of uncertainty, and is particularly dominant for $\Omega_1$. The second most important uncertainty comes from $r_s$ for $\alpha_s$ and $t_2$ for $\Omega_1$, and $e_J$ also has a significant effect on both parameters. As expected, the parameters associated to the transitions on the sides of our fit region, $n_1$ and $t_s$, hardly matter. The renormalization scale parameter $n_s$ for the nonsingular partonic distribution ${\rm d}\hat\sigma_{\rm ns}/{\rm d}C$ also causes a very small uncertainty since the nonsingular terms are always dominated by the singular terms in our fit region. The uncertainties related to the numerical uncertainties of the perturbative constants ($s_2^{\widetilde C}$, $s_3^{\widetilde C}$, $j_3$) as well as the numerical uncertainties in the extraction of the nonsingular distribution for small $C$ values, ($\epsilon_2^{\rm low}$, $\epsilon_3^{\rm low}$) are -- with the possible exception of $j_3$ -- much smaller and do not play an important role. The uncertainty related to the unknown $4$-loop contribution to the cusp anomalous dimension is always negligible. Adding quadratically the symmetrized individual uncertainties shown in Fig.~, we find $0.0007$ for $\alpha_s$ and $0.05$ GeV for $\Omega_1$. This is about one half of the theoretical uncertainty we have obtained by the theory parameter scan for $\alpha_s$ (or five sixths for $\Omega_1$), demonstrating that incorporating correlated variations through the theory parameter scan represents a more realistic method to estimate the theory uncertainty. \subsection{Effects of $\mathbf{\Omega_1}$ hadron-mass resummation} The fit results presented in the previous two sections ignored the small hadron-mass effects. These effects are analyzed in greater detail in this section. We again perform $500$ fits for a theory setup which includes N$^3$LL$^\prime$ accuracy and a power correction in the Rgap scheme, but this time it also includes hadron-mass-induced running. Since the impact of hadron-mass effects is small, one finds that the experimental data in the tail of the distribution is not accurate enough to fit for $\theta(R_\Delta,\mu_\Delta)$ in \Eq{eq:thetadef}, in addition to $\alpha_s(m_Z)$ and $\Omega_1(R_\Delta,\mu_\Delta)$. This is especially true because it enters as a small modification to the power correction, which by itself is not the dominant term. Indeed, fitting for $a(R_\Delta,\mu_\Delta)$ and $b(R_\Delta,\mu_\Delta)$ as defined in Eq.~() gives a strongly correlated determination of these two parameters. The dominant hadronic parameter $\Omega_1^C(R_\Delta,\mu_\Delta)$, which governs the normalization, is still as accurately determined from data as the $\Omega_1$ in Tab.~. However, the orthogonal parameter $\theta(R_\Delta,\mu_\Delta)$ is only determined with very large statistical uncertainties. As discussed in Ref.~, the specific value of $\theta(R_\Delta,\mu_\Delta)$ has a very small impact on the cross section, which is consistent with the inability to accurately fit for it. The results of our fit including hadron-mass effects are \begin{align} &\alpha_s(m_Z) = 0.1119 \pm 0.0006_{\rm exp+had} \pm 0.0013_{\rm pert}\,,\\[1mm] &\Omega_1(R_\Delta,\mu_\Delta) = 0.411 \pm 0.018_{\rm exp+\alpha_s} \pm 0.052_{\rm pert}\,{\rm GeV}\,.\nn \end{align} Note that the meaning of $\Omega_1(R_\Delta,\mu_\Delta)$ here is different from the case in which hadron-mass running effects are ignored because there are extra evolution effects needed to translate this value to that used in the cross section at a given value of $C$, compared to the no-hadron-mass case. In Fig.~ we compare the outcome of the $500$ fits at N$^3$LL$^\prime$ in the Rgap scheme. Results with hadron-mass effects give the red ellipse on the left, and without hadron-mass effects give the blue ellipse on the right. (The latter ellipse is the same as the one discussed above in Sec.~.) The effects of hadron masses on $\alpha_s(m_Z)$ are to decrease its central value by $0.3\ Since the theory uncertainties become slightly smaller when hadron-mass effects are incorporated, one could use this setup as our default. However we take a more conservative approach and consider the $0.3\ In \App{ap:subtractions} we also consider fits performed using the Rgap scheme with C-parameter gap subtractions, rather than our default Rgap scheme with thrust gap subtractions. The two results are fully compatible. As discussed in Ref.~ the thrust gap subtractions give better perturbative convergence, and hence are used for our default cross section. \subsection{Dataset dependence} In this section we discuss how much our results depend on the dataset choice. Our default global dataset accounts for all experimental bins for $Q\ge 35\,$GeV in the intervals $[\,C_{\rm min}, C_{\rm max}\,]=[\,25/Q,0.7\,]$, (more details are given in Sec.~). The upper limit in this range is motivated by the fact that we do not want to include data too close to the shoulder, since we do not anticipate having the optimal theoretical description of this region. The lower limit avoids including data too close to the nonperturbative region, which is near the cross section peak for $Q=m_Z$, since we by default only include the leading power correction $\Omega_1$ in the OPE of the shape function. To consider the impact of this dataset choice we can vary the upper and lower limits used to select the data. In Fig.~ the best fits and the respective total experimental + theory $68\ Interestingly all uncertainty ellipses have very similar correlation and are lined up approximately along the line \begin{align} \frac{\Omega_1}{41.26 \,{\rm GeV}} = 0.1221 - \alpha_s(m_Z) \,. \end{align} As expected, the results of our fits depend only weakly on the $C$ range and the size of the global datasets, as shown in Fig.~. The size and tilt of the total uncertainty ellipses is very similar for all datasets (with the exception of $[\,22/Q, 0.7\,]$, which clearly includes too much peak data). Since the centers and the sizes of the uncertainty ellipses are fully statistically compatible at the $1$-$\sigma$ level, this indicates that our theory uncertainty estimate at N$^3$LL$^\prime$ really reflects the accuracy at which we are capable of describing the different regions of the spectrum. Therefore a possible additional uncertainty that one could consider due to the arbitrariness of the dataset choice is actually already represented in our final uncertainty estimates. \vspace*{-0.3cm} \subsection{Final Results} \vspace*{-0.3cm} As our final result for $\alpha_s(m_Z)$ and $\Omega_1$, obtained at N$^3$LL$^\prime$ order in the Rgap scheme for $\Omega_1(R_\Delta,\mu_\Delta)$, we get \begin{align} \alpha_s(m_Z) & \, = \, 0.1123 \,\pm\, 0.0002_{\rm exp} \\[1mm] & \,\pm\, 0.0007_{\rm hadr} \,\pm \, 0.0014_{\rm pert}, \nonumber\\[2mm] \Omega_1(R_\Delta,\mu_\Delta) & \, = \, 0.421 \,\pm\, 0.007_{\rm exp} \nonumber\\[1mm] & \,\pm\, 0.019_{\rm \alpha_s(m_Z)} \,\pm \, 0.060_{\rm pert}\,\mbox{GeV}, \nonumber \end{align} where $R_\Delta=\mu_\Delta=2$~GeV and we quote individual \mbox{$1$-$\sigma$} uncertainties for each parameter. Here $\chi^2/\rm{dof}=0.99$. Equation~() is the main result of this work. Equation~() accounts for the effect of hadron mass running through an additional (essentially negligible) uncertainty. Also, it neglects QED and finite bottom-mass corrections, which were found to be small effects in the corresponding thrust analysis in Ref.~. Given that we treat the correlation of the systematic experimental uncertainties in the minimal overlap model, it is useful to examine the results obtained when assuming that all systematic experimental uncertainties are uncorrelated. At N$^3$LL$^\prime$ order in the Rgap scheme the results that are analogous to Eq.~() read $\alpha_s(m_Z) = 0.1123 \pm 0.0002_{\rm exp} \pm 0.0007_{\rm hadr} \pm 0.0012_{\rm pert}$ and $\Omega_1(R_\Delta,\mu_\Delta) = 0.412 \,\pm\, 0.007_{\rm exp} \pm 0.022_{\alpha_s} \pm 0.061_{\rm pert}$\,GeV with a combined correlation coefficient of $\rho_{\alpha\Omega}^{\rm total}=-\,0.091$. The results are compatible with Eq.~(), indicating that the ignorance of the precise correlation of the systematic experimental uncertainties barely affects the outcome of the fit. In Fig.~ we show the theoretical fit for the \mbox{C-parameter} distribution in the tail region, at a center-of-mass energy corresponding to the $Z$-pole. We use the best-fit values given in Eq.~(). The band corresponds to the perturbative uncertainty as determined by the scan. The fit result is shown in comparison with experimental data from DELPHI, ALEPH, OPAL, L3 and SLD. Good agreement is observed for this spectrum, as well as for spectra at other center of mass values. \section{Peak and Far Tail Predictions} Even though our fits were performed in the resummation region which is dominated by tail data, our theoretical results also apply for the peak and far-tail regions. As an additional validation for the results of our global analysis in the tail region, we use the best-fit values obtained for $\alpha_s$ and $\Omega_1$ to make predictions in the peak and the far-tail regions where the corresponding data was not included in the fit. Predictions from our full N$^3$LL$^\prime$ code in the Rgap scheme for the C-parameter cross section at the $Z$-pole in the peak region are shown in Fig.~. The nice agreement within theoretical uncertainties (blue band) with the precise data from DELPHI, ALEPH, OPAL, L3, and SLD indicates that the value of $\Omega_1$ obtained from the fit to the tail region is the dominant nonperturbative effect in the peak. The small deviations between the theory band and the experimental data can be explained due to the fact that the peak is also sensitive to higher-order power corrections $\Omega_{k\ge 2}^C$, which have not been tuned to reproduce the peak data in our analysis. In Fig.~ we compare predictions from our full N$^3$LL$^\prime$ code in the Rgap scheme to the accurate DELPHI, ALEPH, L3, and SLD data at $Q=m_Z$ in the far-tail region.\footnote{The OPAL data was excluded from the plot because its bins are rather coarse in this region, making it a bad approximation of the differential cross section.} We find excellent agreement with the data within the theoretical uncertainties (blue band). The key feature of our theoretical prediction that matters most in the far-tail region is the merging of the renormalization scales toward $\mu_S=\mu_J=\mu_H$ at $C\sim 0.75$ in the profile functions. This is a necessary condition for the cancellations between singular and nonsingular terms in the cross section to occur above the shoulder region.\footnote{It is worth mentioning that in the far-tail region we employ the $\overline{\rm MS}$ scheme for $\Omega_1$, since the subtractions implemented in the Rgap scheme clash with the partonic shoulder singularity, resulting in an unnatural behavior of the cross section around $C=0.75$. The transition between the Rgap and $\overline{\rm MS}$ schemes is performed smoothly, by means of a hybrid scheme which interpolates between the two in a continuous way. This hybrid scheme has been discussed at length in Ref.~.} At $Q=m_Z$ the theoretical cross section presented here obtains accurate predictions in the region both below and above the shoulder that agree with the data. Our analysis does not include the full ${\cal O}(\alpha_s^k \Lambda_{\rm QCD}/Q)$ power corrections (for $k < 4$), since they are not part of our master formula. Nevertheless, and in analogy with what was found in the case of thrust, agreement with the experimental data seems to indicate that these missing power corrections may be smaller than naively expected. \section{Universality and Comparison to Thrust} An additional prediction of our theoretical formalism is the universality of $\Omega_1$ between the thrust and \mbox{C-parameter} event shapes. Therefore, a nontrivial test of our formalism can be made by comparing our result for $\Omega_1$ with the determination from the earlier fits of the thrust tail distributions in Ref.~ and the first moment of the thrust distribution in Ref.~. Since we have updated our profiles for thrust, it is expected that the outcome of the $\alpha_s$ and $\Omega_1$ determination is slightly (within theoretical uncertainties) different from that of Ref.~. We also have updated our code to match that of Ref.~ (higher statistics for the two-loop nonsingular cross sections and using the exact result for the two-loop soft function non-logarithmic constant). In addition we have corrected the systematic uncertainty for the ALEPH data, $Q = 91.2$\,GeV of Ref.~. we assumed that two quoted uncertainties where asymmetric uncertainties, but it turns out they are two sources of systematic uncertainties that need to be added in quadrature. This has no significant effect on the results of Ref.~.} When we compare thrust and C-parameter we neglect bottom-mass and QED effects in both event shapes. In this setup, we find an updated result for thrust: \begin{align} \alpha_s(m_Z) & \, = 0.1134 \pm 0.0002_{\rm exp} \\[1mm] & \pm 0.0005_{\rm hadr} \pm 0.0011_{\rm pert}, \nonumber\\[2mm] \Omega_1(R_\Delta,\mu_\Delta) & \, = 0.329 \pm 0.009_{\rm exp} \nonumber\\[1mm] & \pm 0.021_{\rm \alpha_s(m_Z)} \pm 0.060_{\rm pert}\,\mbox{GeV}. \nonumber \end{align} For completeness we also quote an updated thrust result when both QED and bottom-mass effects are taken into account: \begin{align} \alpha_s(m_Z) & \, = 0.1128 \pm 0.0002_{\rm exp} \\[1mm] & \pm 0.0005_{\rm hadr} \pm 0.0011_{\rm pert}, \nonumber\\[2mm] \Omega_1(R_\Delta,\mu_\Delta) & \, = 0.322 \pm 0.009_{\rm exp} \nonumber\\[1mm] & \pm 0.021_{\rm \alpha_s(m_Z)} \pm 0.064_{\rm pert}\,\mbox{GeV}. \nonumber \end{align} Both the results in Eqs.~() and () are fully compatible at 1-$\sigma$ with those in Ref.~, as discussed in more detail in \App{ap:thrustresults}. When testing for the universality of $\Omega_1$ between thrust and C-parameter, there is an important calculable numerical factor of $3\pi/2=4.7$ between $\Omega_1^\tau$ and $\Omega_1^C$ that must be accounted for; see \Eq{eq:O1univ}. If we instead make a direct comparison of $\Omega_1^\tau$ and $\Omega_1^C$, as shown in Fig.~ (lowest blue ellipse vs uppermost green ellipse, respectively) then the results are $4.5$-$\sigma$ away from each other. Accounting for the $3\pi/2$ factor to convert from $\Omega_1^C$ to $\Omega_1^\tau$ the upper green ellipse becomes the centermost red ellipse, and the thrust and C-parameter determinations agree with one another within uncertainties. Due to our high-precision control of perturbative effects, the $\Omega_1$ parameters have only $\sim 15\ A zoomed-in version of this universality plot is shown in Fig.~. The upper red ellipse again shows the result from fits to the C-parameter distribution, while the lower blue ellipse shows the result from thrust tail fits. For both we show the theory uncertainty (dashed lines) and combined theoretical and experimental (dotted lines) 39\ as well as the solid ellipses which correspond to \mbox{$\Delta\chi^2=2.3$} which is the standard 1-$\sigma$ uncertainty for a two-parameter fit (68\ the two analyses are completely compatible at the 1-$\sigma$ level. An important ingredient to improve the overlap is the fact that we define the power corrections in the renormalon-free Rgap scheme. This is shown by contrasting the Rgap result in Fig.~ with the overlap obtained when using the ${\overline {\rm MS}}$ scheme for $\Omega_1$, as shown in Fig.~. \section{Conclusions and Comparison to Other $\mathbf{\alpha_s}$ Determinations} In this paper an accurate determination of $\alpha_s$ from fits to the C-parameter distribution in the resummation region was presented. We fit to the tail of the distribution defined by \mbox{$3\pi\Lambda_{\rm QCD}/Q \ll C \lesssim 3/4$}, where the dominant hadronization effects are encoded in the first moment of the shape function $\Omega_1$, which is a power correction to the cross section. By fitting to data at multiple $Q$'s, the strong coupling $\alpha_s(m_Z)$ and $\Omega_1$ can be simultaneously determined. The key points to our precise theoretical prediction are: a) higher-order resummation accuracy (N$^3$LL$^\prime$), achieved through an SCET factorization theorem, b) $\mathcal{O}(\alpha_s^3)$ matrix elements and fixed-order kinematic power corrections, c) field-theoretical treatment of nonperturbative power corrections, and d) switching to a short-distance Rgap scheme, in which the sensitivity to infrared physics is reduced. As our final result from the \mbox{\it C-parameter global fit} we obtain, \begin{align} \alpha_s(m_Z) & = \, 0.1123 \pm 0.0015\,,\\ \Omega_1(R_\Delta,\mu_\Delta) & = \, 0.421 \pm 0.063\,\mbox{GeV},\nn \end{align} where $\alpha_s$ is defined in the $\msbar$ scheme, and $\Omega_1$ in the Rgap scheme (without hadron-mass effects) at the reference scales $R_\Delta=\mu_\Delta=2$\,GeV. Here the respective total $1$-$\sigma$ uncertainties are shown. The results with individual $1$-$\sigma$ uncertainties quoted separately for the different sources of uncertainties are given in Eq.~(). Neglecting the nonperturbative effects incorporated by $\Omega_1$, the fit yields $\alpha_s(m_Z)=0.1219$ which exceeds the result in Eq.~() by $8\ \begin{align} \alpha_s(m_Z) & \, = \, 0.1128 \,\pm\, 0.0012\,,\\ \Omega_1(R_\Delta,\mu_\Delta) & \, = \, 0.322 \,\pm\, 0.068\,\mbox{GeV}.\nn \end{align} Our theoretical prediction is the most complete treatment of C-parameter at this time, and, to the best of our knowledge, all sources of uncertainties have been included in our final uncertainty. Possible improvements which are expected to be negligible relative to our final uncertainty include finite bottom-mass effects, QED effects, and axial-singlet contributions. From our results there are a number of theoretical avenues that lead to small effects but which would be interesting to investigate further in the future. These are common to almost every event-shape analysis in the literature and include (i) resummation of logarithms for the nonsingular partonic cross section; (ii) the structure of nonperturbative power corrections for the nonsingular contributions (the last two points can be clarified with subleading SCET factorization theorems); (iii) analytic perturbative computations of the ${\cal O}(\alpha_s^3)$ nonlogarithmic coefficients in the partonic soft function and the jet function, as well as the four-loop QCD cusp anomalous dimension (and to a lesser extent, the numerically determined $s_2^{\widetilde C}$ constant of the two-loop partonic soft function); (iv) a better understanding of hadron-mass effects, and in particular their resummation beyond NLL; (v) a better theoretical description of the region around and above the shoulder. Concerning (i), and following the common lore, we have incorporated in our analysis the nonsingular contributions in fixed-order perturbation theory. However we have estimated the uncertainty related to the higher-order logarithms through the usual renormalization scale variation. Concerning (ii) we observe that the effect of these neglected power corrections is much smaller than naively expected, as can be seen from a comparison of our theoretical prediction and LEP data in the far-tail region. A first step towards clarifying (i) and (ii) has been taken in Refs.~, for the case of thrust. The computation of missing perturbative terms (iii) is a priori feasible with current computational knowledge but they do not dominate our perturbative uncertainties. Concerning (iv) we have shown that hadron-mass effects have a very small impact on the determination of $\alpha_s$, and hence unless the rest of the sources of uncertainty become substantially smaller, our lack of knowledge does not constitute a problem. As for (v), our fits do not include data above the shoulder, so this problem has no impact on our fit. Nevertheless an analysis of these subleading effects would be interesting. The same theoretical program carried out for thrust and C-parameter can be applied to other event-shapes, and the most prominent one is Heavy-Jet-Mass. This has partially worked out already in Ref.~ at the purely perturbative level using fully canonical profiles. Their determination of $\alpha_s$ is discussed below. For recoil-sensitive observables such as Jet Broadening~, one needs to deal with rapidity singularities, which imply that additional logs need to be resummed, and more complicated nonperturbative power corrections. The former has been pushed to the N$^2$LL order in Ref.~, and the latter has been studied in Ref.~. Recoil-insensitive versions of Broadening have also been derived~, but not yet studied experimentally. Finally, it is very straightforward to generalize our theoretical treatment to the case of oriented event shapes~, in which one additionally measures the angle between the beam and the thrust axes. At this point we compare our result for $\alpha_s$ with other determinations from event shapes at $\mathcal{O}(\alpha_s^3)$. To the best of our knowledge, the only analyses which fit to the tail of the C-parameter distribution using three-loop input are Ref.~ (using purely fix-order perturbation theory) and Ref.~ (including NLL resummation). Both analyses use Monte Carlo (MC) event generators to estimate hadronization effects, and fit $\alpha_s$ for different $Q$ values, finding values $\alpha_s(m_Z)=0.1288\,\pm\, 0.0043$ and $0.1252\,\pm\, 0.0053$ respectively for a fit to the \mbox{$Q = 91.2\,$GeV} data. These larger $\alpha_s(m_Z)$ values are consistent with our fits which neglect power corrections, and following Ref.~ we can conclude from this that MCs does not provide a reasonable estimate of the power corrections when including the higher-order perturbative contributions. In Ref.~ two-parameter global fits to the first five moments of the C-parameter distribution were performed. Hadronization effects are included via the frozen coupling model, and the value obtained, $\alpha_s(m_Z) = 0.1181\pm 0.0048$, is fully consistent with our result in \Eq{eq:allerror} at 1-$\sigma$. A graphical comparison with other event-shape determinations is shown in Fig.~. The figure includes determinations where power corrections are estimated with MC generators, labeled by 7-10. Analyses 1-6 correspond to those in which power corrections were incorporated with an analytic method (either a shape function or the dispersive model). In the analyses 1-6 global fits are performed, whereas in the 7-10 analyses $\alpha_s$ was determined at multiple $Q$ values and the final result is an average of those. Only analyses 1~, 3 (this work), and 4~ used a completely field-theoretical approach for the power corrections. We also show both results from fits to the event-shape distributions (3-10) and from fits to moments of the event shape distribution (1 and 2). Although all analyses included $\mathcal{O}(\alpha_s^3)$ matrix elements, different levels of resummation have been achieved. Analyses 2~ and 10~ did not included resummation; 6~ and 9~ included NLL resummation; 5~ include N$^2$LL resummation; and analyses 3,4,7, and 8 included N$^3$LL resummation. Analyses 2, 9, and 10 simultaneously fit to many event shapes, whereas the others focused on a single observable: thrust (1, 4-6 and 8~), Heavy-Jet-Mass (7~), and \mbox{C-parameter} (3, which is this work). The analyses 1, 3, 4, 7 and 8 used SCET to perform Sudakov log resummation. All results that used an analytic treatment of power corrections have smaller values of $\alpha_s$. This is consistent with a simple dimensional analysis argument (see Refs.~). Higher order resummation results in a convergent perturbation series and smaller uncertainties, and the Rgap scheme also reduces uncertainties. Accounting for the fact that results relying on MC for the treatment of power corrections should likely have larger hadronization uncertainties, all results are compatible among one another. The most precise results are however clearly in disagreement with the world average, which is dominated by lattice QCD results (see below) and shown as a translucent green band. We conclude this work by comparing our result for $\alpha_s(m_Z)$ with the results of a selection of recent analyses using other techniques and observables, as shown in Fig.~. We include a N$^3$LO analysis of data from deep inelastic scattering from the ABM group~, the global PDF fits of the MSTW group~ and the NNPDF collaboration~; the most recent (and accurate) determination from the HPQCD lattice collaboration~, from the analysis of Wilson loops and pseudoscalar correlators; a determination analyzing the lattice prediction for the QCD static potential ; a reanalysis of electroweak precision observables by the Gfitter collaboration~; the most recent analysis of tau decays in which the recently released ALEPH data was used together with the OPAL data; the previous determinations from fits to the thrust distribution~ and moments of the thrust distribution~; and of course the current world average~ (shown as the green band). The ABM (DIS) and thrust results are compatible with our determination, while in contrast the disagreement with either lattice QCD or the world average is $4$-$\sigma$. Many other determinations lie between these two values. The source of this disagreement is an important open question. \begin{acknowledgments} We thank the Erwin-Schr{\"o}dinger Institute (ESI) for partial support in the framework of the ESI program ``Jets and Quantum Fields for LHC and Future Colliders''. This work was supported by the offices of Nuclear and Particle Physics of the U.S. Department of Energy (DOE) under Contract DE-SC0011090, and the European Community's Marie-Curie Research Networks under contract PITN-GA-2010-264564 (LHCphenOnet). IS was also supported in part by the Simons Foundation Investigator grant 327942. VM was supported by a Marie Curie Fellowship under contract PIOF-GA-2009-251174 while part of this work was completed. AH, VM, and IS are also supported in part by MISTI global seed funds. VM thanks H.~Stenzel for explaining how to interpret the systematic uncertainties in the ALEPH dataset and Thomas Hahn for computer support. \end{acknowledgments} \appendix \section{Profile Formulae} In this appendix, we give the details for the profile functions that control the renormalization scales as laid out in \Sec{subsec:profiles}. For the soft profile function, we use the form, \begin{equation} \!\!\!\mu_{S} = \left\{ \begin{tabular}{p{.5\columnwidth} l} $\mu_0$ & $0 \le C < t_0$ \\ $\zeta(\mu_0,\,0,\,0,\,\frac{r_s\,\mu_H}{6},\,t_0,\,t_1,\,C)$ & $t_0 \le C < t_1$ \\ $r_s \,\mu_H\, \frac{C}{6}$ & $t_1 \le C < t_2$ \\ $\zeta(0,\,\frac{r_s\,\mu_H}{6},\,\mu_H,0,\,t_2,\,t_s,\,C)$ & $t_2 \le C < t_s$ \\ $\mu_H$ & $t_s \le C < 1$ \end{tabular} \right.\!, \end{equation} where the physical meaning of the parameters is explained in Sec.~. The function $\zeta(a_1,b_1,a_2,b_2,t_1,t_2,\,t)$ (with $t_1 < t_2$), which smoothly connects two straight lines of the form $l_1(t) = a_1 \,+\, b_1\,t$ for $t < t_1$ and $l_2(t) = a_2 \,+\, b_2\,t$ for $t > t_2$ is given by \begin{align} \zeta(t) &= \left\{ \!\begin{tabular}{p{.5\columnwidth} l} $\hat a_1 + b_1(t - t_1) + e_1(t - t_1)^2$ & $~t_1 \le t \le t_m$ \\ $\hat a_2 + b_2(t - t_2) + e_2(t - t_2)^2$ & $~t_m \le t \le t_2$ \end{tabular}\nonumber \right.\!,\\[0.1cm] \hat a_1 & = a_1 + b_1\,t_1\,,\qquad \hat a_2 = a_2 + b_2\,t_2\,,\\ e_1 &=\frac{4\,(\hat a_2-\hat a_1)-(3\,b_1 + b_2)\,(t_2-t_1)}{2\,(t_2-t_1)^2}\,,\nn\\[0.1cm] e_2 &=\frac{4\,(\hat a_1-\hat a_2)+(3\,b_2 + b_1)\,(t_2-t_1)}{2\,(t_2-t_1)^2}\,.\nn \end{align} For the jet scale, we use the form \begin{equation} \!\!\!\!\!\mu_J(C) = \left\{\!\! \begin{array}{lr} \big[\,1 + e_J (C-t_s)^2\,\big] \sqrt{ \mu_H\, \mu_{S} (C)} & C \le t_s\\ \,\mu_H & C > t_s \end{array} \right.\!, \end{equation} which allows a slight modification of the natural relation between the scales $\mu_J = \sqrt{\mu_H \mu_S}$ in order to account for theoretical uncertainties. For the subtraction scale, we have \begin{equation} \!\!\!\!\!R(C) = \left\{\!\! \begin{array}{ll} R_0 & 0 \le C < t_0 \\ \zeta(R_0,\,0,\,0,\,\frac{r_s\,\mu_H}{6},\,t_0,\,t_1,\,C) & t_0 \le C < t_1 \\ \mu_S(C) & t_1 \le C \le 1 \end{array} \right.\!\!. \end{equation} As explained earlier, we take $R=\mu_S$ in the resummation region to avoid large logs and $R \neq \mu_S$ in the nonperturbative region to remove the renormalon. The $\zeta$ function here interpolates smoothly between these two regions. It is necessary to vary the profile parameters to estimate the theory uncertainty. We hold the difference between the parameters associated with the purely nonperturbative region constant: \mbox{$\mu_0 - R_0 = 0.4\,$GeV}, and we set as default values \mbox{$\mu_0 = 1.1$ {\rm GeV}}, $R_0 = 0.7$ {\rm GeV}. We are then left with nine profile parameters to vary during the theory scan, whose central values and variation ranges used in our analysis are: \mbox{$r_s = 2\times 1.13^{\pm1}$}, \mbox{$n_0=12\,\pm\,2$}, $n_1=25\,\pm\,3$, $t_2 = 0.67\,\pm\,0.03$, \mbox{$t_s=0.83\,\pm\,0.03$}, $e_J = 0\,\pm\, 0.5$, $e_H = 2^{\pm1}$ and $n_s = 0\pm1$. These variations are shown in Tab.~. \section{Comparison of thrust and C-parameter subtractions} In Fig.~ we compare fits performed in the Rgap scheme with C-parameter gap subtractions as the upper red ellipse, and for our default fits in the Rgap scheme with thrust gap subtractions as the lower blue ellipse. At N$^3$LL$^\prime$ order with C-parameter subtractions the results are $\alpha_s(m_Z) = 0.1126 \pm 0.0002_{\rm exp} \pm 0.0007_{\rm hadr} \pm 0.0022_{\rm pert}$ and $\Omega_1(R_\Delta,\mu_\Delta) = 0.447 \pm 0.007_{\rm exp} \pm 0.018_{\alpha_s} \pm 0.065_{\rm pert}$~GeV, with $\chi^2_{\rm min}/{\rm dof} = 0.988$. One can see that, even though both extractions are fully compatible, the thrust subtractions lead to smaller perturbative uncertainties. This is consistent with the better perturbative behavior observed for the cross section with thrust subtractions in Ref.~. \section{Comparison of thrust results with Ref.~} In Fig.~ we compare global fits for the thrust distribution using the profiles of Ref.~ (shown by the right ellipse in blue) and the profiles used here (shown by the left ellipse in red). As mentioned earlier, the profiles used here have several advantages over those of Ref.~ in terms of their ability to independently impact the different regions of the thrust distribution, and in particular do a better job in the nonperturbative region (which is outside our fit region). The two versions of the profiles are consistent within their variations, and the fit results shown for 39\ \bibliography{../thrust3} |
1501.04112 | Title: Can long-range interactions stabilize quantum memory at nonzero
temperature?
Abstract: A two-dimensional topologically ordered quantum memory is well protected
against error if the energy gap is large compared to the temperature, but this
protection does not improve as the system size increases. We review and
critique some recent proposals for improving the memory time by introducing
long-range interactions among anyons, noting that instability with respect to
small local perturbations of the Hamiltonian is a generic problem for such
proposals. We also discuss some broader issues regarding the prospects for
scalable quantum memory in two-dimensional systems.
Body: \title{Can long-range interactions stabilize quantum memory at nonzero temperature?} \author{Olivier Landon-Cardinal} \affiliation{Institute for Quantum Information and Matter and Walter Burke Institute for Theoretical Physics, California Institute of Technology, Pasadena, California 91125, USA } \author{Beni Yoshida} \affiliation{Institute for Quantum Information and Matter and Walter Burke Institute for Theoretical Physics, California Institute of Technology, Pasadena, California 91125, USA } \author{David Poulin} \affiliation{D\'{e}partement de Physique, Universit\'{e} de Sherbrooke, Qu\'{e}bec, Canada } \author{John Preskill} \affiliation{Institute for Quantum Information and Matter and Walter Burke Institute for Theoretical Physics, California Institute of Technology, Pasadena, California 91125, USA } \date{\today} \begin{abstract} A two-dimensional topologically ordered quantum memory is well protected against error if the energy gap is large compared to the temperature, but this protection does not improve as the system size increases. We review and critique some recent proposals for improving the memory time by introducing long-range interactions among anyons, noting that instability with respect to small local perturbations of the Hamiltonian is a generic problem for such proposals. We also discuss some broader issues regarding the prospects for scalable quantum memory in two-dimensional systems. \pacs{03.67.Pp, 03.67.Lx} \end{abstract} \maketitle \section{Introduction} Protecting quantum information from damage will be essential for future quantum technologies. Thus, devising a \emph{self-correcting quantum memory}, capable of storing qubits for a very long time without active error correction~, would be a scientific milestone with far-reaching implications for the scalability of quantum computing and the security of quantum communication protocols. We define a self-correcting quantum memory as a quantum many-body system with a local Hamiltonian such that (i) the ground state is degenerate in the limit of a large system, (ii) at a sufficiently small nonzero temperature, quantum information stored in the degenerate ground state is protected from error for a time which grows without bound as the system size increases, (iii) both of these features are stable with respect to generic small local static perturbations of the Hamiltonian. For a quantum memory to be useful, one will also need a {\em practical} method for storing a quantum state in memory and reading it out, but we do not impose that criterion in our definition. Rather, we regard (ii) as satisfied if it is possible in principle to recover the stored quantum state from a thermally fluctuating state. Topologically ordered systems have a perturbatively stable ground-state degeneracy~, and would meet all three of our criteria if the topological order persists at nonzero temperature~. Known mathematical models of two-dimensional topologically ordered media (such as the toric code~) typically do not satisfy (ii)~. The rate per unit area for thermal production of anyons is suppressed by the energy gap, but does not depend on the system size; once produced, thermal anyons can propagate without further energy penalty, quickly destroying the quantum information stored in the ground state~. Three-dimensional models have been proposed in which the quantum memory time is enhanced because the propagation of excitations is impeded~, but so far the known models of self-correcting quantum memory require at least four spatial dimensions~. (A recently proposed model based on a low-distortion three-dimensional embedding of a four-dimensional code~ has not been shown to be perturbatively stable.) For a recent review of quantum memories at nonzero temperature, see~. To suppress errors arising from the propagation of thermally produced anyons, Hamma \emph{et al.} proposed coupling the toric code to massless bosonic fields, inducing long-range attractive interactions among the anyons~; however, their scheme requires the coupling strength to diverge with the system size. Pedrocchi {\emph{et al.} proposed an alternative scheme, in which coupling anyons to three-dimensional massless bosons induces an infrared divergent renormalization of the anyon chemical potential, forcing the rate for thermal anyon production to approach zero as the system size grows~. Another suggestion is that disorder in a two-dimensional topological medium could localize the anyons and hence extend the memory time~. In this paper, we emphasize a general issue confronting schemes for stabilizing a two-dimensional quantum memory by coupling to massless bosons --- the criterion (iii) typically fails because generic small local perturbations introduce a mass gap for the bosons and hence compromise the scheme. We consider the model of Pedrocchi \emph{et al.}~ as an instructive example, but our arguments apply more generally, and we also assess some alternative proposals for enhancing memory times. We offer no rigorous arguments excluding self-correcting memory in two or three dimensions, but we hope our discussion helps to clarify the question. Nor do we mean to deny the possibility that clever engineering of interactions or disorder in anyonic systems could substantially enhance the memories times of anyonic systems; our focus is on whether arbitrarily long memory times are achievable as a matter of principle. This paper is organized as follows. In Section~, we discuss how Hamiltonian perturbations and thermal fluctuations can corrupt the information stored in a quantum memory, and in Section~ we recall some proposals for enhancing the storage time. In Section~, we discuss the idea of stabilizing a two-dimensional topologically ordered system at nonzero temperature by introducing long-range interactions between anyons, using the ``toric-boson model'' introduced in~ as an explicit example, and in particular we explain how to recover the conclusions of~ using an effective field theory description of the system. In Section~ and , we argue that the infinite energy barrier of the toric-boson model, and hence its thermal stability, are unstable with respect to generic perturbations of the system's microscopic Hamiltonian. Section~ contains some concluding remarks and open questions regarding self-correcting quantum memory. \section{Logical errors in a quantum memory} To clarify the concept of self-correcting quantum memory, we recall in this Section some ways for a quantum memory to fail, either due to imperfect control of the Hamiltonian or due to thermal fluctuations. \subsection{Spectral instability} Desideratum (iii) of our definition of a self-correcting memory requires the degenerate ground space to be stable with respect to small local perturbations of the Hamiltonian. If adding a small local perturbation can lift the degeneracy, then the environment might couple to this local operator, producing fluctuations in the relative phases of energy eigenstates, hence driving decoherence in the energy-eigenstate basis. This failure mechanism relies only on the properties of the nearly-degenerate ground space, not the properties of the excitations supported by the system. Alternatively, the memory could fail because a weak local perturbation destroys the energy gap separating the ground space from the rest of the energy-eigenstate spectrum. When the energy gap collapses, the structure of the ground space may undergo a qualitative change, allowing the information stored in the ground state to be corrupted. We use the term ``spectral instability'' to refer to either of these failure modes. Sufficient conditions for spectral stability, based on topological order, were identified in~ and then generalized in~. \subsection{Thermal relaxation} Desideratum (ii) of our definition requires the stored quantum information to be well protected from error at a sufficiently small nonzero temperature. Ideally, we would assess whether (ii) is satisfied by modeling the thermal environment, simulating how the stored information is affected by thermal fluctuations, and computing the fidelity of a suitably chosen decoding map. Since this program is quite difficult to carry out, and in any case the results depend on the details of the environment and decoding procedure, we usually assess (ii) using simpler criteria. \subsubsection{Energy and free energy barriers} One way to simplify the story is to imagine that the environment applies a sequence of local quantum operations to the quantum memory. After many such steps, we envision cooling the memory, so that it relaxes back to the ground-state space, and then we compare this final state to the initial state deposited in the memory. We say the error sequence is ``harmful'' if the final state deviates significantly from the initial state. Each error sequence has an energy height, the maximal value of the expectation value of the Hamiltonian attained during the sequence, and we say that the memory's \emph{energy barrier} is the lowest energy height for any harmful error sequence. Naively, a high energy barrier means a long memory time, because only rare thermal fluctuations will be able to surmount the barrier and damage the stored quantum state. An Ising ferromagnetic in two or more spatial dimensions is a thermally stable classical memory storing one bit of information in the sign of its magnetization; for the magnetization to flip, thermal fluctuations must produce a large droplet of flipped spins with size comparable to the system size, surmounting an energy barrier scaling like the droplet's surface area. The divergent memory time, though, is not perturbatively stable. A small magnetic field favors one sign of the magnetization; hence bubbles of flipped spins, arising as thermal fluctuations in the disfavored phase at a constant rate per unit time and volume, destroy the stored classical information in a constant time. The three-dimensional toric code, on the other hand, is an example of a self-correcting classical memory, as the one bit detected by its membrane-like logical operator is robust against small local perturbations of the Hamiltonian. As in the classical case, to ensure a long storage time for a quantum memory we would like the energy barrier to diverge as the system size increases. However, a divergent energy barrier is not by itself sufficient for thermal stability. Haah's three-dimensional cubic code~ has a logarithmically divergent barrier and Michnicki's welded code~ has a polynomially divergent barrier, yet in both cases the memory time is bounded above by a constant independent of the system size at any nonzero temperature~. These examples caution us that even if harmful error sequences are strongly Boltzmann suppressed, the suppression can be overwhelmed if the harmful sequences have high entropy. We should consider not just the energy cost, but rather the free energy cost of encoded errors. In the case of the cubic code, for example, the Boltzmann suppression dominates when the system size is sufficiently small; hence the logarithmically growing energy barrier implies that the memory time $\tau$ increases with system size $L$ as $\tau\sim \exp( c \beta \log L )$, where $c$ is a constant and $\beta$ is the inverse temperature. However once $L$ is larger than a critical system size $L^*\sim e^{c'\beta}$, entropy becomes dominant, and the memory time actually declines with increasing system size. The memory time is optimized for $L\sim L^*$, yielding $\tau < e^{\tilde c\beta^2}$ for some constant $\tilde c$; criterion (ii) for self correction is not satisfied. Conceivably, entropic effects might enhance the memory time under some circumstances, perhaps allowing the memory to be self correcting even though the energy barrier is finite, as suggested in~. A model in which entropy suppresses harmful error sequences was proposed in~ (see also Sec.~). \subsubsection{Anyon propagation} Typical topologically ordered systems in two dimensions support anyons --- pointlike quasiparticle excitations obeying exotic statistics. If the system resides on a torus, a logical error occurs when a pair of anyons is created, and then one member of the pair propagates around a homologically nontrivial cycle of the torus before reannihilating with its partner. Thus the energy barrier is the (constant) energy cost of creating and separating a pair of anyons. At nonzero temperature, thermal fluctuations produce a nonvanishing density of anyons per unit area. We can imagine decoding the state by bringing pairs of anyons together to annihilate, returning the system to a ground state. Starting with an initial noisy encoded state, we may allow the system to evolve for a while in contact with a heat bath, before applying the decoding map to the final configuration. By combining the anyon world lines arising from the decoding of the initial state, the decoding of the final state, and the thermally induced anyon propagation, creation, and annihilation between the initial and final time, we obtain a cycle, a set of closed loops. If this cycle is homologically nontrivial, a logical (encoded) error occurs. When the anyons propagate a distance comparable to the typical (constant) distance between particles in the thermal anyon gas, a logical error will occur with nonnegligible probability. Therefore, we do not normally expect a two-dimensional quantum memory to be self correcting, unless supplemented by an additional mechanism to impede the propagation or creation of anyons. The three-dimensional version of the toric code also has pointlike ``anyon'' excitations, and a logical error occurs when an anyon propagates across the system. Therefore this code, too, has a constant energy barrier and is not self correcting. The cubic code~ has pointlike excitations, but reaching a configuration in which one particle is separated from all others by a distance at least $R$ requires energy scaling like $\log R$; thus thermal particle propagation is suppressed and the memory has a logarithmically growing energy barrier. Nevertheless, as noted in Sec.~, the free energy cost of a logical error becomes independent of system size for a sufficiently large system at a fixed temperature. In the three-dimensional welded code~, $v > 2$ three-dimensional toric code subblocks are sewn together at each weld. Anyons propagate freely within each subblock, but for an anyon to pass through a weld from one subblock to another, at least $v-2$ additional anyons must be created. Therefore the energy barrier is high if the code has many such subblocks. However, logical errors can arise at nonzero temperature from a process in which $v$ thermal anyons propagating in distinct subblocks meet at a weld and fuse to the vacuum~. This process is entropically enhanced because the fusion can occur at any point along the weld; the resulting logical error rate is independent of system size and therefore the code is not self correcting. \section{Improving the storage time} Having discussed how a quantum memory can fail, we now briefly consider some possible ways to improve the stability of a quantum memory and hence enhance its storage time. \subsection{Anyon localization} Even for a topologically ordered system with a spectral gap, in which small local perturbations split the ground state degeneracy by an amount which is exponentially suppressed for a large system size, a quantum memory may fail due to small local perturbations of the Hamiltonian. The problem is that, although information encoded in the ground space may be well protected in principle, if the perturbation is unknown it may be difficult to prepare precisely the ground state of the perturbed system. For example, if the ideal system has the toric code Hamiltonian, we might write to the memory by preparing one of the toric code ground states. But if the Hamiltonian is slightly deformed by a weak perturbation, the ground state of the ideal Hamiltonian is a coherent superposition of excited states of the perturbed Hamiltonian; this superposition dephases quickly, destroying the encoded information in a constant time independent of system size~. If the perturbation were exactly known, then in principle we could decode the state by running the Hamiltonian evolution backward in time. But if (more realistically) the perturbation is unknown, then an uncorrectable logical error is likely to occur. This error mechanism operates even at zero temperature; the error arises not from thermal fluctuations but rather from imperfect control of the system's Hamiltonian. One way to evade this problem is to respond to the deformation of the Hamiltonian by adjusting the procedure for encoding and decoding the state. Under a generic perturbation, the static pointlike anyons of the exact toric code become propagating dressed anyons with nonzero size. Using a physical device that detects these dressed anyons, we can prepare a ground state of the perturbed system, which will store quantum information reliably at zero temperature. In principle we could read out the encoded state at a later time using dressed string operators. Alternatively, these zero-temperature errors can be suppressed by introducing disorder in the system Hamiltonian, ensuring that in low-lying excited states the anyons are spatially localized rather than freely propagating~. In fact, localization may suffice to make the memory time exponentially large in the system size, when the errors are due to unknown Hamiltonian perturbations~. However, at any nonzero temperature, {\em single-particle} localization does not prevent anyons from diffusing at a nonzero rate due to thermally activated hopping through the disordered system. Thus at nonzero temperature, though localization may enhance the memory time by a constant factor, we still expect the logical errors to occur after a constant time independent of system size --- the memory is not self correcting. We note that a disordered topologically ordered system may exhibit {\em many-body localization}, in which anyons are spatially localized even in highly excited states of the disordered Hamiltonian~. However, many-body localization is a property of a perfectly isolated Hamiltonian system. Even a many-body-localized two-dimensional topological phase would fail to be self correcting, because anyons would still be able to diffuse when the system is in contact with a thermal bath. For the case of a two-dimensional commuting Hamiltonian like the disordered toric code model, the finding that topological order is destroyed at nonzero temperature can be formulated rigorously. Hastings~ showed that the Gibbs state of a (not necessarily translation-invariant) 2D commuting local Hamiltonian can be (approximately) transformed to a product state by a constant-depth local quantum circuit. To analyze the memory time, we can draw guidance from studies of the disordered 1D Ising model, whose stability properties are similar to those of the 2D toric code. For a one-dimensional classical spin system with Hamiltonian \begin{align} H_{Ising}= -\sum_{i}J_{i}Z_{i}Z_{i+1} \end{align} where the nearest-neighbor couplings $\{J_{i}\}$ are positive but otherwise arbitrary, the memory time $\tau$ is independent of system size, scaling like $\tau \sim \exp(2\beta \bar J)$ where $\bar J\equiv(J_{min}+J_{max})/2$ is the average of the minimal and maximal values of the coupling~. Thus the memory time is not significantly enhanced in the regime $J_{max}\gg J_{min}$ where defects are strongly localized. The conclusion $\tau = O(1)$ also applies to the (disordered) toric code, subject to a master equation whose jump operators are Pauli operators; in that case the jumps transform energy eigenstates to energy eigenstates, so that the convergence to a fixed point can be analyzed using classical methods~. \subsection{Entropically suppressed propagation} Brown \emph{et al.} described an anyon model in which anyon propagation is impeded by entropic considerations~. In their model, a two-dimensional topologically ordered medium is divided into domains separated by defect lines. Each domain supports both heavy and light anyons, where the decay of a heavy anyon into two light anyons is allowed by both kinematics and the fusion rules. Furthermore, upon crossing a defect line, a heavy anyons is transformed into a light one, and a light one is transformed into a heavy one. Propagation of light anyons across defect lines is suppressed at low temperature, because an energy boost is needed to cross to the other side. Furthermore, it is entropically favored for heavy anyons to decay to pairs of light ones. Therefore, anyons cannot easily travel across the system and the memory time is correspondingly enhanced. Numerical evidence suggests that if the system size is properly optimized, the memory time $\tau$ scales with the inverse temperature $\beta$ as $\tau \sim e^{c \beta^2}$, just as in the three-dimensional cubic code, even though the energy barrier is constant rather than logarithmically growing. This is a notable improvement compared to the scaling $\tau\sim e^{c \beta}$ in the toric code. Though not self correcting, this system illustrates how in some cases favorable entropic effects can significantly enhance the memory time. \subsection{Long-range anyon interactions} Since logical errors in a two-dimensional topologically ordered quantum memory arise from anyon propagation, any mechanism that impedes anyon propagation can in principle enhance the memory time. Aside from localization due to disorder as discussed above, there is another mechanism that could limit how far anyons can travel --- attractive interactions between anyons. The interactions may impose a hefty energy penalty on harmful error sequences, encouraging confinement of the anyons to small dilute clusters which do not produce logical errors. For the system to be truly self correcting, though, these interactions must have infinite range. Otherwise only a constant amount of energy would be needed for an anyon to break away from a cluster, leaving it free to propagate unmolested by attractive forces applied by other anyons, thus allowing logical errors to occur at a constant rate at any nonzero temperature. For a system with a local Hamiltonian, long-range forces arise from the exchange of gapless excitations. Hamma \emph{et al.}~ proposed a model in which the anyon density provides a source for a massless scalar field. Scalar exchange produces long-range attractive interactions between anyons, but each anyon also has a logarithmically infrared divergent negative self energy, rendering the system unstable. The divergent self-energy can be cancelled by introducing other fields with carefully tuned couplings, but then the Hamiltonian contains terms with infrared divergent coefficients. Improving on , Pedrocchi \emph{et al.} proposed an alternative model with a local Hamiltonian and finite couplings, in which the anyonic chemical potential is infrared divergent ~. Thus the energy barrier, the cost of creating a distantly separated anyon pair, diverges with system size, as does the memory time. We review this scheme in Sec.~. It works, if the Hamiltonian is carefully tuned. But as we explain in Sec.~ it does not meet our criteria for self correction, because the divergent energy barrier is unstable with respect to generic small perturbations of the local Hamiltonian. We also note that schemes for ``active'' rather than ``passive'' error correction have been proposed, inspired by the idea that long-range attractive interactions might impede the propagation of diffusing anyons . In a scheme recently discussed in , the anyon motion is guided by a classical field, which is regularly updated using a local evolution rule as the anyons propagate. Such active error correction schemes, though effective and interesting, do not meet our criteria for self correction for several reasons. For one thing, to achieve an arbitrarily low logical error rate, the scheme requires a ``speed of light'' (rate of information propagation in the classical field) which diverges relative to the ``speed of sound'' (rate of anyon motion), a requirement not satisfied by a system with a local Hamiltonian and bounded couplings. Furthermore, the dissipative dynamics of the field which controls the anyons may not be realizable by any Hamiltonian system in contact with a thermal bath. \section{Toric code coupled to a scalar field} \subsection{Toric code} To prepare for our discussion of models with long-range anyon interactions, let's first recall Wen's formulation~ of Kitaev's toric code model~. Consider an $L\times L$ square lattice with one qubit per site, where $L$ is an even integer. The Hamiltonian is \begin{align} H_{Toric} = -J \sum_{r} W_{r}, \quad W_{r}=X_{r,1}Z_{r,2}X_{r,3}Z_{r,4}, \end{align} where $r$ labels lattice plaquettes and $W_{r}$ acts on the four qubits contained in plaquette $r$ ($X$ and $Z$ denote Pauli matrices). On the torus ({\em i.e.} a square with periodic boundary conditions), the model has four degenerate ground states which can encode two qubits. The energy cost of creating an anyon pair is $4J$, and the anyons are noninteracting. A logical error occurs if a pair is created and one anyon propagates around a nontrivial cycle of the torus before the anyons reannhihilate. Therefore the (constant) energy barrier is $4J$. The typical separation between particles in the thermal anyon gas, and hence also the memory time, scale like the inverse Boltzmann factor $e^{2\beta J}$. \subsection{Coupling the toric code to massless bosons} Now, following Pedrocchi \emph{et al.}~, we consider coupling the two-dimensional toric code to an auxiliary three-dimensional bosonic system which induces long-range attractive forces among the plaquette variables. The Hamiltonian proposed by is \begin{equation} \begin{split} &H = -A \sum_{r}W_{r}(a_{r}+a^{\dagger}_{r}) + H_{boson},\\ &H_{boson} = \sum_{q}\epsilon_{q}a^{\dagger}_{q}a_{q},\\ &a_q = \frac{1}{\sqrt{N}}\sum_s a_s e^{iqs}.\\ \end{split} \end{equation} Here the index $r$ labels the sites of the two-dimensional topologically ordered medium and $s$ labels the $N$ sites of a three-dimensional lattice which contains the two-dimensional lattice; $a_r$ and $a_r^\dagger$ are the annihilation and creation operators for bosonic modes residing at the location of plaquette operator $W_r$, and $q$ labels the discrete Fourier modes for a bosonic field defined on the three-dimensional lattice. We assume $\epsilon_q > 0$, because a bosonic mode with negative $\epsilon_q$ would be unstable. The precise behavior of $\epsilon_q$ for large $q$ does not matter for analyzing the long-distance physics of the model, but the behavior of $\epsilon_q$ for small $q$ does matter. We assume a gapless spectrum for the Fourier modes with $\epsilon_{q}\sim q^2$ for small $q$. We will call the Hamiltonian in Eq.~() the ``toric-boson model.'' Coupling anyons to the three-dimensional massless boson field produces an attractive force between anyons, with potential energy inversely proportional to separation. After a polaron transformation (\emph{i.e.}, completing the square) the Hamiltonian may be rewritten as~ \begin{align} H = - \sum_{r,r'}J_{r,r'}W_{r}W_{r'} + \sum_{q}\epsilon_{q}\tilde{a}^{\dagger}_{q}\tilde{a}_{q}, \end{align} where \begin{align} J_{r,r'} = \frac{A^2}{N}\sum_{q} \frac{1}{\epsilon_{q}} e^{iq(r-r')} \sim \frac{A^2}{|r-r'|}; \end{align} here $N$ is the number of lattice sites in the three-dimensional bosonic system, \begin{equation} \tilde{a}_{q} = a_{q}-\frac{A}{\epsilon_{q}\sqrt{N}} \sum_{r} W_{r}e^{iqr},\end{equation} and the conclusion $J_{r,r,'} \sim {A^2}/{|r-r'|}$ applies for asymptotically large separation. Note that $\tilde{a}_{q}$ and $\tilde{a}_{q}^{\dagger}$ both obey ordinary bosonic commutation relations because the $W_{r}$'s residing at distinct plaquettes are commuting operators. The long-range behavior of the potential energy, scaling like $\sim 1/R$ where $R$ is the separation, arises from the pole in $1/\epsilon_{q}$, which dominates the Fourier integral for large $R$. Because the Hamiltonian in Eq.~() is the sum of an anyonic part $H_{A}$ and bosonic part $H_{B}$, which are completely decoupled, the energy eigenstates can be written as $|\psi\rangle = |\psi_{qubit}\rangle \otimes |\phi_{boson}\rangle$, where $|\psi_{qubit}\rangle$ is an eigenstate of $H_A$ and $|\phi_{boson}\rangle$ is an eigenstate of $H_B$. Since $J_{r,r'}>0$, the interactions between plaquette variables in Eq.~() are ``ferromagnetic,'' favoring alignment. $H_A$ has a $\mathbb{Z}_{2}$ global symmetry, under which $W_r$ changes sign, which is spontaneously broken at zero temperature --- there is one ground state sector with $W_{r}=1$ for all $r$ and another sector with $W_r = -1$ for all $r$, where both sectors have a topological degeneracy as well. The spontaneous breakdown of the global $\mathbb{Z}_{2}$ symmetry persists at sufficiently small nonzero temperature, and we may assume without loss of generality that the thermal expectation value $\langle W_r\rangle$ is positive . As emphasized in~, the crucial property of Eq.~() is that the energy cost of creating a single isolated anyon is infrared divergent, \emph{i.e.}, increases with the system size. To compute this energy cost, we follow~ and suppose that the bosonic bath is at any moment in a Gibbs state with respect to the effective Hamiltonian $\langle \alpha | H | \alpha \rangle $, where $|\alpha\rangle$ is the instantaneous anyonic state, \emph{i.e.}, an eigenstate of all operators $W_r$. In that case, this energy cost of creating a single isolated anyon only depends on the anyonic part $H_A$ of the Hamiltonian in Eq.~\eqref{eq:Toric_Boson} (see~ for the detailed calculation). The divergence arises because each plaquette interacts with many other distantly separated plaquettes on the two-dimensional lattice; flipping the sign of one $W_r$ relative to all the others increases the potential energy by an amount \begin{align} \mu \sim \sum_{r'}J_{r,r'} \sim O(L). \end{align} This effective chemical potential $\mu$ is $O(L)$ because we sum the $1/R$ interaction energy over the $L\times L$ lattice. In contrast to the standard toric code, where the energy of an isolated anyon is $2J$ and the thermal density of anyons at inverse temperature $\beta$ scales like $e^{2\beta J}$, in the model Eq.~() the thermal density scales like $e^{c \beta L}$ (where $c$ is a constant), and correspondingly the memory time grows exponentially with $L$ at any nonzero temperature. The infrared divergent chemical potential and the resulting self-correcting behavior of Eq.~() were the main conclusions of . In fact, the infinite chemical potential means that the spontaneous breakdown of the $\mathbb{Z}_{2}$ global symmetry, and hence the self-correcting behavior, hold at arbitrarily high temperature. \subsection{Derivation using effective field theory} Because the conclusion $\mu = O(L)$ follows from only the long-distance properties of the Hamiltonian Eq.~(), it can be recovered using an effective field theory description which captures the infrared behavior of the model. The effective Hamiltonian becomes \begin{equation} H_{EFT}=\tfrac{1}{2}\int d^{3}x\:\left(\nabla\phi\right)^{2}-A\int d^{2}y\: w(y)\phi(y)+ \cdots, \end{equation} where $\phi(x)$ is a suitably normalized scalar field. The $\tfrac{1}{2}\left(\nabla\phi\right)^{2}$ term in the Hamiltonian density arises from the dispersion relation $\epsilon_{q}\sim q^2$ for small $q$; higher-order terms in $q$ correspond to terms with more derivatives in $H_{EFT}$, which are not shown because they produce only small corrections for long-wavelength modes of the scalar field. The source $w(y)$ describes the anyon configuration. To be compatible with the conventions of Sec.~, we assume $w(y)$ takes the positive value $w_0>0$ in the ground state of the two-dimensional topological medium, and has the value $w(y) = w_0 - 2 n(y)$ when anyons are present, where $n(y)$ is a suitably smoothed and normalized anyon density. We assume $w(y) = 0$ outside the two-dimensional medium. The energy is minimized by a solution to the static classical field equation, which is Poisson's equation \begin{equation} \nabla^{2}\phi(x) =-Aw(x). \end{equation} The solution is \begin{equation} \phi(x)=A\int d^{2}y\left(\tfrac{1}{-\nabla^{2}}\right)_{x,y}w(y)+\phi_{0}(x) \end{equation} where $\phi_0(x)$ is a solution to the homogeneous field equation, and the Green function $\tfrac{1}{-\nabla^{2}}$ is determined by imposing appropriate boundary conditions. We assume that the two-dimensional anyonic source for the three-dimensional scalar field $\phi$ has bounded support, and that $\phi$ vanishes at spatial infinity. The boundary condition justifies an integration by parts with no surface term, and using the field equation we may obtain an alternative expression for the minimal energy of the field configuration in the presence of an anyon source: \begin{eqnarray} E[w] & = & -\int d^{3}x\,\tfrac{1}{2}\phi\nabla^{2}\phi-A\int d^{2}y\: w(y)\phi(y)\nonumber\\ & = & \tfrac{A}{2}\int d^{3}x\, w(x)\phi(x)-A\int d^{2}y\: w(y)\phi(y)\nonumber\\ & = & -\tfrac{A}{2}\int d^{2}x\: w(x)\phi(x) \end{eqnarray} with the integral now supported only on the two-dimensional surface where the source field $w$ is nonzero. Inserting Eq.~() into this expression, one finds \begin{equation} E[w]=-\tfrac{A^{2}}{2}\int d^{2}x\, d^{2}y\: w(x)\left(\tfrac{1}{-\nabla^{2}}\right)_{x,y}w(y). \end{equation} The boundary condition determines the Green function to be \begin{equation} \left(\tfrac{1}{-\nabla^{2}}\right)_{x,y}=\frac{1}{4\pi \left|x-y\right|}, \end{equation} and so we obtain \begin{equation} E[w]=-\tfrac{1}{2}\int d^{2}x\, d^{2}y\: w(x)J(x,y)w(y), \end{equation} where \begin{equation} J(x,y)= \frac{A^2}{4 \pi|x-y|}, \end{equation} as in Eq.~(). To compute the effective anyon chemical potential, we write $w(y)=w_0-2n(y)$, finding that $E[w]$ contains a term linear in the anyon density \begin{equation} \int d^{2}x\: \mu(x) n(x) \end{equation} where \begin{equation} \mu(x) = 2w_0\int d^{2}y\, J(x,y). \end{equation} Assuming for simplicity that $x$ is the center of a disk with radius $L/2$, we obtain \begin{eqnarray} \mu & = & 2w_0 A^{2}\int_{0}^{2\pi}d\theta\int_{0}^{L/2}rdr\tfrac{1}{4\pi r}=\tfrac{w_0A^{2}}{2}L, \end{eqnarray} recovering the linear divergence found in Eq.~(). \section{Perturbative instability of massless bosons} Having seen that the model in Eq.~() has an infrared divergent energy barrier, and hence an exponentially increasing quantum storage time at sufficiently small nonzero temperature, we next consider the stability of the energy barrier with respect to small perturbations of the system's local Hamiltonian. For assessing perturbative stability, it is important to note that the divergent anyon chemical potential arises because the response of the scalar field to the anyon vacuum is infrared divergent. A small perturbation of the Hamiltonian can tame this divergence, rendering the energy barrier finite. The bosonic ground state $|\phi_{boson}\rangle$ of Eq.~() satisfies \begin{equation} \tilde{a}_{q}|\phi_{boson}\rangle = 0 \end{equation} for all $q$, which implies \begin{align} \langle a_{r}\rangle_{0} \sim \frac{A}{N}\sum_{q}\frac{1}{\epsilon_{q}} \sum_{r'} e^{iq(r'-r)} =O(L) \end{align} where $\langle a_{r}\rangle_{0}$ denotes the ground-state expectation value, and we have substituted $W_{r'} = 1$. The bosonic mode $a_r$ has an unbounded occupation number in the anyon ground state, which is why the cost of creating an isolated anyon at site $r$ is divergent. Additional terms in the bosonic Hamiltonian, which are generically present, will prevent the occupation number from diverging. Suppose, for example, the Hamiltonian contains a term \begin{align} V = \epsilon \sum_{r} n_{r} = \epsilon \sum_q a_q^\dagger a_q, \end{align} where $n_r= a_r^\dagger a_r$ is the number operator for the mode localized at site $r$. We may assume $\epsilon > 0$ because otherwise the bosonic field would be unstable, an indication that we are expanding around the wrong bosonic vacuum. Now the perturbation $V$, which disfavors a large occupation number, competes with the Hamiltonian Eq.~(), and the ground-state expectation value of $a_r$ takes the finite value \begin{align} \langle a_{r}\rangle_{0} \sim \frac{A}{N}\sum_{q}\frac{1}{\epsilon+\epsilon_{q}} \sum_{r'} e^{iq(r'-r)} = O(\epsilon^{-1/2}), \end{align} because the perturbation removes the pole at $q=0$. For this particular perturbation the bosonic Hamiltonian is quadratic and therefore exactly solvable, but the conclusion that a generic perturbation removes the infrared divergence holds more generally. In the effective field theory language, the solution to the static field equation Eq.() becomes \begin{eqnarray} &\phi(x)=A\int d^{2}y\left(\tfrac{1}{-\nabla^{2}}\right)_{x,y}w(y)\nonumber\\ &= Aw_0\int d^{2}y \frac{1}{4\pi |x-y|} = O(L) \end{eqnarray} in the anyonic vacuum such that $w(y) = w_0$ in the two-dimensional topological medium and $w(y) = 0$ elsewhere. The infrared divergent behavior of $\phi(x)$ arises because the scalar field is exactly massless in Eq.~(). Generically a nonzero mass term will be present, perturbing the Hamiltonian according to \begin{align} H_{EFT}\to H_{EFT} + \int d^{3}x\: \tfrac{1}{2} m^2\phi^2. \end{align} We may assume $m^2 > 0$, because otherwise the scalar field would be unstable, indicating that we are expanding around the wrong bosonic vacuum. The static vacuum solution becomes \begin{eqnarray} \phi(x)=Aw_0\int d^{2}y\left(\tfrac{1}{-\nabla^{2}+m^2}\right)_{x,y}=O(m^{-1}); \end{eqnarray} now the $y$ integral converges because the Green function decays like $e^{-m|x-y|}$ for large separation. The effective field theory approach is useful for analyzing the energy barrier of a quantum memory because whether the energy barrier diverges or not hinges on the behavior of the model in the far infrared, where the field theory approximation should be reliable. The degrees of freedom described by the field theory may be \emph{emergent}, not corresponding in any very direct with the microscopic degrees of freedom of the underlying Hamiltonian model. But just as the divergent energy barrier in the model Eq.~() is realized only when the free parameter $m^2$ of the effective field theory is tuned to zero, we may anticipate that a divergent energy barrier arises from the exchange of massless bosons only on (at best) a codimension-one surface in the space of all Hamiltonian models. A generic perturbation moves the Hamiltonian off this surface, destroying the infinite energy barrier. To evade this conclusion, we need models in which bosons with the appropriate couplings remain gapless when the microscopic Hamiltonian is deformed. Not just the infinite energy barrier but also the topological order of the two-dimensional medium might be threatened by small perturbations of the Hamiltonian. In addition to coupling to plaquette variables, the scalar field could couple to individual qubits; for example the perturbed Hamiltonian might contain terms such as \begin{equation} V = \epsilon' \sum_r (a_r + a_r^\dagger)(X_{r,1}+X_{r,2}+X_{r,3}+X_{r,4}), \end{equation} If in the ground state the expectation value of the field scales like $\langle a_r \rangle_0\sim m^{-1} \sim \epsilon^{-1/2}$ as in Eq.~() and (), then an effective magnetic field \begin{equation} B\sim \frac{\epsilon'}{\sqrt{\epsilon}} \end{equation} is applied to the qubits. Although enhanced by the large strength of the field, this magnetic field is still perturbatively small compared to the energy cost $\sim 1/\sqrt{\epsilon}$ of creating an isolated anyon, so it is plausible that the topological order would survive. In any case, we expect the topological order to persist within a region of finite volume in the space of Hamiltonians. \section{Stabilizing massless bosons} The scheme of Pedrocchi \emph{et al.}~ for stabilizing a quantum memory requires massless bosonic fields, and we have argued that under generic conditions bosonic fields acquire mass unless the Hamiltonian is tuned to a codimension-one surface. But there are important exceptions to this rule, and we should consider whether these exceptions can be exploited by quantum engineers. In particular, the masslessness of Goldstone bosons and gauge bosons follows from general symmetry principles. \subsection{Goldstone bosons} A spontaneously broken exact continuous symmetry is always associated with an exactly massless Goldstone boson, whose masslessness is preserved by any perturbation that respects the symmetry (and also maintains its spontaneous breakdown). If $\phi(x)$ is a Goldstone field, then the symmetry acts according to \begin{equation} \phi(x) \mapsto \phi(x) + \mbox{constant}, \end{equation} disallowing a mass term $\tfrac{1}{2}m^2\phi^2$ in the effective Hamiltonian. The symmetry not only forbids a mass term for the Goldstone boson, but also constrains its coupling to other fields. For example, a term $-A\phi(x)w(x)$ describing the coupling of the Goldstone field to the anyonic density $w(x)$ is not allowed, but a derivative coupling $-A\nabla \phi(x) w(x)$ is compatible with the symmetry, and can be reexpressed as $A\phi(x)\nabla w(x)$ after integrating by parts. (This term violates rotational invariance, but could be allowed in an anisotropic system.) Then, arguing as in Sec.~, we find in place of Eq.~() that the energy of the bosonic ground state in the presence of a source becomes (in $D$ spatial dimensions) \begin{eqnarray} &E[w]=-\tfrac{A^{2}}{2}\int d^{D}x\, d^{D}y\: \nabla w(x)\left(\tfrac{1}{-\nabla^{2}}\right)_{x,y}\nabla w(y)\nonumber\\ &= -\tfrac{A^{2}}{2}\int d^{D}x\, w(x)^2, \end{eqnarray} the infrared-finite integral of a local source-dependent energy density. Because they are derivatively coupled, the exchange of Goldstone bosons generates short-range contact interactions between anyons rather than a long-range forces. Perturbations that explicitly break the continuous symmetry would gap out the Goldstone bosons in any case. But even if we are willing to impose a symmetry that protects the massless Goldstone bosons, there would be no resulting divergent energy barrier to suppress logical errors. \subsection{Gauge bosons} In some cases, exactly massless deconfined $U(1)$ gauge bosons can emerge from an underlying spin model with no fundamental gauge symmetry. From an effective field theory viewpoint, since the gauge symmetry is not an exact feature of the underlying microscopic Hamiltonian, nothing seems to forbid adding a gauge non-invariant term to the Hamiltonian of the long-distance effective field theory, which would gap out the gauge bosons. But if the microscopic local Hamiltonian energetically favors a sector of the theory which satisfies gauge constraints, these constraints could be robust against generic small perturbations of the microscopic Hamiltonian~. A model with such features can support a stable ``Coulomb phase,'' where a ``photon'' remains exactly massless without any need to impose symmetries or carefully tune the Hamiltonian. Emergent photons can mediate long-range forces between charged particles, which are attractive for particles of opposite charge, and repulsive for particles of like charge. Chesi \emph{et al.} suggested that long-range repulsive interactions among anyons in a two-dimensional medium could produce a divergent free energy barrier, because for a nonzero anyon density the cost of introducing an additional isolated anyon would diverge with system size~; thus the anyon density should approach zero in the thermodynamic limit. However, anyon models in which gauge-mediated interactions are repulsive for each anyon pair do not seem to be self consistent. Because the photon couples to a locally conserved charge, anyon pairs which can be locally created from the vacuum must have opposite ``electric charges.'' At nonzero temperature, anyons coupled to emergent photons would form a plasma with charge screening and no divergent energy barrier It may be possible in principle for anyons to carry dipole moments which couple to emergent photons. In that case, if the gauge field is three-dimensional there would be a long-range interaction with potential energy falling of with distance $R$ like $1/R^3$. Even if in a suitably engineered Hamiltonian the 3D gauge field couples to the plaquette variables of the 2D toric code, these interactions decay too rapidly to produce an infrared divergent energy barrier. For the 3D toric code, a logarithmically divergent chemical potential $\mu \sim \log L$ due to dipolar interactions is conceivable. It's not clear, however, how to arrange the needed coupling between the topological medium and the gauge theory; see Sec.~ for further discussion. \subsection{Other long-range forces} Massless scalars can also be stabilized by imposing both supersymmetry and chiral symmetry -- the chiral symmetry enforces the masslessness of a fermion species, and the supersymmetry requires the scalar and fermion to be degenerate. Massless spin-two bosons (gravitons) can be stabilized by general covariance. If we could find spin models in which supersymmetry and/or general covariance are emergent and robust with respect to generic local perturbations of the microscopic Hamiltonian, that would be very interesting from the perspective of fundamental physics, apart from the potential applications to self-correcting quantum memory! Putting aside the microscopic origin and perturbative stability of the long-range interactions, we may consider the consequences of long-range forces between ``anyons'' in $D$ dimensions described by Eq.~(), where \begin{align} J_{r,r'} \sim \frac{1}{|r-r'|^{\alpha}}. \end{align} For $0<\alpha<D$, the cost of creating an isolated anyonic excitation diverges with the system size, so the model has a quantum memory time which increases rapidly with system size, even at arbitrarily high temperature. The precise scaling of the lifetime of the memory with system size is discussed in~. For $\alpha >D$, energy cost remains finite, and the memory time is bounded above by a constant at any nonzero temperature. The case $\alpha=D$ is marginal, with a logarithmically divergent energy barrier. By similar reasoning, the one-dimensional Ising model with $0<\alpha<1$ is a good classical memory at arbitrarily high temperature. It is also known that this system has a finite-temperature phase transition for $1<\alpha < 2$, and is magnetically disordered at any nonzero temperature for $\alpha > 2$~. \section{Discussion} The main thrust of this paper is contained in Sec.~-, where we have described schemes for stabilizing a two-dimensional topologically ordered quantum memory at nonzero temperature using long-range interactions, and have argued that such schemes have a significant and generic drawback --- the Hamiltonian needs to be precisely tuned for the memory time to scale favorably with the system size. Our arguments are incomplete, and leave many open questions, including the important and enticing question whether self-correcting quantum memory is achievable in a suitably engineered three-dimensional system. This concluding section contains a few questions and remarks which may help to guide future research on self-correcting quantum memory. \subsection{Quasi-topological phases} In mathematical models of topologically ordered systems, we usually assume the system has an energy gap separating the topologically degenerate ground state from the rest of the energy spectrum. Then we can imagine integrating out all quasiparticle excitations to obtain an effective theory with no propagating excitations, a topological quantum field theory (TQFT) which accurately describes the physics of the system up to corrections which are exponentially small in the system size. But strictly speaking this picture does not necessarily apply to the topologically ordered systems observed in the laboratory, which may support gapless photon or phonon excitations. Furthermore, disordered systems may have many low-lying localized bulk quasiparticle excitations with energy well below the topological gap. Bonderson and Nayak proposed the term \emph{quasi-topological phase} for a system that has some of the features of a topological phase, but also has gapless excitations and therefore cannot be precisely described at low energy by a TQFT ~. For example, Laughlin states are quasi-topological; their fractionally charged quasiparticles couple to the massless photon, and correspondingly the topological degeneracy on the torus is split by an amount that scales polynomially rather than exponentially with the system size. Furthermore, the anyon braiding statistics is modified by nonuniversal dynamical phases which are not described by the TQFT. Bonderson and Nayak also proposed the term \emph{strong quasi-topological phase} for a system in which some sector of the theory decouples from the gapless excitations and can be accurately described by a TQFT. For example, the Moore-Read state has a strongly quasi-topological sector. Although the Ising anyons carry electric charges which produce non-topological exchange phases, the distinguishable fusion channels for a system of many Ising anyons span a topologically protected Hilbert space well described by the Ising TQFT. The toric-boson model, the toric code coupled to an exactly massless scalar field as in Eq.~(), is a strong quasi-topological phase. One might have thought that, just as distantly separated anyons have a potential energy scaling algebraically with their separation, so large punctures on the lattice, which we may think of as fattened anyons, should also have an algebraically decaying interaction, and corresponding the splitting of the degeneracy on the torus should be algebraically decaying as for Laughlin states. But in the exactly solvable toric-boson model, the topological degeneracy on the torus is exact. Deforming the model with a small local perturbation might conceivably destroy the topological order as discussed in Sec.~, but in the ``weak magnetic field'' regime of Eq.~() and () the perturbation may produce only an exponentially small splitting of the ground state degeneracy, just as for the toric code without long-range interactions. \subsection{Topological order at low temperature in a finite system} Our focus throughout this paper has been on the scaling of the memory time with system size, and in particular on whether arbitrarily long memory times are achievable as a matter of principle. But we should emphasize that a two-dimensional topologically ordered system of fixed size could be a useful quantum memory if the temperature $T$ is far below the topological gap $\Delta$. The signatures of topological order, such as quantized Hall conductance and nontrivial topological entanglement entropy, disappear on an $L\times L$ lattice for system size larger than $L_0(T)$, where~ \begin{equation} \log L_0(T) \approx \Delta /T. \end{equation} For $L > L_0(T)$ the system is likely to contain propagating thermally excited anyons, causing rapid decay of stored quantum information. But just as the quantized Hall conductance is a robust physical phenomenon, even though it disappears in the thermodynamic limit at any nonzero temperature, so a topological quantum memory with storage time enhanced by the inverse Boltzmann factor $e^{\Delta/T}$ may be a valuable resource, even though it fails to meet our criteria for self correction. \subsection{Self correction and fast propagation} The Lieb-Robinson bound, which applies to any local Hamiltonian which is a sum of finite-norm terms, asserts that information propagates at finite speed. This bound is violated by the toric-boson model Eq.~(); therefore this model cannot arise from an underlying model whose Hamiltonian is a sum of terms, each with bounded norm and bounded range. When the toric-boson model on a $L\times L$ lattice is perturbed by a suitable local Hamiltonian, anyons can propagate at speed $O(L)$. Consider dividing the lattice into two domains separated by a horizontal domain wall, as in Fig.~. The lower half $A$ is in the phase where plaquettes take the value $W_{r}=+1$, and the upper half $B$ is in the phase with $W_r= -1$. At time $t=0$, two anyonic excitations (each with $W_r=+1$) are created deep within the $B$ phase, and at the same time magnetic fields are applied to spins on a vertical line which extends from the lower anyon to the domain wall, and beyond into the $A$ phase. This anyon is subjected to a long-range force attracting it toward the $A$ phase. Were no magnetic field applied, the configuration with an anyon pair would be an exact energy eigenstate and would not evolve. But the applied magnetic field gives the lower anyon a finite effective mass, so it can respond to the attractive force by falling vertically. By the time it reaches the domain wall, this falling anyon has accelerated to speed $O(L)$. Thus anyons can travel at unbounded speed in the limit of infinite system size, showing that the Lieb-Robinson bound does not apply. When the anyon reaches the domain wall, it plunges through, becoming an anyon with $W_r=-1$ in the $A$ phase. Forces attracting it back toward the $B$ phase slow the anyon down until it comes to rest deep with the $A$ phase, and then proceeds to oscillate between positions in the $A$ and $B$ phases. Since the divergent anyon chemical potential of the toric-boson model is directly related to the long-range forces which can accelerate anyons to unbounded speed, one wonders whether this connection holds more generally. Generalized Lieb-Robinson bounds have been derived for systems which have long-range interactions scaling with separation $r$ as $r^{-\alpha}$, showing that in $D$ dimensions the time required for information to propagate a distance $r$ is bounded below by a logarithm of $r$ for $ \alpha > D$~ and by a sublinear power of $r$ for $\alpha > 2D$~. No limitations on propagation speed have been derived for $ \alpha \le D$, the regime in which long-range interactions can cause the chemical potential to diverge. Since the infinite anyon chemical potential results from the response of the topological medium arbitrarily far away, it's tempting to assert that a divergent chemical potential can occur only in a system with constant propagation time independent of $r$, though we do not have a rigorous general argument for that conclusion. Not only bosons, but also fermions can induce long-range interactions when fermion wave functions are delocalized. A familiar example is the RKKY interaction between nuclear magnetic dipole moments induced by exchange of conduction electrons, leading to an effective interaction strength proportional to \begin{equation} \sim \frac{1}{r^4} [2kr\cos(2kr) - \sin(2kr)] \end{equation} where $r$ is the separation of the dipoles and $k$ is the conduction electron's wave number. Even though such effective long-range forces can arise from fermion exchange, these forces cannot produce an infinite propagation speed or an infinite anyon chemical potential. The microscopic Hamiltonian is a sum of terms with bounded range and bounded norm, so the Lieb-Robinson bound implies a finite propagation speed. Furthermore, the energy cost of creating an isolated anyon is finite, because only a constant number of terms in the Hamiltonian, each with constant norm, are sensitive to the presence of the excitation. Indeed, this remark applies not just to fermions but also to spin models where the Hamiltonian is a sum of terms with bounded range and bounded norm. \subsection{Self correction as a phase transition} We expect the stored quantum information in a quantum memory to decay quickly at sufficiently high temperature. Therefore, a system which is self correcting is expected to undergo a phase transition at a critical temperature. In the low-temperature phase the memory time increases without bound as the system size increases, and in the high-temperature phase the memory time is bounded above by a constant independent of system size. If the system has no phase transition at nonzero temperature, then we do not expect it to be self correcting at low temperature. For example, though Haah's cubic code~ and Michnicki's welded code~ both have divergent energy barriers, in both cases the partition function can be computed exactly, and one finds that singularities occur only at zero temperature. The absence of a phase transition at nonzero temperature is compatible with the observation that the memory time is constant at nonzero temperature, despite the increasing energy barrier. In contrast, the four-dimensional toric code, which is self correcting, has a phase transition at nonzero temperature associated with condensation of string excitations . In fact the connection between self correction and phase transitions can be rather subtle, as Hastings \emph{et al.} observed~. They noticed that in the toric code in six or more dimensions, the critical temperature $T_c$ associated with thermodynamic singularities is higher than the percolation temperature $T_p$, above which thermally excited extended defects can have infinite size with nonnegligible probability; furthermore, the separation between the two temperatures grows parametrically with increasing dimension $D$. A related phenomenon occurs if we consider the toric code in $D=4$ dimensions, but where the variables residing on lattice plaquettes take values in $Z_N$ rather than $Z_2$. As has been known for some time, for $N\ge 5$ the four-dimensional $Z_N$ gauge theory has two phase transitions as the temperature varies . In the low-temperature phase, defects are typically small and dilute, and the memory time increases exponentially with system size. In the high-temperature phase, stored information decays in constant time. To gain insight into the memory time in the intermediate-temperature region, it is helpful to consider classical spin systems storing classical information which also exhibit defect percolation below the critical temperature. For example, for the $N$-state Potts model~ with $N\ge 5$, the defect percolation temperature $T_p$, above which thermally fluctuating domain walls have unbounded size, is less than the critical temperature $T_c$, above which the system does not spontaneously magnetize. In the intermediate-temperature phase, although defects percolate, the free-energy barrier for logical errors is $O(\log L)$, and correspondingly the memory time scales polynomially with system size~. This physical picture suggests that the memory time may also scale polynomially for high-dimensional toric codes, when the temperature is in between the percolation temperature and the critical temperature. Thus, while $T < T_c$ may be necessary for an exponentially scaling memory time, it does not seem to be sufficient in general. Hastings \emph{et al.} also suggested that high-dimensional toric codes may be self correcting even in the superheated regime above the critical temperature, due to hysteretic behavior ~. Mean-field theory predicts that the Potts model is a good classical memory even for $T > T_c$~, but this mean-field prediction is not supported by numerical simulations. The memory time of the (slightly) superheated two-dimensional Ising model increases polynomially with system size $L$ up to an optimal size $L^*$ comparable to the correlation length, but the $N$-state Potts model has a first-order thermal phase transition for $N\ge 5$, and its stored classical information decays rapidly when the temperature is just slightly above the critical temperature. Another connection between self correction and phase transitions can be obtained from the analysis of master equations describing thermally fluctuating systems. For any classical spin system subject to Glauber dynamics, in which \emph{all} correlations functions decay exponentially with distance, if the Gibbs state is the unique fixed point then the convergence to the Gibbs state is rapid~. Here ``rapid'' means scaling polynomially with system size; for some classical spin systems stronger results can be derived~, and in any case the classical memory time may in some cases be much shorter than the convergence time to the Gibbs state. Since exponentially decaying correlations are a typical feature of high-temperature phases, we may expect rapid mixing to the Gibbs state above the critical temperature, and therefore rapid decay of stored classical information. \subsection{Perturbative origin of long-range forces} In Sec.~ we discussed how long-range interactions between plaquette operators might enhance the quantum memory time of a two-dimensional topologically ordered system. But until now we have not considered in detail how such long-range forces might arise from a microscopic Hamiltonian which is a sum of terms with bounded range and bounded norm. Before we discuss the microscopic origin of long-range interactions, let's first recall how topological order can arise in a two-dimensional spin system. The toric code provides a beautiful and instructive example of a frustration-free commuting Hamiltonian with a topologically ordered ground state. But each term in the Hamiltonian acts nontrivially on four qubits, while the naturally occurring interactions in a spin Hamiltonian are typically two-body terms. In fact, frustration-free commuting models realizing any doubled (non-chiral) anyon model can be constructed , but only by including even higher-weight many-body terms in the Hamiltonian. What spin models with physically plausible two-local interactions will exhibit topological order? It is known that the ground state of any two-body frustration-free commuting Hamiltonian has only short-range entanglement, and therefore cannot be topologically ordered~. This conclusion also applies to frustration-free commuting Hamiltonians with three-local interactions among qubits on any graph, or for three-local interactions among qutrits on a nearly Euclidean lattice~. In this sense the toric code, with four-local interactions among qubits on a square lattice, is the optimal model of two-dimensional topological order based on an exactly solvable commuting Hamiltonian. Therefore, to obtain a topologically ordered ground state from a two-local Hamiltonian, we must be willing to consider noncommuting Hamiltonians, which may not be exactly solvable. One approach to constructing such models is the method of perturbative gadgets~. The idea is to approximate a target Hamiltonian $H_0$, which is commuting and has a topologically ordered ground state, with a two-local Hamiltonian $H$. We may express $H$ as \begin{equation} H = H_{0} + V; \end{equation} because the topological order of $H_0$ is perturbatively stable, the quantum memory properties of the two-local Hamiltonian $H$ will mimic those of the target Hamiltonian $H_0$ if the perturbation $V$ is a sum of sufficiently small local terms. But can we also obtain long-range plaquette-plaquette interactions of the form $-\sum_{r,r'} J_{r,r'}W_{r}W_{r'}$ as a perturbative approximation to a two-body Hamiltonian? On a square lattice, each term in this Hamiltonian involves eight lattice sites, and the number of terms scales like the square of the system's area, where each term involves the exchange of at least one gadget particle, so the ``density'' of gadget particles per unit area is divergent. We have not found a way for the gadget particles to couple to the plaquette variables in the desired way, or for the nonlocal Hamiltonian to be perturbatively close to a two-local one. In fact, as already noted in Sec.~, we do not expect the infinite anyon chemical potential of the toric-boson model to be realizable by any Hamiltonian which is a sum of terms with bounded range and bounded norm, simply because the energy cost of a localized excitation should be finite if the excitation occupies a finite lattice volume. Or at any rate, the toric-boson model could provide a good description only in a scaling limit where the lattice spacing becomes small compared to the anyon size. A further potential problem is that long-range interactions, whatever their origin, might interfere with the perturbative stability of the target Hamiltonian, and hence invalidate the perturbative gadget method for achieving topological order. Arguments for the perturbative stability of topological order formulated in~ do not apply if the perturbation $V$ includes terms which decay algebraically (rather than exponentially) with distance. Indeed, systems with long-range interactions can often be well described using mean-field approximations in which entanglement is assumed to be short range. If a topologically ordered system with long-range plaquette-plaquette interactions can arise from a two-local Hamiltonian at all, approximation schemes going beyond conventional perturbation theory may be needed to analyze the stability of the system's stored quantum information~. \acknowledgments We thank Jeongwan Haah, Michael J. Kastoryano, Daniel Loss, Kamil Michnicki, Fernando Pastawski, Fabio Pedrocchi, and Kristan Temme for helpful discussions. BY is supported by the David and Ellen Lee Postdoctoral fellowship. OLC is partially supported by Fonds de Recherche Qu\'ebec-Nature et Technologies. DP is partially supported by Canada's NSERC and the Canadian Institute for Advanced Research. This work was supported in part by NSA/ARO grant W911NF-09-1-0442, and AFOSR/DARPA grant FA8750-12-2-0308. We also acknowledge funding provided by the Institute for Quantum Information and Matter, an NSF Physics Frontiers Center with support of the Gordon and Betty Moore Foundation (NSF Grants No. PHY-0803371 and PHY-1125565). Part of this work was done while DP was visiting IQIM. \bibliography{myref_nyaan_2015} \appendix \section{Equivalence of lattice model to 3D massless bosons with 2D classical source} The goal of this section is to prove the equivalence between the lattice model \begin{eqnarray} H & = & H_{b}+A\sum_{p}W_{p}\otimes\left(a_{p}+a_{p}^{\dagger}\right)\\ H_{b} & = & \varepsilon_{0}\sum_{i}a_{i}^{\dagger}a_{i}-t\sum_{\langle i,j\rangle}a_{i}^{\dagger}a_{j} \end{eqnarray} and the Hamiltonian of 3D free massless bosons, linearly coupled to a 2D source \begin{equation} H=\tfrac{t}{2}\int d^{3}x\:\tfrac{1}{2}\left(\nabla\phi\right)^{2}+A\int d^{2}y\: w(y)\phi(y) \end{equation} This equivalence will highlight that the fine-tuning condition \begin{equation} \varepsilon_{0}=6t \end{equation} sets the mass of bosons to zero, thus allowing long-range interactions. The simple idea is to promote the lattice operator $a_{i}+a_{i}^{\dagger}$ to a scalar field $\phi$. The linear term $w\times\phi$ is straightforwardly equivalent to the term $\sum_{p}W_{p}\otimes\left(a_{p}+a_{p}^{\dagger}\right)$. We now argue that $H_{b}$ gives rise to the gradient term $\frac{1}{2}\left(\nabla\phi\right)^{2}$. By using finite differences and denoting the nearest-neigbors of site $i$ as $\mathcal{N}(i)$, one gets \begin{eqnarray} \tfrac{1}{2}\left(\nabla\phi\right)^{2} & \leftrightarrow & \tfrac{1}{2}\sum_{i}\sum_{j\in\mathcal{N}(i)}\left( (a_{j}+a_{j}^{\dagger})-(a_{i}+a_{i}^{\dagger})\right)^{2}\\ & = & 1+\tfrac{1}{2}\sum_{i}\sum_{j\in\mathcal{N}(i)}a_{j}^{2}+a_{j}^{\dagger\,2}+a_{i}^{2}+a_{i}^{\dagger\,2}\\ & & +\sum_{i}\sum_{j\in\mathcal{N}(i)}a_{j}^{\dagger}a_{j}+a_{i}^{\dagger}a_{i}-a_{j}^{\dagger}a_{i}-a_{i}^{\dagger}a_{j} \end{eqnarray} The terms in line () are insignificant since they do not conserve the number of bosons. The significant terms in line () can be rearranged by counting the number of diagonal terms $a_{i}^{\dagger}a_{i}$ and crossed terms $a_{i}^{\dagger}a_{j}$. Since there are 6 nearest neighbors in 3D, one gets \begin{eqnarray} \eqref{eq:significant_terms} & = & 12\sum_{i}a_{i}^{\dagger}a_{i}-2\sum_{\langle i,j\rangle}a_{i}^{\dagger}a_{j} \end{eqnarray} The ratio between the coefficients of the hopping terms and the on-site potentials is precisely the same as in the fine-tuning condition, Eq. (). Thus, imposing this ratio ensures that only a kinetic term $\propto\left(\nabla\phi\right)^{2}$ appears. Otherwise, a quadratic term $\propto\phi^{2}$ would appear and give a mass to bosons. In other words, the fine-tuning condition, Eq. (), is crucial to induce long-range (rather than short-range) interactions between anyons. \section{Entropy effects} \subsection{Localization introduces a bumpy energy landscape} Randomizing coupling strengths in a quantum memory Hamiltonian can localize propagations of anyonic excitations and stabilize quantum memory against quantum decoherence at $T=0$~. For concreteness, consider the Toric code. We assume that the actual Hamiltonian is given by \begin{align} H = H_{toric} + V \end{align} with some imperfection $V$ which may be viewed as a local perturbation. If one knows the details of the imperfection $V$, qubits can be encoded into the four lowest energy states of the actual Hamiltonian. For sufficiently small $V$, the energy splitting between four lowest energy states is exponentially suppressed, and the encoded qubits are essentially free from dephasing. However, since one has no knowledge on the imperfection $V$, one needs to encode logical qubits into ground states $|\psi_{toric}\rangle$ of $H_{toric}$ instead of the actual Hamiltonian $H$. Since the state $|\psi_{toric}\rangle$ is a superposition of excited states of $H$, it quickly dephases and encoded qubits will be lost in $O(1)$ time~. (If one knows $V$, one may reverse the time evolution to recover the original state). Randomizing the coupling strengths of the Toric code resolves the above problem of quantum decoherence~ \begin{align} H^{*}_{toric}= -\sum_{r}J_{r}W_{r} \end{align} where $J_{r}>0$ are some random constants. The key idea is that the energy landscape of anyonic excitations becomes ``bumpy'' and excited states become localized even in the presence of small perturbation $V$. In such a localized regime, quantum memory time at $T=0$ becomes exponentially long~. The capability of localizing anyonic excitations is also crucial for topological quantum computation since anyons must be coherently controlled. \subsection{Enhancement due to bumpy energy landscape} We ask whether entropy effects may stabilize quantum memory at $T>0$. Brown and collaborators have argued that bumpy energy landscape leads to significant enhancement of quantum memory time by studying $\mathbb{Z}_{5}$ Toric code with defect lines~. Their model does not exhibit a diverging energy barrier. Yet, numerical evidence point to a partial self-correcting behaviour similar to the Cubic code~, i.e., a memory time $\tau\sim e^{c\beta^2}$ by optimizing the lattice size at a given temperature. Their model highlights that although there exists a sequence of local transformations connecting ground states which does go through a diverging energy barrier, this sequence can be rather improbable. More generally, the energy barrier criterion does not capture the entropy effects which cannot be neglected at non-zero temperature. |
1501.04116 | Title: A Simple Model for Long-Range Interacting Pendula
Abstract: We show that the Hamiltonian mean field (HMF) model describes the equilibrium
behavior of a system of long pendula with flat bobs that are coupled through
long-range interactions (charged or self gravitating). We solve for the
canonical partition function in the coordinate frame of the pendula angles. The
Hamiltonian in the angles coordinate frame looks similar to the form of the HMF
model but with the inclusion of an index dependent phase in the interaction
term. We also show interesting non-equilibrium behavior of the pendula angles,
namely that a quasistationary clustered state can exist when pendula angles are
initially ordered by their index.
Body: \title{A Simple Model for Long-Range Interacting Pendula} \author{Owen Myers} \affiliation{Materials Science Program, University of Vermont, Burlington, Vermont 05405, USA.} \affiliation{Department of Physics, University of Vermont, Burlington, Vermont 05405, USA.} \email{oweenm@gmail.com} \author{Adrian Del Maestro} \affiliation{Materials Science Program, University of Vermont, Burlington, Vermont 05405, USA.} \affiliation{Department of Physics, University of Vermont, Burlington, Vermont 05405, USA.} \author{Junru Wu} \affiliation{Materials Science Program, University of Vermont, Burlington, Vermont 05405, USA.} \affiliation{Department of Physics, University of Vermont, Burlington, Vermont 05405, USA.} \author{Jeffrey S. Marshall} \affiliation{School of Engineering, University of Vermont, Burlington, Vermont 05405, USA.} \date{\today} \begin{abstract} We show that the Hamiltonian mean field (HMF) model describes the equilibrium behavior of a system of long pendula with flat bobs that are coupled through long-range interactions (charged or self gravitating). We solve for the canonical partition function in the coordinate frame of the pendula angles. The Hamiltonian in the angles coordinate frame looks similar to the form of the HMF model but with the inclusion of an index dependent phase in the interaction term. We also show interesting non-equilibrium behavior of the pendula angles, namely that a quasistationary clustered state can exist when pendula angles are initially ordered by their index. \end{abstract} \pacs{} \maketitle \section{Introduction} Systems with long-range interactions are a source of unique problems in the field of statistical mechanics and thermodynamics. This is due to several properties of long-range systems which fall outside of the conditions normally needing to be satisfied when applying the methodologies of thermodynamics. Simply from the words ``long-range'' the first infringement can be deduced, that long-range systems are not additive. If two systems with short-range interactions are brought together to form a larger system then the energy difference between the conglomerate system and the sum of its constituents is the new potential energy from the boundary between them. In the thermodynamic limit, the potential energy of the boundary is small compared to the bulk and can be neglected, making short-range systems additive. In the case of long-range interactions, one particle will feel a significant potential created by every other particle, so the additional potential energy of two systems added together does not scale as the boundary but in a more complicated way that depends on the specific nature of the interactions . Directly related to the lack of additivity is the fact that systems with long-range interactions are not extensive because their energy diverges in the thermodynamic limit . Although these characteristics compel cautious use of the usual tools of statistical mechanics, they are also the source of many interesting dynamical and statistical features. Depending on the system of interest, such features include canonical and microcanonical ensemble inequivalence and related negative specific heat , quasistationary states (different than metastable states which lie on local extrema of equilibrium potentials) whose lifetimes increase with the number of particles , an interesting dependence of the largest Lyapunov exponent on particle number in a long-range Fermi-Pasta-Ulam model , and spontaneous creation of macroscopic structures in non-equilibrium states . In some cases, long-range interactions can greatly simplify problems. For instance, mean field models depend on one of two premises: (i) interactions are short-range but the system is embedded in a space of infinite dimension so that all bodies in the system are nearest neighbors, or (ii) interactions are infinitely long. For some time, the primary motivation for the study of long-range interactions was to understand galaxies, galaxy clusters and the general thermodynamic properties of self-gravitating systems. Aside from mean field models, interest has further built since the observation of modified scattering lengths in Bose-Einstein Condensates (BEC) through the use of Feshbach resonances . Using this technique, a BEC can be made to be almost non-interacting by tuning the scattering length to zero. One could even tune the scattering length to a negative value, making the BEC collapse. More recently, O'dell et al. has shown that it may be possible to produce an attractive $1/r$ potential between atoms in a BEC by applying an ``extremely off resonant'' electromagnetic field. This has opened the possibility of creating table-top methods which physically model aspects of cosmological behavior on a laboratory scale, as well as the possible development of entirely new dynamics in BEC. The challenges in understanding long-range systems drive the development of solvable models that could help better explain some of the aforementioned phenomena. Campa et al. have recently published a collection of important solvable models. One particularly significant model, which is important to this work, is the Hamiltonian Mean Field (HMF) $XY$ spin model , often written in the form \begin{equation} H = \sum_{i=1}^{N} \frac{p_i^2}{2} + \frac{\gamma}{2N}\sum_{i,j=1}^{N} \left[ 1- \cos{(\theta_i-\theta_j)} \right] , \end{equation} \noindent where $\theta_i$ is the angular position of the $i^{\mathrm{th}}$ particle (spin), as shown in Fig.~, and $p_i$ is its conjugate angular momentum. The HMF model is generally used to describe two different classes of systems: 1) a mean field $XY$ classical spin model, and 2) a one dimensional periodic system of itinerant particles with long-range interactions. Though the connection between the HMF model and the second class of systems mentioned could be thought of as contrived given the simplifications under which the model is realized, it has been shown that the model produces useful insights into how non-neutral plasmas and self gravitating systems behave . In this paper, we study the dynamics of an array of $N$ pendula with long-range interacting bobs. By considering long pendula with flat bobs undergoing small oscillations and having parallel planes of rotation, we produce a model related to the HMF model through a coordinate transformation. The transformation introduces a dependence on the indices of the particle labels. A cartoon of the physical picture is shown in Fig.~. The index dependence in the Hamiltonian, that will be described in detail in the next section, is a consequence of the pendula pivots being slightly offset from one another and appears as a phase in the cosine term of the HMF model. It inspires the investigation of non-equilibrium ``repulsive'' behavior in the angle coordinate frame where we find an interesting quasistationary state when the angles of the pendula are initially ordered according to their indices. We find the clustered positions in the usual HMF coordinate frame (biclusters), but in the angle coordinate frame clustering is only found for the initially ordered angles and, unlike the biclusters, these are clearly quasistationary states. A quasistationary state is defined as a dynamical state that can persist for a length of time which goes to infinity as the thermodynamic limit is approached . In addition to discussing the clustered angle states exhibited by the system, we also solve for the canonical partition function in the pendulum angle coordinate frame, finding that in equilibrium with a heat bath, the probability distributions of the angles can be described by the original HMF model. This finding is similar to the work done by on a model sometimes called the HMF $\alpha$-model. In the HMF $\alpha$ model, a $1/r_{ij}^\alpha$ dependence between the classical spins is introduced , where $r_{ij}$ is the distance between the $i^{\mathrm{th}}$ and $j^{\mathrm{th}}$ spins on a lattice. Though the physical motivations behind studying these various models can be very different, it is interesting that their equilibrium behavior is the same or nearly the same. We believe that the work in this paper further suggests that the HMF model universally describes an entire class of long-range interacting systems in equilibrium. \section{The Model} \subsection{Coordinates} In Fig.~, we show an array of pendula rotating in the same plane with bobs that interact through a long-range potential. If we consider the case where all the pendula only undergo small oscillations, we may write the horizontal location of the $i^{\mathrm{th}}$ particle, $x_i$, as $x_i = id + \ell\theta_i$, where $d$ is the distance between the pivots of neighboring pendula and $\ell$ is the length of each pendulum. The small $\theta$ regime makes the problem one dimensional in $x$. We choose periodic boundary conditions and rescale the system by $2\pi/Nd$ so that \begin{equation} x\rightarrow \frac{2\pi}{Nd} x \end{equation} making the total system length a dimensionless $2\pi$ where N is the number of particles in one period. We will refer to a periodic space with length $2\pi$ as a unit circle. The position of the $i^{\mathrm{th}}$ particle (bob) is now \begin{equation} x_i = \frac{2\pi i}{N} + \frac{2\pi}{N}\frac{\ell}{d}\theta_i . \end{equation} For reasonable choices of $\ell$ and $d$ ($\ell/d << N$), the second term on the RHS is suitably small such that the Hamiltonian can be written with terms that are quadratic in $\theta$. However, we are primarily interested in a regime where $\ell/d \rightarrow \infty$ as the thermodynamic limit is approached. Physically this corresponds to the small oscillations of very long pendula with suspension points that are close together compared to their lengths. In order to simplify the calculations that follow, we define $\phi_i$ to be the last term on the RHS of Eq.~(), namely $\phi_i \equiv 2\pi \ell \theta_i / Nd $. Given the choice of large $\ell/d$, $\phi_i$ can take any value in the range $[0,2\pi)$. This is only true because $\ell/d$ is large, \textit{not} because the $\theta_i$s are. In terms of $\phi_i$, the positions can be rewritten as \begin{equation} x_i = \frac{2\pi i}{N} + \phi_i . \end{equation} \subsection{Density Approximation} We have not yet explicitly stated the physical mechanism through which the bobs interact. Connecting the interactions with specific physical motivations should be discussed with some discretion because the development of the model leaves these motivations up to some freedom of interpretation. Imagine that the bobs all carry some charge. We will not distinguish between particles in any other way than their indices, so in the case where all particles carry the same charge, repulsive behavior is expected. On the other hand one could make the bobs attract one another, which could be thought of as the self-gravitating case. To be solvable, the model requires some simplifications. For the sake of brevity we will speak of the particle charge or mass density as the ``density''. The approximation that we invoke is similar to that used when justifying the HMF model (Eq.~()) to describe free particles in a one-dimensional ring . The distribution of the bobs is such the mass density, $\rho(x)$, is given by \begin{equation} \rho(x) = \sum_{i=1}^{N}\delta(x-x_i) - \frac{1}{2\pi} . \end{equation} The constant $1/2\pi$ subtracted from the delta function is necessary to produce a meaningful expression for the potential $\Phi$ and corresponds to the inclusion of a neutralizing (of opposite sign) homogeneous background density. Restricting the problem further to that of solving Poisson's equation for a one-dimensional potential physically amounts to choosing large and flat bob geometries oriented with their smallest axis parallel to the $x$ axis. Writing the delta function as a cosine Fourier series, Poisson's equation becomes \begin{equation} \nabla^2\Phi(x) = \frac{\gamma}{\pi}\sum_{i=1}^{N}\sum_{n=1}^{\infty} \cos{[n(x-x_i)]} . \end{equation} The parameter $\gamma$ contains the particle (bob) charge or mass and becomes the interaction strength in the Hamiltonian. We can see that the zeroth-order term in the Fourier series canceled the constant neutralizing background that was superficially added. The most important simplification in this paper is truncating the sum of the Fourier coefficients used to represent the delta function after the $n = 1$ coefficient. Antoni et al. defend the truncation by asserting that the ``large scale collective properties'' do not greatly change when higher order terms of the sum (including interactions at the smaller length scales) are included, and discuss the consequences of the approximation in some detail . The simplification also warrants a brief discussion of the way that it could be physically interpreted. The truncation of the sum is equivalent to smearing out the density of each particle over the system so that it is peaked at its given location, $x_i$, but also having a negative density peak on the opposite side of the unit circle. This could be thought of as doubling the number of particles and enforcing that each particle has a negative partner that always remains on the opposing side of the unit circle. After this doubling, the now nebulous masses are dispersed such that a pair's density is described by a cosine function with the positive peak centered at $x_i$. \subsection{Solving Poisson's Equation} Integrating Poisson's equation once, we obtain: \begin{equation} \nabla\Phi(x) = \frac{\gamma}{\pi}\sum_{i=1}^{N} \left\{ \sin{(x-x_i)} + c_1 \right\} . \end{equation} In order to determine the constant $c_1$ from the integration, the physical picture should be examined. A sensible requirement is that when all of the bobs are hanging at their equilibrium positions, directly below their pivot (all $\phi_i=\theta_i=0$), the net force experienced by any bob is zero. This is a valid requirement if the bobs are attractive or repulsive, the only difference being that the configuration would be unstable or stable, respectively. The force that the $j^{\mathrm{th}}$ particle experiences when $\phi_j$ and all $\phi_i$ are zero is given by \begin{equation} -\nabla\Phi\left(x_j \right) = -\frac{\gamma}{\pi}\sum_{i=1}^{N} \left\{ \sin { \left[ \frac{2\pi (j-i)}{N} \right] } + c_1 \right\} . \end{equation} The sum $\sum_i \sin{[2\pi (j-i)/N]}$ equals zero for any $j$, so $c_1$ must be zero. Integrating once more to obtain the potential yields \begin{equation} \Phi(x) = \frac{\gamma}{\pi}\sum_{i=1}^{N} \left[ c_2-\cos{\left(x-\frac{2\pi i}{N} - \phi_i \right)} \right] . \end{equation} To determine $c_2$ we stipulate that if all $\phi_i=0$, then $\Phi(0)=0$. Inserting Eq.~() (or Eq.~()) for $x_i$ yields \begin{equation} \Phi(0) = \frac{\gamma}{\pi}\sum_{i=1}^{N} \left[ c_2-\cos{\left(\frac{2\pi i}{N}\right)} \right] . \end{equation} The sum over the cosine is zero, therefore $c_2 = 0$ and we can now write the potential energy of the $j^{\mathrm{th}}$ particle as \begin{equation} \Phi(x_j) = -\frac{\gamma}{\pi}\sum_{i=1}^{N} \cos{\left[\frac{2\pi (j-i)}{N}+\phi_j - \phi_i \right]} . \end{equation} \noindent \subsection{The Hamiltonian} The Hamiltonian can be written as \begin{equation} H = H_0 + H_I , \end{equation} where $H_0$ is the kinetic energy piece \begin{equation} H_0 = \sum_{i=1}^{N} \frac{p_i^2}{2} , \end{equation} and \begin{equation} H_I = -\frac{\gamma}{2N}\sum_{i,j=1}^{N} \cos{ \left[ \frac{2\pi (i-j)}{N} + \phi_i-\phi_j \right] } \end{equation} is the interaction piece, so \begin{equation} H = \sum_{i=1}^{N} \frac{p_i^2}{2} -\frac{\gamma}{2N}\sum_{i,j=1}^{N} \cos{ \left[ \frac{2\pi (i-j)}{N} + \phi_i-\phi_j \right] } . \end{equation} The mass of the bobs has been set to unity, $\gamma$ is the interaction strength, a factor of $1/2$ accounts for the double counting, and the $1/\pi$ coefficient in the potential energy has been absorbed into $\gamma$. The factor of $1/N$ is a rescaling of the potential energy that ensures that as the thermodynamic limit is approached, the potential energy of the system does not diverge. The $1/N$ scaling is known as the Kac prescription . The Kac serves to keep both the energy and entropy of a system proportional to the number of particles in the system, an important prerequisite for phase transitions . \subsection{Relationship to the HMF model and the Spin Interpretation} Due to the simple bijective relationship between $x_i$ and $\phi_i$ one can simply solve the equations of motion for the HMF model and find the dynamics for $\phi_i$ via the coordinate transform $x_i \rightarrow \phi_i$. Previously it was mentioned that the HMF model is used to describe free particles on a ring with long-range repulsion or attraction, as well as describing a classical $XY$ spin model. The $\theta_i$ played the role of either the position of the $i^{\mathrm{th}}$ particle on the ring or the orientation of the $i^{\mathrm{th}}$ spin. Therefore, it is interesting to speculate about the type of spin system the model describes in the $\phi_i$ picture. Thus far, the rescaled angle $\phi_i = 2\pi \ell \theta_i / Nd$ describes the distance of a pendulum bob from the point directly below its pivot, but it could also be interpreted as the orientation of spin. In the spin interpretation of Eq.~(), the potential energy of the $i^{\mathrm{th}}$ and $j^{\mathrm{th}}$ spin pair depend on both their relative orientation as well as the difference between their indices. In the following discussion it will sometimes be convenient to speak about $\phi_i$ in the spin language. We will prove that in the $\phi$ picture, the system in equilibrium with a heat bath is equivalent to the HMF model (the $x$ picture) in equilibrium with a heat bath by solving the partition function in the $\phi_i$ coordinate frame. In the process of simplifying the Hamiltonian to solve for the partition function, we will find expressions of the form $\cos{\phi_i}$ and $\sin{\phi_i}$ which we talk about as the horizontal and vertical components of a magnetization $\vec{m}_i = (\cos{\phi_i},\sin{\phi_i})$. It could easily be stated that in the spin analogy, the $\phi_i$ are orientations of the spins, but we should make a more concrete connection between this idea and the original presentation of the model. We would like to remind the reader that even though the angles $\theta_i$ of the pendula are small, the long suspensions of the bobs ($\ell$) allow $\phi_i$ to cover the entire system which, rescaled, has dimensionless length $2\pi$. The system is also periodic, so the bobs can be thought of as moving on a unit circle where the position of the $i^{\mathrm{th}}$ bob is $x_i=2\pi i/N + \phi_i$. In order to think of $\phi_i$ as the spin orientations, we start by considering each bob as living on its own individual unit circle. An example of these unit circles is shown in Fig.~, a visual aid to the following. Imagine stacking horizontal circles in the vertical direction and rotating each by an angle $2\pi/N$ from the one below. The projection of these circles onto the horizontal plane would be the system viewed in $x$, i.e. the HMF model. If we twist the stack so there is no rotation between adjacent circles and then project onto the horizontal plane it creates the picture viewed in $\phi$, where the pivot points are all aligned. The reason for this artificial construction of stacked circles is partly to pictorially depict the transformation between $x$ and $\phi$ and partly to show how $\vec{m}_i$ (as defined) is just the orientation of the $i^{\mathrm{th}}$ spin in the $\phi$ picture. Said differently, each circle in the $\phi$ picture represents a spin with an orientation in the horizontal plane determined by $\phi_i$; an infinite-range classical mean field spin model described by Eq.~(). \section{Equilibrium} In this section, we solve for the canonical partition function, in the $\phi$ coordinate frame using the Hamiltonian in Eq.~() and show that, in equilibrium, the HMF model describes the angles of long pendula with long-range interacting bobs. In order to solve the configurational piece of the partition function the Hamiltonian must be modified. Using the cosine and sine sum and difference identities twice, the potential interaction piece of the Hamiltonian $H_I$ can be written as \begin{multline} H_i = \frac{-\gamma}{2N} \sum_{i,j} \Bigg\{ \cos{\left[ \frac{2\pi (i-j)}{N} \right]} [ \cos{\phi_i}\cos{\phi_j} \\ + \sin{\phi_i}\sin{\phi_j} ] \\ - \sin{\left[ \frac{2\pi (i-j)}{N} \right]} \left[ \sin{\phi_i}\cos{\phi_j} - \cos{\phi_i}\sin{\phi_j} \right] \Bigg\} . \end{multline} \noindent The coefficients in the Hamiltonian $\cos{[2\pi (i-j)/N]}$ and $\sin{[2\pi (i-j)/N]}$ should be thought of as matrices with components $C_{ij}$ and $S_{ij}$ respectively. The Hamiltonian can now be written in the form \begin{multline} \frac{\gamma}{2N}\sum_{i,j} ( \cos{\phi_i}C_{ij} \cos{\phi_j} + \sin{\phi_i}C_{ij}\sin{\phi_j} \\ - \sin{\phi_i}S_{ij}\cos{\phi_j} + \cos{\phi_i}S_{ij}\sin{\phi_j} ), \end{multline} \noindent which is suggestive because it can be regarded as the matrix equation \begin{multline} H_I = \frac{\gamma}{2N} \Bigg[ (\cos{\phi_1},\cos{\phi_2},...,\cos{\phi_N}) C \begin{pmatrix} \cos{\phi_1} \\ \cos{\phi_2} \\ \vdots \\ \cos{\phi_N} \end{pmatrix} \\ + (\sin{\phi_1},\sin{\phi_2},...,\sin{\phi_N}) C \begin{pmatrix} \sin{\phi_1} \\ \sin{\phi_2} \\ \vdots \\ \sin{\phi_N} \end{pmatrix} \\ - (\sin{\phi_1},\sin{\phi_2},...,\sin{\phi_N}) S \begin{pmatrix} \cos{\phi_1} \\ \cos{\phi_2} \\ \vdots \\ \cos{\phi_N} \end{pmatrix} \\ + (\cos{\phi_1},\cos{\phi_2},...,\cos{\phi_N}) S \begin{pmatrix} \sin{\phi_1} \\ \sin{\phi_2} \\ \vdots \\ \sin{\phi_N} \end{pmatrix} \Bigg] . \end{multline} \noindent It is helpful to consider the particles positions on the unit circle with respect to their pivot ($\phi$) as magnetizations. Defining \begin{equation} \vec{m}_i \equiv (\cos{\phi_i},\sin{\phi_i}), \end{equation} \noindent and with $m^T_{\mu} = (m_{0,\mu},m_{1,\mu},...,m_{N-1,\mu})$ where $\mu$ holds the place of $x$ or $y$, the Hamiltonian becomes \begin{multline} H_I = \frac{\gamma}{2N} ( m^T_x C m_x + m^T_y C m_y \\ - m^T_y S m_x + m^T_x S m_y ) . \end{multline} A closer examination of the structure of the coefficient matrices $C$ and $S$ indicates that they take the special form of circulant matrices, and thus can be simultaneously diagonalized by a unitary matrix $U$. A circulant matrix has the form \begin{equation} \left( \begin{matrix} a_1 & a_2 & a_3 & \ldots & a_{N-1} & a_N \\ a_N & a_1 & a_2 & \ldots & a_{N-2} & a_{N-1} \\ a_{N-1} & a_N & a_1 & \ldots & a_{N-3} & a_{N-2} \\ \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\ a_3 & a_4 & a_5 & \ldots & a_1 & a_2 \\ a_2 & a_3 & a_4 & \ldots & a_N & a_1 \\ \end{matrix} \right) , \end{equation} a special kind of Toeplitz matrix, where each subsequent row is a cyclic permutation of the row above or below it. Any matrix $A$ with elements $a_{ij}$ that can be written in terms of some function $f(i-j)$ is a circulant matrix. Because a circulant matrix is a normal matrix it can be diagonalized by a unitary matrix. We show that $C$ and $S$ are simultaneously diagonalizable by showing that they commute, i.e. $[C,S]=0$ where $[C,S] = CS-SC$. Starting with the second term, $-SC = S^TC^T=(CS)^T$ which is found by arguing that $C$ is symmetric since cosine is an even function and does not change under the exchange of $i$ and $j$, whereas $S$ is odd because sine is an odd function and does change sign under exchange of $i$ and $j$. The commutation becomes $[C,S] = CS + (CS)^T$. Also, an odd function multiplied by an even function results in an odd function so the entire matrix $CS$ is odd. Therefore $(CS)^T = -CS$ bringing us to the final expression $[C,S] = CS-CS=0$. We have shown that $C$ and $S$ can be simultaneously diagonalized by $U$. The matrix $U$ is known for circulant matrices and is called the Fourier Matrix. The matrices $C$ and $S$ can be rewritten as $C= U^\dagger D^C U$ and $S = U^\dagger D^S U$, where $D^C$ and $D^S$ are diagonal matrices with diagonal elements that are the eigenvalues of $C$ and $S$, respectively, which we denote as $\lambda^C_i$ and $\lambda^S_i$. From here on we label the indices $i$ from $0$ to $N-1$. It is worth pointing out that due to $S$ being antisymmetric, $U$ must be complex. Equation~() becomes \begin{multline} H_I = \frac{\gamma}{2N} \Big( m^T_x U^\dagger D^C U m_x + m^T_y U^\dagger D^C U m_y \\ - m^T_y U^\dagger D^S U m_x + m^T_x U^\dagger D^S U m_y \Big) . \end{multline} We will move back to the index notation using the following relations: \begin{equation} D^{C,S} = \lambda_i^{C,S}\delta_{ij} , \end{equation} \begin{equation} X_j \equiv \sum_{k=1}^N U_{jk} m_k^x , \end{equation} and \begin{equation} Y_j \equiv \sum_{k=1}^N U_{jk} m_k^y. , \end{equation} where $X$ is not to be confused with $x$. Using the Kronecker delta, we set all $i=j$ since these are the only nonzero terms. The Hamiltonian is now given by \begin{multline} H_i = \frac{-\gamma}{2N}\sum_{j=0}^{N-1} \Big( \|X_j\|^2 \lambda_j^C + \|Y_j\|^2 \lambda_j^C \\ -Y_j^* X_j \lambda_j^S + X_j^* Y_j \lambda_j^S \Big) , \end{multline} with $\|X\|= XX^*$ and $\|Y\|= YY^*$. The inclusion of the eigenvalues $\lambda^C$ and $\lambda^S$ simplifies the Hamiltonian further. We will now solve for $\lambda^C$ and $\lambda^S$. Looking at the form of a circulant matrix shown in Eq.~() reminds us that the elements of a circulant matrix can be defined with a single label. We write the single labeled elements of the cosine and sine matrices respectively as \begin{equation} c_l = \cos{\frac{2\pi l}{N}} , \end{equation} and \begin{equation} s_l = \sin{\frac{2\pi l}{N}} , \end{equation} where $l=0,1,2,...,N-1$. The eigenvalues, $\lambda^A$, of a $N\times N$ circulant matrix $A$ can be written in terms of the single label elements $a_l$. The $j^{\mathrm{th}}$ eigenvalue of $A$ is known to be $\lambda^A_j=\sum_{l=0}^{N-1} a_l \exp{(2\pi i l j / N)}$, where $i$ is $\sqrt{-1}$ (not an index) and $l=0,1,2,...,N-1$. Therefore, \begin{equation} \lambda_j^C = \sum_{l=0}^{N-1} \cos{\left(\frac{2\pi l}{N}\right)} e^{2\pi i l j /N} , \end{equation} and \begin{equation} \lambda_j^S = \sum_{l=0}^{N-1} \sin{\left(\frac{2\pi l}{N}\right)} e^{2\pi i l j /N} . \end{equation} Writing cosine and sine in their exponential forms gives \begin{equation} \lambda_j^C = \frac{1}{2} \sum_{l=0}^{N-1} \left[ e^{i 2 \pi l (j+1)/N} + e^{i 2 \pi l (j-1)/N} \right] , \end{equation} \begin{equation} \lambda_j^S = \frac{-i}{2} \sum_{l=0}^{N-1} \left[ e^{i 2 \pi l (j+1)/N} + e^{i 2 \pi l (j-1)/N} \right] , \end{equation} The above representations of the eigenvalues show that $C$ and $S$ each have only two non-zero eigenvalues corresponding to $j=1,N-1$ given by $\lambda_1^C = \lambda_{N-1}^C = N/2$ and $\lambda_1^S=(\lambda_{N-1}^{S})^* = iN/2$. The Hamiltonian simplifies greatly to \begin{equation} H_I = \frac{\gamma}{2} - \left( \| X_1+iY_1 \|^2 + \| X_{N-1} - iY_{N-1} \|^2 \right). \end{equation} The representation of $H_I$ in Eq.~() must be further modified before the partition function can be found. We do this by splitting the Fourier matrix $U$ into its real and imaginary components, $a_{ik}$ and $b_{ik}$, given by \begin{equation} a_{ik} \equiv \frac{1}{\sqrt{N}} \cos{\left(\frac{2 \pi i k}{ N} \right)}, \end{equation} and \begin{equation} b_{ik} \equiv \frac{1}{\sqrt{N}} \sin{\left(\frac{2 \pi i k }{ N} \right)} . \end{equation} This was done to write the absolute squares in Eq.~() in terms of the squares of $a_{ik}$ and $b_{ik}$. By noticing that $a_{1k} = a_{(N-1)k}$ and $b_{1k} = -b_{(N-1)k}$ we write the configurational partition function as \begin{multline} Z_I = A \int d^N\phi e^ { \frac{\beta \gamma}{2} \left( \sum_k \left[ a_{1k} m_k^x - b_{1k} m_k^y \right] \right)^2 } \\ \times e^ { \frac{\beta \gamma}{2} \left( \sum_k \left[ b_{1k} m_k^x + a_{1k} m_k^y \right] \right)^2 } , \end{multline} \noindent where $\beta=1/k_B T$. The Hubbard-Stratonovich transformation is now applied twice, once to each quadratic quantity in the partition function. The integration variables introduced through this transformation are $z_1$ and $z_2$ with subscripts for first and second quadratic quantities, respectively. After after switching the order of integration, we find \begin{multline} Z_I = \frac{A }{2 \pi \beta \gamma} \int_{-\infty}^{\infty} dz_1 dz_2 e^ { -(z^2_1+z^2_2)/ 2 \beta \gamma } \prod_k \\ \times \int_{-\pi}^{\pi} d \phi_k e^ { (z_{1}a_{1k}+z_{2}b_{1k})\cos{\phi_k} + (z_{2}a_{1k}-z_{1}b_{1k})\sin{\phi_k} }. \end{multline} The integration can be performed using the identity \begin{equation} \int_{-\pi}^{\pi} d\phi e^{\xi \cos{\phi} + \eta \sin{\phi}} = 2\pi I_0 \left( \sqrt{\xi^2 + \eta^2} \right) \end{equation} where \begin{equation} \xi^2 + \eta^2 = (z_1 a_{1k} + z_2 b_{1k})^2 + (z_2 a_{1k} - z_1 b_{1k})^2 \end{equation} which simplifies when $a$ and $b$ are included to \begin{multline} \left[ z_1 \frac{1}{\sqrt{N}}\cos{\left(\frac{2\pi k}{N}\right)} + z_2 \frac{1}{\sqrt{N}}\sin{\left(\frac{2\pi k}{N}\right)} \right]^2 \\ + \left[ z_2 \frac{1}{\sqrt{N}}\cos{\left(\frac{2\pi k}{N}\right)} - z_1 \frac{1}{\sqrt{N}}\sin{\left(\frac{2\pi k}{N}\right)} \right]^2 \\ = \frac{1}{N}(z_1^2 + z_2^2) \end{multline} It is convenient to make a change to polar coordinates by introducing $z=\sqrt{z_1^2 +z_2^2}$, following which the partition function can be written as \begin{equation} Z_I = \frac{A }{ \beta \gamma} \int_{-\infty}^{\infty} dz e^ { -z/ 2 \beta \gamma } \prod_k 2 \pi I_0 \left( \frac{\sqrt{z}}{\sqrt{N}} \right) . \end{equation} Equation~() is recognized to be an intermediate step of the solution to the canonical partition function for the HMF model. From here we jump to the main results, the details of which are included in the HMF literature . The integration over $z$ in Eq.~() can be preformed using the saddle point approximation. The rescaled free energy per particle follows as \begin{equation} -\beta F = - \frac{\beta}{2} + \inf_z \left[ \frac{- z^2}{2\beta} + \ln{2\pi I_0(z)} \right] \end{equation} The expression above permits a convenient path to finding the phase transition. Solving for the minimum values of $z$ in order to satisfy the last term in Eq.~() results in the equation \begin{equation} \frac{z}{\beta} - \frac{I_1(z)}{I_0(z)} = 0 , \end{equation} \noindent which can be solved self consistently for $z$ and represented graphically for different values of $\beta$ as in Fig.~. The reader will see that after $\beta$ is increased passed the critical value ($\beta=2$) there are two well-defined solutions. The Hubbard-Stratonovich transformation decouples spin-spin (squared terms in the Hamiltonian) contributions to the partition function at the price of needing to create a linear interaction between each spin with an auxiliary field $z$ . Again, a more detailed procedure can be found in where discussion of the internal energy in the equilibrium state is followed by non-equilibrium behavior of the system prepared in microcanonical ensembles. Here we will simply touch on the most important point of the equilibrium behavior, being that for $\beta<2$ the system is paramagnetic but for $\beta \ge 2$ a pitchfork bifurcation occurs resulting in two stable solutions. At this point there is a discontinuity in the free energy, a second order phase transition occurs and the system can maintain finite magnetization. In this case, the order parameter is the total magnetization $\vec{M}=\frac{1}{N}\sum_{j=1}^{N} \vec{m}_i$ where $\vec{m}_i$ was defined to be $(\cos{\phi_i},\sin{\phi_i})$. Showing that the canonical partition function in the $\phi$ coordinate frame model and the HMF model are equivalent necessitates a more detailed discussion of the equilibrum behaviors in the $\phi$ frame. Campa et al. , in their review of the HMF model, rigorously show ensemble equivalence between the canonical and microcanoical ensemble of the HMF model. In light of this fact, a large $N$ microcanonical simulation should be able to produce equilibrium behavior like the phase transition mentioned above. The temperature in a numerical simulation of a system with many particles would be ``set'' through a choice of the initial momenta distribution. In this type of simulation, it is common practice to compute the order parameter and free energy , begging the question: does a large microcaonical simulation of Eq.~() approximate the expected equilibrium behavior? Also, since the index-dependent model in equilibrium with a heat bath can be described by the HMF model, would the dynamics of such a simulation qualitatively resemble those in the HMF model? The answer to both of these questions is \textit{no} if one were to find the equations of motions in $\phi_i$ for some large $N$ and then compare them to an HMF model or the $x$ coordinate frame. As stated, this discrepancy may appear to detract from our result. Indeed, it uncovers a conceptual omission in the model, but it is one whose rectification gives insight into the models ensemble equivalence property of the model, or lack thereof. The omission was in the arbitrary scaling of $x$ which we will now rectify. We introduce the parameter $L$ which generalizes the scaling in Eq.~() to \begin{equation} x\rightarrow \frac{2\pi L}{Nd} x , \end{equation} making the position of the $i^{\mathrm{th}}$ particle \begin{equation} x_i = \frac{2\pi L i}{N} + \frac{2\pi L}{N}\frac{\ell}{d}\theta_i . \end{equation} and changing the definition of $\phi_i$ to $\phi_i\equiv 2\pi L\ell\theta_i/Nd$. It can be shown that the introduction of $L$ only changes the final result of the partition function by a constant factor of $L$ due to the enlarged limits of integration. Numerically, we find is that if $L>>1$, then the simulations in $\phi$ closely reproduce the dynamics of HMF model simulations (dynamics in $x$). Therefore, for large $L$ the mirocanonical simulations can approximate equilibrium and the answers to the previous questions - does a large microcaonical simulation of Eq.~() approximate the expected equilibrium behavior, and since the index-dependent model in equilibrium with a heat bath can be described by the HMF model, would the dynamics of such a simulation qualitatively resemble those in the HMF model? - becomes \textit{yes}. Alternatively, the coordinate frame inequivelence is most extreme for small $L$. These numerical results were found using initial conditions that are randomly distributed $\phi_i$ about the domain $[-L\pi,L\pi)$. It should be stated that for the rest of this paper we work with $L=1$ becuase we are inetersed in cases where the $\phi$ coordinate frame is markedly differnet than the $x$ coordinate frame. \section{Non-Equilibrium Results} For a system of pendula, it is interesting to study an initial configuration where all pendula are set to random small displacements from $\phi_i=0$. Specifically we initialize the $i^{\mathrm{th}}$ pendulum angle, $\phi_i$, randomly in the range $[-\pi/N,\pi/N)$. In $x$ the indices are ordered in $x$ such that $x_1<x_2<x_3<...<x_N$ and the $i^{\mathrm{th}}$ bob is randomly distributed in the range $[2\pi i/N-\pi/N,2\pi i/N+\pi/N)$. It is possible to make some general statements about the dynamics of this configuration in $x$ using the equations of motion. Expressing the Hamiltonian with terms that are quadratic in $\phi$ yields \begin{multline} H_I = \frac{\gamma}{2N}\sum_{ij} \Bigg\{ \sin { \left[ \frac{2 \pi (i-j)}{N} \right] } (\phi_i - \phi_j) \\ -\cos { \left[ \frac{2 \pi (i-j)}{N} \right] } \left( 1-\frac{\phi_j^2}{2} - \frac{\phi_i^2}{2}+\phi_i \phi_j \right) \Bigg\} . \end{multline} \noindent With this expression, the equations of motion for the $i^{\mathrm{th}}$ particle can be written as \begin{equation} \ddot{\phi}_i = \dot{p}_i = \frac{\partial H}{\partial \phi_i}, \end{equation} \noindent from which we obtain \begin{multline} \ddot{\phi}_i = \frac{-\gamma}{2N}\sum_{j} \Bigg\{ \cos{\left[\frac{2\pi(i-j)}{N}\right]}(\phi_j-\phi_i) \\ + \sin{\left[ \frac{2\pi (i-j)}{N} \right]} \Bigg\} . \end{multline} \noindent In the above equation, the last term and the $\cos{[2\pi(i-j)/N]}\phi_i$ term sum to zero, leading to \begin{equation} \ddot{\phi}_i = \frac{-\gamma}{2N} \sum_j \cos{\left[\frac{2\pi(i-j)}{N}\right]}\phi_j . \end{equation} \noindent Using the difference formula, we write the acceleration as \begin{equation} \ddot{\phi}_i = \frac{-\gamma}{2} \left[ \cos{\left( \frac{2\pi i}{N} \right)} \langle \mu_1 \rangle + \sin{\left( \frac{2\pi i}{N} \right)} \langle \mu_2 \rangle \right] , \end{equation} \noindent where $\mu_1 = \phi_j\cos{(2\pi i/N)}$ and $\mu_2 = \phi_j\sin{(2\pi i/N)}$. The mass (moment of inertia) has been set to unity so the above expression is the force as a function of index, $\ddot{\phi_i} = F(i)$. If $\langle \mu_1 \rangle$ and $\langle \mu_2 \rangle$ are known, then the initial dynamics of the system are elucidated by Eq.~(), but in the case of randomly initialized $\phi_i$ the $\langle \mu_1 \rangle$ and $\langle \mu_2 \rangle$ are also random and can be different from one another in both magnitude and sign. However, a general description of the results can be given without exactly knowing these coefficients. Equation~() shows that the initial force on a given particle depends on its position because the indices are ordered in $x$. In the continuum (thermodynamic limit), the force takes the form \begin{equation} F(x) \equiv \frac{-\gamma}{2} \left( \langle \mu_1 \rangle \cos{x} + \langle \mu_2 \rangle \sin{x} \right) . \end{equation} \noindent Therefore $\langle \mu_1 \rangle$ and $\langle \mu_2 \rangle$ partly play the role of the amplitude of this force as a function of $x$, but also can shift the $\cos{x}+\sin{x}$ spatial dependence, which is periodic over the system length. In Fig.~, we show a fit of the force as a function of $x$ using Eq.~() as well as the actual force calculated for an example set of initial conditions. The domain in Fig.~ can be split into two pieces (independent of $\mu$)- one where the particles experience a positive force, the other in which the particles experience a negative force. As time is increased, the movements of the particles evolve the coefficients $\langle \mu_1 \rangle$ and $\langle \mu_2 \rangle$ in such a way that the magnitude of the force decreases to zero for all particles and then switches sign complementary to the original force. This results in a standing compression wave of the particles with a wavelength $2\pi$. The compression wave is not stable and eventually two clusters form about each node. These two clusters are often referred to as a ``bicluster'', or the antiferromagnetic state in the HMF model, and have been explained by Barr\'{e} et al. by analysing the Vlasov equation. They find that the initial compression wave (referred to by a different name) creates an effective double-well potential giving rise to the bicluster . The question of the bicluster stability has not yet been definitely answered, but for a detailed discussion we refer the reader to Leyvraz et al. . Given the simple mapping between the $\phi$ and $x$ coordinate frames, we should also be able to show the initial form of the force in $x$ as well. As presented in Eq.~(), the Hamiltonian in the $x$ coordinate frame only differs from the HMF model by a constant $\gamma/2$. In $x$, $H_I$ is \begin{equation} H_I= \frac{\gamma}{2N}\sum_{i,j=1}^{N}\cos{( x_i - x_j )}. \end{equation} Using the difference identity, we find the equations of motion for the $i^{\mathrm{th}}$ particle to be \begin{equation} \ddot{x}_i \frac{-\gamma}{2N} \left( -\sin{x_i}\sum_j \cos{x_j} + \cos{x_i}\sum_j \sin{x_j} \right). \end{equation} The sums over cosine and sine of $x_j$ play the same role as $\langle \mu_1 \rangle$ and $\langle \mu_2 \rangle$, and the force at a given position $x_i$ is clearly of the same form as that shown in Eq.~(). Depending on $\langle \mu_1 \rangle$ and $\langle \mu_2 \rangle$, all $\phi_i$ oscillate about zero with amplitudes and phases that depend on their location $x_i$ as discussed above. As the clustering in $x$ begins, the $\phi_i$ begin to spread out over the full domain $[0,2\pi)$ and continue to do so until it is covered. The more interesting case in $\phi$ is when all $\phi_i$ are initially randomly distributed in ranges that depend on their index, specifically when $\phi_i$ are chosen in the ranges. $[2\pi i/N-\pi/N,2\pi i/N+\pi/N)$ so that $\phi_1<\phi_{2}<\phi_{3}<...<\phi_{N}$. It should be noted that in this new configuration the dynamics in $x$ are nearly identical to the configuration previously discussed for ordered $x_i$. The dynamics in $\phi$ differ \textit{drastically} between the two cases though. In this ordered angle case, we find some interesting grouping of the scaled angles. Initially the bobs oscillate with an amplitude that depends sinusoidally on their position in $\phi$, similar to the previous discussion in the $x$ picture but with four nodes where the $\phi_i$ remain relatively stationary. Once again this behavior could be thought of as a standing compression wave, but in $\phi_i$ it has a wave length of $\pi$ whereas in the $x$ picture it had a wavelength of $2\pi$. As the system evolves, all $\phi_i$ slowly begin to shift towards the nodes of this standing wave until there are four clusters of the angles. After some time, the angles begin to re-distribute themselves randomly about the domain. The distribution of $\phi_i$ in these three regimes is summarized in three histograms shown in Fig.~. Aside from the number of clusters, there are two primary differences between the clustering in $\phi$ and the clustering in $x$: (i) The clustering in $\phi$ \textit{only} occurs when the angles are ordered in the method described above, whereas the dynamics in $x$ look identical regardless of the distribution in $x$, presuming it is somewhat homogeneous about the domain. (ii) The clustering in $\phi$ is a quasistationary state whereas the clustering in $x$ exists for much longer times regardless of the system size. Since the clustering in $\phi$ is quasistationary, a properly prepared system could exist in the clustered angle state for an arbitrarily long time but only for large $N$. We can view the effect of increasing $N$ and therefore the lifetimes of the clustered states by observing the order of the particle index as a function of time. In Fig.~(a)-(c), we show that as $N$ is increased, the time it takes for particles to fully mix increases. This is shown by plotting the indices on a color scale from 0 (blue) to $N-1$ (red) along the horizontal axis as time is increased along the vertical axis. In Fig.~(d), we show how the ordering of the particles changes at the very beginning of clustering for $N=100$. Figure.~ also shows that the compression wave is not quasistationary since it quickly reduces to the clustered state regardless of $N$. \section{Conclusion} Though the application of statistical mechanics and thermodynamics to systems with long-range interactions may not always be appropriate, we find that the canonical partition function improves our understanding of a system of pendula with long-range interacting bobs. Solving for the canonical partition function of the Hamiltonian in Eq.~(), we show that the equilibrium behavior in the $\phi$ coordinate frame is equivalent to the $x$ coordinate frame, i.e. the HMF model. As we have argued that the Hamiltonian in Eq.~() describes the behavior of the angles of repulsive or attracting pendulum bobs (see Fig.~), then the proven equivalence of the canonical partition function of Eq.~() and the Hamiltonian mean field model suggests that the Hamiltonian mean field model sufficiently describes the angles of a system of pendula in equilibrium. Ensemble equivalence between the microcanonical ensemble and the canonical ensemble is known for the Hamiltonian mean field model model and because of this, the microcanonical simulations could be used to approximate equilibrium behavior. We find numerically that in the case of large system lengths, $L$, the dynamics of the system in $\phi$ resemble the dynamics of the Hamiltonian mean field model, equivalently the behavior of the system in $x$. Therefore for large system sizes of long pendula in equilibrium, the HMF model describes their dynamics and statistics. In this paper we also briefly discuss two particular sets of non-equilibrium results. In one case, the system is initialized with small $\phi_i$ so that $x_i$ are distributed relatively evenly throughout the $x$ domain. This initial configuration essentially gives rise to the ``repulsive'' low temperature HMF model which exhibits interesting non-equilibrium behavior and is described in great detail by . In the second case, in which $\phi_i$ are ordered by their index $i$, we show there is a compression wave in $\phi$, followed by clustering, and finally a mixed index state displaying no apparent order or structure. This is in contrast to the dynamics produced by a randomly distributed set of initial $\phi_i$ which begins and then remains in a random disordered state. The clustering that can occur in $\phi$ is different from the clustering in $x$ because it only occurs when the angles are initially ordered and because it is quasistationary; the lifetime increases with the number of particles in the system. \begin{acknowledgments} We would like to acknowledge the support and generosity of Anand Sharma. This work is partially supported by a contract from NASA (NNX13AD40A). \end{acknowledgments} \bibliography{long_range_refs} |
1501.04117 | Title: Coherent control of plasma dynamics
Abstract: Coherent control of a system involves steering an interaction to a final
coherent state by controlling the phase of an applied field. Plasmas support
coherent wave structures that can be generated by intense laser fields. Here,
we demonstrate the coherent control of plasma dynamics in a laser wakefield
electron acceleration experiment. A genetic algorithm is implemented using a
deformable mirror with the electron beam signal as feedback, which allows a
heuristic search for the optimal wavefront under laser-plasma conditions that
is not known a priori. We are able to improve both the electron beam charge and
angular distribution by an order of magnitude. These improvements do not simply
correlate with having the `best' focal spot, since the highest quality vacuum
focal spot produces a greatly inferior electron beam, but instead correspond to
the particular laser phase that steers the plasma wave to a final state with
optimal accelerating fields.
Body: \maketitle \begin{affiliations} \item{Center for Ultrafast Optical Science, University of Michigan, Ann Arbor, MI 48109-2099 USA} \item{Polytech Paris-Sud - Universit\'e Paris-Sud, 91405 Orsay, France} \end{affiliations} \begin{abstract} Coherent control of a system involves steering an interaction to a final coherent state by controlling the phase of an applied field. Plasmas support coherent wave structures that can be generated by intense laser fields. Here, we demonstrate the coherent control of plasma dynamics in a laser wakefield electron acceleration experiment. A genetic algorithm is implemented using a deformable mirror with the electron beam signal as feedback, which allows a heuristic search for the optimal wavefront under laser-plasma conditions that is not known \emph{a priori}. We are able to improve both the electron beam charge and angular distribution by an order of magnitude. These improvements do not simply correlate with having the `best' focal spot, since the highest quality vacuum focal spot produces a greatly inferior electron beam, but instead correspond to the particular laser phase front that steers the plasma wave to a final state with optimal accelerating fields. \end{abstract} \section{Introduction} The concept of coherent control --- precise measurement or determination of a process through control of the phase of an applied oscillating field --- has been applied to many different systems, including quantum dynamics, trapped atomic ions, chemical reactions, Cooper pairs, quantum dots and THz generation to name but a few. A plasma wave is a coherent and deterministically evolving structure that can be generated by the interaction of laser light with plasma. It is therefore natural to assume that coherent control techniques may also be applied to plasma waves. Plasma waves produced by high power lasers have been studied intensively for their numerous applications, such as the production of ultrashort pulses by plasma wave compression, generation of extremely high power pulses by Raman amplification, for inertial confinement fusion ignition schemes, as well as for fundamental scientific investigations. In particular, laser wakefield acceleration of ultra-relativistic electron beams, has been a successful method for accelerating electrons to relativistic energies over a very short distance. In laser wakefield acceleration, an electron bunch `surfs' on the electron plasma wave generated by an intense laser and gains a large amount of energy. The accelerating electric field strength that the plasma wave can support can be many orders of magnitude higher than that of a conventional accelerator, which makes laser wakefield acceleration an exciting prospect as an advanced accelerator concept. However, although highly competitive in terms of accelerating gradient, beams from laser wakefield accelerator experiments are currently inferior to conventional accelerators in terms of other important characteristics, such as energy spread and stability. In addition, due to constraints in laser wakefield technology, experimental demonstrations have predominantly been performed in single shot operation, far below the kHz-MHz repetition rates of conventional accelerators. In recent years, deformable mirror adaptive optical systems have been successfully implemented in high intensity laser experiments to increase the peak laser intensity by improving the beam focusability, especially in systems using high numerical aperture optics. The shape of the deformable mirror is generally determined in a closed loop where either a direct measurement of the wavefront is performed or some nonlinear optical signal is used as feedback in an iterative algorithm. The objective of adaptive optics has largely been optimization of the laser focal shape to a near diffraction-limited spot, thus producing the highest possible intensity. Adaptive optics can also be useful for certain focal profile shaping, optimization of a laser machining process or harmonic generation. In the following, we demonstrate that orders of magnitude improvement to electron beam properties from a laser wakefield accelerator operating at kHz repetition rate can be made, through the use of a genetic algorithm coupled to a deformable mirror adaptive optical system to coherently control the plasma wave formation. The electron image from a scintillator screen was processed and used in the fitness function as feedback for the genetic algorithm. Using this method, we were able to improve the beam properties significantly. This result was not simply due to an improvement in focal quality since a laser pulse with the `best' (highest intensity/lowest $M^2$) focus in vacuum produced a greatly inferior electron beam compared with a laser pulse optimized using the electron beam properties themselves. It was found that the focal spot optimized for electron beam production had pronounced intensity `wings'. Modifications to the phase front of the tightly focusing laser alter the light propagation, which experiences strong optical nonlinearities in the plasma, and therefore affect the plasma wave dynamics in a complex but deterministic manner. \section{Results} \subsection{Experimental setup and procedure} The experiment was performed using the relativistic Lambda-Cubed ($\lambda^3$) laser system (see Methods). The output laser beam was reflected from a deformable mirror and focused onto a free-flowing argon gas plume to produce an electron beam by laser wakefield acceleration (see Methods) at 500 Hz. Electrons were measured using a scintillating screen imaged onto a lens-coupled CCD camera. The experimental setup is shown schematically in Fig.~. We first implemented a genetic algorithm for laser focus optimization using the second-harmonic signal generated from a beta barium borate ($\beta$-BBO) crystal (setup A in Fig.~). The laser spot was optimized such that highest peak intensity is achieved when the second harmonic generation is strongest. Subsequently, we modified the fitness function to use a figure of merit (FOM, refer to equation in Methods) from the electron scintillation data, calculating the inverse distance weighting (with power parameter n) to a single point $\mathbf{r_0}$ for all pixel intensities within an electron image. The pixel of the optimization point $\mathbf{r_0}$ was \emph{dynamically} adjusted during the genetic algorithm to concentrate all electron signal to the peak location of the charge distribution during each generation. The genetic algorithm was initialized using a `flat' mirror shape with 30 V for all actuators to allow immediate deformation in both directions. \subsection{Optimization of the electron spatial profile} For comparison, electron beams produced by the `best' laser focus (by optimizing the intensity) and the initial mirror shape at 30 V are shown in Fig.~a and b respectively. The optimized electron beam profiles are shown in Fig.~e-j for various weighting parameters, $n$. The genetic algorithm converged to the best electron beam using $n=8$ in terms of beam divergence and peak charge density. The peak charge density was increased by a factor of 20 compared to the initial electron beam profile before optimization (see Fig.~d). The optimized electron profile is highly stable and collimated, with a full-width-at-half-maximum (FWHM) divergence of $\Delta x=7.4\pm 0.6$ mrad and $\Delta y=12.8 \pm 1.4$ mrad. The shot-to-shot pointing (defined by the centroid position) fluctuation of the electron beam is less than 1 mrad (root mean square, r.m.s.). The integrated charge was increased by more than two-fold from the electron beams generated by a laser focus of highest intensity. The high repetition rate and real-time diagnostics permit implementation of the algorithm within a practical time frame using a standard personal computer. Typical optimization takes only a few minutes ($\sim$40 iterations) to reach convergence (see Fig.~c). The second harmonic optimization generates a near-diffraction-limited focal spot as shown in Fig.~a for the far-field laser intensity profile in vacuum. In Figs.~a and b, we compare the transverse intensity distribution within the focal region over the length scale of the gas jet. The laser profile (Fig.~b) that produces the best electron beam exhibits several low intensity side lobes around the central peak, and has a peak intensity about half that of the optimized focus. The complex laser profiles appear to have a very dramatic effect on the structure of the plasma waves produced and consequently the electron beam profile. Fig.~c shows the relative wavefront change recorded by a Shack-Hartmann wavefront sensor. Calculation using the reconstructed wavefront gives a Strehl ratio of about 0.5, which is in agreement with the far-field intensity measurement. This small wavefront modification of the driver pulse can lead to a significant improvement in the electron beam properties through the relativistic nonlinear optics of the plasma. The relative position between the focal plane and the center of the gas flow was controlled by moving the nozzle. Scanning the gas nozzle both before and after genetic algorithm confirms the optimal focal position does not change, excluding the possibility that the improvement may be due to optimizing the focal position. \subsection{Control of energy distribution} Furthermore, we extended the genetic algorithm optimization to \emph{control the electron energy distribution}. Through control of the light propagation, the plasma wave amplitude will be affected and therefore also the strength of the accelerating gradient. Hence, we can expect to be able to modify the energy spectrum. A high resolution energy analyzer using a dipole magnet pair was used to obtain the electron energy spectrum as the electrons were dispersed in the horizontal plane in the magnetic field. A 150 $\mu$m pinhole was placed 2.2 cm from the electron source to improve the energy resolution of the spectrometer. The schematic setup is shown in Fig.~a. The energy resolution limited by the entrance pinhole and transverse emittance of the beam is estimated to be 2 keV for the energy range of measurement. Three rectangular masks are set in the low-, mid- and high-energy region on the dispersed data, namely masks I, II, and III in Fig.~b. We employed a fitness function (see Methods) to preferentially maximize the total counts inside the mask. Raw spectra from the genetic algorithm optimization are displayed in Fig.~b, showing that the brightest part has shifted congruently. The resulting spectra have mean energies of 89 keV, 95 keV and 98 keV respectively for masks I, II, and III, noting that they do not fall on the visual centroid of the image because the scintillator sensitivity is not included in the presentation of the raw data, however it was taken into account for computing the mean energies. Our results show that manipulation of electron energy distribution using the deformable mirror is somewhat restricted. The final result after optimization does not reach the objective mask completely despite that the mean energies can be varied by up to 10\ \subsection{Numerical simulations} Although the details of the initial conditions required for optimal beams are difficult to determine and are therefore found using the genetic algorithm, we can at least demonstrate how modifications to the phase front of the laser pulse can improve the beam properties with an example. To illustrate the underlying physics of the plasma wave dynamics determined by the conditions of the driving laser pulse, we performed two dimensional (2D) particle-in-cell (PIC) simulations using the OSIRIS framework. Parameters similar to the experiment conditions were used, with a Gaussian plasma density profile to enable trapping of electrons in the density down ramp (see ref.~\citen{He_NJP_2013} and Methods for details on the simulation). It was previously shown in ref.~\citen{Cor_PRSTAB_2011} that the focusing fields of laser plasma accelerators can be controlled by tailoring the transverse intensity profile of the laser pulse using higher-order modes, where generalization to 3D was also discussed. Here, we simulated a laser pulse with a fundamental Gaussian mode (TEM$_{00}$) or a coherent superposition of a fundamental (TEM$_{00}$) and a second-order Hermite-Gaussian (TEM$_{02}$) mode (Fig.~a). Although the plasma wave has a larger amplitude when it is driven by a single mode laser pulse, the wake phase front evolves a backward curvature when electrons are trapped and accelerated (see top panel in Supplementary Movie 1 and Fig.~b). Contrastingly, the evolution of the wakefield driven by the laser pulse with additional mode forms a flatter plasma phase front \emph{at the point of trapping} (Fig.~d). In Fig.~c and e, the momentum distribution of the forward accelerated electrons has a larger transverse spread for the single mode laser pulse compared to the one with the addition of higher order modes. This is a consequence of the different trapping conditions and accelerating fields from the coherent plasma wakefield structure, which is governed by the structure of the driving laser pulse. In a comparative test to show this effect is not simply due to a lower intensity, we repeated the simulation using a single fundamental Gaussian mode laser pulse with a larger focal spot with the same peak intensity as that with the superimposed modes. The wakefield evolution shows very similar response as Fig.~b and does not develop a flatter phase front as seen in Fig.~d. The subsequently accelerated electrons have very similar divergence to that in Fig.~c, eliminating the possibility that the improvement comes from either a high intensity effect or a simple change in $f$-number. \emph{Note that we are not saying that this mixed-Hermite-Gaussian mode is the optimal pulse in the experiment; this is simply an illustration of how small changes in pulse shape can have significant effects on electron beam properties.} \section{Discussion} When a particular wavefront of laser light interacts with plasma, it can affect the plasma wave structures and trapping conditions of the electrons in a complex way. For example, Raman forward scattering, envelope self-modulation, relativistic self-focusing, and relativistic self-phase modulation and many other nonlinear interactions modify both the pulse envelope and phase as the pulse propagates, in a way that cannot be easily predicted and that subsequently dictates the formation of plasma waves. Moreover, under realistic experimental conditions, ionization dynamics before the laser pulse reaches the vacuum focus can also modify the phase of the driving pulse. Ideally, the light interacts in such a way as to generate large amplitude plasma waves with electric field structures that accelerate electrons with small divergence, high charge etc. Because of the complicated interaction, it is difficult to determine a laser phase profile that will lead to such a plasma structure. However, such unforeseeable conditions were successfully revealed by \emph{using the evolutionary algorithm method}, with the result that the electron charge can be increased and emitted in a very well collimated beam. Here we have implemented coherent control of a nonlinear plasma wave and demonstrated an order of magnitude improvement in the electron beam parameters. The laser beam optimized to generate the best electron beam was not the one with the `best' focal spot. Control and shaping of the electron energy distribution was observed to be less effective, but was still possible. The capability for wavefront control was also limited by the number of actuators and maximum deformation of the deformable mirror used in our experiments. In addition, this work was performed using adaptive optics, but it is clear that coherent control of plasma waves should be possible in a variety of configurations, for example by using an acousto-optic modulator to control the temporal phase of the driving pulse. Recently developed techniques for single-shot diagnosis of plasma wave structures may provide an avenue for direct control of the plasma evolution. The concept of coherent control for plasmas opens new possibilities for future laser-based accelerators. Although still at the stage of fundamental research, laser wakefield accelerators are showing significant promise. In principle, such improvements could be integrated into next generation high-power laser projects, such as ICAN, based on coherent combination of many independent fibers, taking advantage of both their high repetition rate and controllability. The stability and response of the wakefield to laser conditions, such as phase front errors, is not well understood, but is crucial for the success of laser wakefield acceleration as a source of relativistic electrons and secondary radiation. For example, the presence of an asymmetric laser pulse was shown to affect the betatron oscillations and properties of x-rays produced in laser wakefield accelerators. Implementing the methods of this study should enable a significantly improved understanding and control of the wakefield acceleration process with regard to stability, dark current reduction and beam emittance. \begin{methods} \subsection{Laser System} The Relativistic Lambda Cubed laser ($\lambda^3$) produces 30 fs pulses of 800 nm light at a repetition rate of 500 Hz with an ASE (Amplified-Spontaneous-Emission) intensity contrast of $\sim10^8$ around 1 ns before the main pulse. The system is seeded by a Femto-Laser Ti:sapphire oscillator, which generates 12 fs pulses and has a companion carrier envelope phase locking system. An RF addressable acousto-optic filter called a Dazzler controls the spectral amplitude and phase of these pulses. Selected pulses from the Dazzler train are stretched to 220 ps in a low-aberration stretcher and amplified to 7 mJ in a cryogenically cooled large-mode regenerative amplifier (Regen). The energy dumped from the Regen cavity is `cleaned' in a Pockels cell and used to seed a 3-pass amplifier as an upgrade from the laser system described in ref.~\citen{Hou:08}, which delivers up to 28 mJ pulses before compression. Following 71\ \subsection{Electron Acceleration and Detection} The focused laser pulse drives plasma waves by interacting with an argon gas jet flowing continuously from a 100 $\mu$m inner diameter fused silica capillary. Typically the laser axis is 300 $\mu$m above the orifice of the tubing. The laser pulses were focused by an $f/2$ off-axis parabolic mirror to a spot size of 2.5 $\mu$m FWHM with a maximum of 10 mJ energy on target. The plasma electron density is measured to be in the range (0.5-2)$\times 10^{19}$ cm$^{-3}$ using transverse interferometry. Electrons are accelerated in the density down ramp with a final energy in the 100 keV range and detected by a high resolution scintillating screen (J6677 FOS by Hamamatsu), which is placed about 35 cm downstream from the source and imaged with a lens coupled 12-bit CCD camera for a 4$\times$4 cm effective area. The scintillator sensitivity was calibrated using an electron microscope for the energy range in the spectrum measurement. Electron beam charge was estimated using the calibrated scintillator response, manufacturer-provided information for the CCD camera (gain, quantum efficiency etc.) and the measured effective numerical aperture of the imaging system. \subsection{Focal characterization and wavefront measurement} The amplified laser beam was attenuated by using a half-wave plate and the polarization dependent properties of the compressor grating of the laser system. \O 25 mm neutral density filters (Thorlabs, Inc.) were inserted in the exit beam after the compressor and before a telescope beam expander. The laser focus was imaged by a 60$\times$ microscope objective lens (Newport Corporation, M-60X) onto a 8-bit CCD camera for focal characterization (cf setup A in Fig.~). A Shack-Hartmann wavefront sensor (Flexible Optical BV) was used to determine the relative wavefront change between different deformable mirror configurations. The sensor, which consists of a $30\times 30$ microlens array having a focal length of 3.5~mm and 150~$\mu$m pitch, was directly placed in the path of the converging beam after the focusing parabolic mirror. Reconstruction of the wavefront was performed using the FrontSurfer analysis software (Flexible Optical BV), typically with $\approx450$ measured local wavefront slopes and RMS error on the order of 0.05$\lambda$. Rotating the neutral density filters and the half-wave plate did not change the focal spot or the wavefront measurement significantly, insuring the wavefront distortion introduced by attenuation was negligible. \subsection{Deformable Mirror and Genetic Algorithm} The deformable mirror (AOA Xinetics) has a 47-mm clear aperture of a continuous face sheet with 37 piezoelectric actuators arranged on a square grid spaced 7 mm apart. The maximum stroke used in this experiment is about 2 $\mu$m. The mirror shape is controlled by a genetic algorithm, which is a method mimicking the process of natural selection and routinely used to generate optimal solutions in complex systems with a large number of variables. The \emph{genetic representation} in our experiments comprises a set of 37 independent voltage values for the deformable mirror actuator array. A \emph{fitness function} is designed to produce a single \emph{figure of merit} (FOM) to evaluate how close the solution is to the goal. In the electron beam profile optimization experiment, FOM is computed as follows: \begin{equation} FOM=\sum_{\substack{{(i,j)}\\\mathbf{r_{ij}}\neq \mathbf{r_0}}}\frac{I_{ij}}{|\mathbf{r_{ij}}-\mathbf{r_0}|^n} \end{equation} where $I_{ij}$ is the pixel intensity for every pixel $(i,j)$ in the whole image and $\mathbf{r_0}$ is a coordinate point in the image used as an optimization target. The power factor $n>0$ gives higher weighting to those pixels closer to the target (inverse distance weighting). In the experiment to control the energy spectrum, FOM is calculated using the following formula given a pre-defined image mask, \begin{equation} {\rm FOM}=\left(1-\frac{{\rm mean\; intensity\;}{\rm outside\; mask}}{{\rm mean\;intensity\;}{\rm of \;whole\; image}}\right) \times {\rm mean\; intensity\;inside\;mask} \end{equation} The mean intensity is the sum of the pixel counts divided by the number of pixels for a defined region. A rectangular mask was used in the experiment as specified by the region enclosed by the red dashed lines in Fig.~b. \subsection{Numerical simulations} The 2D PIC simulations were performed in a stationary box of the dimensions $573\times 102$ $\mu$m with $10000\times 600$ cells and 4$\times$4 particles-per-cell. A Gaussian plasma density profile was used in the propagation dimension ($x_1$), peaked at $x_1=200~\mu$m with a full-width-at-half-maximum of 120 $\mu$m and a maximum electron density of 0.005$n_c$, where $n_c$ is the plasma critical density. The laser pulse was initialized at the left edge of the simulation window and focused at 215 $\mu$m in the density down ramp. In 2D geometry, the transverse intensity profile of the laser pulse for fundamental Gaussian mode (TEM$_{00}$) has the form $a_0^2\exp(-2x_2^2/w_0^2)$, and the second-order Hermite-Gaussian mode $a_2^2\exp(-2x_2^2/w_2^2)(8x_2^2/w_2^2-2)^2$, where $a_{0,2}$ is the normalized vector potential and $w_{0,2}$ is the beam waist parameter. The two modes are coherently superimposed in the same plane of polarization. Here we used even-order Hermite-Gaussian mode (TEM$_{02}$) for its symmetric property. A phase difference of $\pi/8$ was applied at the beam waist between the two modes to simulate variations in the optical phase front condition. The beam waist was positioned to account for focal shift as a result of coherent superposition of two modes such that the location of the maximum on-axis laser intensity was the same (at $x_1=215~\mu$m) for all simulation runs. The simulation parameters are $a_0=1.08$ and $w_0=3.31$ $\mu$m for the Gaussian mode alone, or $a_0=1.0$, $a_2=0.15$ and $w_0=w_2=3.31$ $\mu$m for the superimposed mode. \end{methods} \begin{thebibliography}{10} \expandafter\ifx\csname url\endcsname\relax \def\url#1{\texttt{#1}}\fi \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi \providecommand{\bibinfo}[2]{#2} \providecommand{\eprint}[2][]{} \bibitem{Warren_Sci_1993} \bibinfo{author}{Warren, W.~S.}, \bibinfo{author}{Rabitz, H.} \& \bibinfo{author}{Dahleh, M.} \newblock \bibinfo{title}{Coherent control of quantum dynamics: The dream is alive}. \newblock \emph{\bibinfo{journal}{Science}} \textbf{\bibinfo{volume}{259}}, \bibinfo{pages}{1581--1589} (\bibinfo{year}{1993}). \bibitem{Wineland_JRNIST_1998} \bibinfo{author}{Wineland, D.} \emph{et~al.} \newblock \bibinfo{title}{Experimental issues in coherent quantum-state manipulation of trapped atomic ions}. \newblock \emph{\bibinfo{journal}{J. Res. Natl. Inst. Stand. Technol.}} \textbf{\bibinfo{volume}{103}}, \bibinfo{pages}{259} (\bibinfo{year}{1998}). \bibitem{Assion_Sci_1998} \bibinfo{author}{Assion, A.} \emph{et~al.} \newblock \bibinfo{title}{Control of chemical reactions by feedback-optimized phase-shaped femtosecond laser pulses}. \newblock \emph{\bibinfo{journal}{Science}} \textbf{\bibinfo{volume}{282}}, \bibinfo{pages}{919--922} (\bibinfo{year}{1998}). \bibitem{Nakamura_Nat_1999} \bibinfo{author}{Nakamura, Y.}, \bibinfo{author}{Pashkin, Y.} \& \bibinfo{author}{Tsai, J.} \newblock \bibinfo{title}{Coherent control of macroscopic quantum states in a single-cooper-pair box}. \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{398}}, \bibinfo{pages}{786--788} (\bibinfo{year}{1999}). \bibitem{Reithmaier_Nat_2004} \bibinfo{author}{Reithmaier, J.} \emph{et~al.} \newblock \bibinfo{title}{Strong coupling in a single quantum dot-semiconductor microcavity system}. \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{432}}, \bibinfo{pages}{197--200} (\bibinfo{year}{2004}). \bibitem{Press_Nat_2008} \bibinfo{author}{Press, D.}, \bibinfo{author}{Ladd, T.~D.}, \bibinfo{author}{Zhang, B.} \& \bibinfo{author}{Yamamoto, Y.} \newblock \bibinfo{title}{Complete quantum control of a single quantum dot spin using ultrafast optical pulses}. \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{456}}, \bibinfo{pages}{218--221} (\bibinfo{year}{2008}). \bibitem{Xie_PRL_2006} \bibinfo{author}{Xie, X.}, \bibinfo{author}{Dai, J.} \& \bibinfo{author}{Zhang, X.-C.} \newblock \bibinfo{title}{Coherent control of THz wave generation in ambient air}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{96}}, \bibinfo{pages}{075005} (\bibinfo{year}{2006}). \bibitem{Strickland_1985} \bibinfo{author}{Strickland, D.} \& \bibinfo{author}{Mourou, G.} \newblock \bibinfo{title}{Compression of amplified chirped optical pulses}. \newblock \emph{\bibinfo{journal}{Optics Communications}} \textbf{\bibinfo{volume}{56}}, \bibinfo{pages}{219 -- 221} (\bibinfo{year}{1985}). \bibitem{Faure_PRL_2005} \bibinfo{author}{Faure, J.} \emph{et~al.} \newblock \bibinfo{title}{Observation of laser-pulse shortening in nonlinear plasma waves}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{95}}, \bibinfo{pages}{205003} (\bibinfo{year}{2005}). \bibitem{Schreiber_PRL_2010} \bibinfo{author}{Schreiber, J.} \emph{et~al.} \newblock \bibinfo{title}{Complete temporal characterization of asymmetric pulse compression in a laser wakefield}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{105}}, \bibinfo{pages}{235003} (\bibinfo{year}{2010}). \bibitem{Malkin_PRL_1999} \bibinfo{author}{Malkin, V.~M.}, \bibinfo{author}{Shvets, G.} \& \bibinfo{author}{Fisch, N.~J.} \newblock \bibinfo{title}{Fast compression of laser beams to highly overcritical powers}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{82}}, \bibinfo{pages}{4448--4451} (\bibinfo{year}{1999}). \bibitem{Trines_NP_2011} \bibinfo{author}{Trines, R. M. G.~M.} \emph{et~al.} \newblock \bibinfo{title}{Simulations of efficient Raman amplification into the multipetawatt regime}. \newblock \emph{\bibinfo{journal}{Nat. Phys.}} \textbf{\bibinfo{volume}{7}}, \bibinfo{pages}{87--92} (\bibinfo{year}{2011}). \bibitem{Fisch_PRL_2002} \bibinfo{author}{Malkin, V.~M.} \& \bibinfo{author}{Fisch, N.~J.} \newblock \bibinfo{title}{Collective deceleration of relativistic electrons precisely in the core of an inertial-fusion target}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{89}}, \bibinfo{pages}{125004} (\bibinfo{year}{2002}). \bibitem{Willingale_PRL_2011} \bibinfo{author}{Willingale, L.} \emph{et~al.} \newblock \bibinfo{title}{High-power, kilojoule class laser channeling in millimeter-scale underdense plasma}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{106}}, \bibinfo{pages}{105002} (\bibinfo{year}{2011}). \bibitem{Tajima1979} \bibinfo{author}{Tajima, T.} \& \bibinfo{author}{Dawson, J.~M.} \newblock \bibinfo{title}{Laser electron accelerator}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{43}}, \bibinfo{pages}{267--270} (\bibinfo{year}{1979}). \bibitem{Mangles_Nat_2004} \bibinfo{author}{Mangles, S. P.~D.} \emph{et~al.} \newblock \bibinfo{title}{{Monoenergetic beams of relativistic electrons from intense laser-plasma interactions.}} \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{431}}, \bibinfo{pages}{535--538} (\bibinfo{year}{2004}). \bibitem{Geddes_Nat_2004} \bibinfo{author}{Geddes, C.} \emph{et~al.} \newblock \bibinfo{title}{{High-quality electron beams from a laser wakefield accelerator using plasma-channel guiding}}. \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{431}}, \bibinfo{pages}{538--541} (\bibinfo{year}{2004}). \bibitem{Faure_Nat_2004} \bibinfo{author}{Faure, J.} \emph{et~al.} \newblock \bibinfo{title}{{A laser--plasma accelerator producing monoenergetic electron beams}}. \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{431}}, \bibinfo{pages}{541--544} (\bibinfo{year}{2004}). \bibitem{Bahk_OL_2004} \bibinfo{author}{Bahk, S.-W.} \emph{et~al.} \newblock \bibinfo{title}{Generation and characterization of the highest laser intensities (10$^{22}$ W/cm$^2$)}. \newblock \emph{\bibinfo{journal}{Opt. Lett.}} \textbf{\bibinfo{volume}{29}}, \bibinfo{pages}{2837--2839} (\bibinfo{year}{2004}). \bibitem{Albert_OL_2000} \bibinfo{author}{Albert, O.}, \bibinfo{author}{Wang, H.}, \bibinfo{author}{Liu, D.}, \bibinfo{author}{Chang, Z.} \& \bibinfo{author}{Mourou, G.} \newblock \bibinfo{title}{Generation of relativistic intensity pulses at a kilohertz repetition rate}. \newblock \emph{\bibinfo{journal}{Opt. Lett.}} \textbf{\bibinfo{volume}{25}}, \bibinfo{pages}{1125--1127} (\bibinfo{year}{2000}). \bibitem{Planchon_OL_2006} \bibinfo{author}{Planchon, T.~A.} \emph{et~al.} \newblock \bibinfo{title}{Adaptive correction of a tightly focused, high-intensity laser beam by use of a third-harmonic signal generated at an interface}. \newblock \emph{\bibinfo{journal}{Opt. Lett.}} \textbf{\bibinfo{volume}{31}}, \bibinfo{pages}{2214--2216} (\bibinfo{year}{2006}). \bibitem{Zeng_JMO_1999} \bibinfo{author}{Zeng, Z.}, \bibinfo{author}{Ling, N.} \& \bibinfo{author}{Jiang, W.} \newblock \bibinfo{title}{The investigation of controlling laser focal profile by deformable mirror and wave-front sensor}. \newblock \emph{\bibinfo{journal}{Journal of Modern Optics}} \textbf{\bibinfo{volume}{46}}, \bibinfo{pages}{341--348} (\bibinfo{year}{1999}). \bibitem{Agmy_OE_2005} \bibinfo{author}{El-Agmy, R.}, \bibinfo{author}{Bulte, H.}, \bibinfo{author}{Greenaway, A.~H.} \& \bibinfo{author}{Reid, D.} \newblock \bibinfo{title}{Adaptive beam profile control using a simulated annealing algorithm}. \newblock \emph{\bibinfo{journal}{Opt. Express}} \textbf{\bibinfo{volume}{13}}, \bibinfo{pages}{6085--6091} (\bibinfo{year}{2005}). \bibitem{Campbell_JO_2007} \bibinfo{author}{Campbell, S.}, \bibinfo{author}{Triphan, S. M.~F.}, \bibinfo{author}{El-Agmy, R.}, \bibinfo{author}{Greenaway, A.~H.} \& \bibinfo{author}{Reid, D.~T.} \newblock \bibinfo{title}{Direct optimization of femtosecond laser ablation using adaptive wavefront shaping}. \newblock \emph{\bibinfo{journal}{Journal of Optics A: Pure and Applied Optics}} \textbf{\bibinfo{volume}{9}}, \bibinfo{pages}{1100} (\bibinfo{year}{2007}). \bibitem{Bartels_Nat_2000} \bibinfo{author}{Bartels, R.} \emph{et~al.} \newblock \bibinfo{title}{{Shaped-pulse optimization of coherent emission of high-harmonic soft X-rays}}. \newblock \emph{\bibinfo{journal}{Nature}} \textbf{\bibinfo{volume}{406}}, \bibinfo{pages}{164--166} (\bibinfo{year}{2000}). \bibitem{Yoshitomi_APB} \bibinfo{author}{Yoshitomi, D.} \emph{et~al.} \newblock \bibinfo{title}{Phase-matched enhancements of high-harmonic soft x-rays by adaptive wave-front control with a genetic algorithm}. \newblock \emph{\bibinfo{journal}{Applied Physics B}} \textbf{\bibinfo{volume}{78}}, \bibinfo{pages}{275--280} (\bibinfo{year}{2004}). \bibitem{osiris} \bibinfo{author}{Fonseca, R.} \emph{et~al.} \newblock \bibinfo{title}{Osiris: A three-dimensional, fully relativistic particle in cell code for modeling plasma based accelerators}. \newblock In \bibinfo{editor}{Sloot, P.}, \bibinfo{editor}{Hoekstra, A.}, \bibinfo{editor}{Tan, C.} \& \bibinfo{editor}{Dongarra, J.} (eds.) \emph{\bibinfo{booktitle}{Computational Science -- ICCS 2002}}, vol. \bibinfo{volume}{2331} of \emph{\bibinfo{series}{Lecture Notes in Computer Science}}, \bibinfo{pages}{342--351} (\bibinfo{publisher}{Springer Berlin Heidelberg}, \bibinfo{year}{2002}). \bibitem{He_NJP_2013} \bibinfo{author}{He, Z.-H.} \emph{et~al.} \newblock \bibinfo{title}{High repetition-rate wakefield electron source generated by few-millijoule, 30~fs laser pulses on a density downramp}. \newblock \emph{\bibinfo{journal}{New Journal of Physics}} \textbf{\bibinfo{volume}{15}}, \bibinfo{pages}{053016} (\bibinfo{year}{2013}). \bibitem{Cor_PRSTAB_2011} \bibinfo{author}{Cormier-Michel, E.} \emph{et~al.} \newblock \bibinfo{title}{Control of focusing fields in laser-plasma accelerators using higher-order modes}. \newblock \emph{\bibinfo{journal}{Phys. Rev. ST Accel. Beams}} \textbf{\bibinfo{volume}{14}}, \bibinfo{pages}{031303} (\bibinfo{year}{2011}). \bibitem{Mori_JQE_1997} \bibinfo{author}{Mori, W.} \newblock \bibinfo{title}{The physics of the nonlinear optics of plasmas at relativistic intensities for short-pulse lasers}. \newblock \emph{\bibinfo{journal}{Quantum Electronics, IEEE Journal of}} \textbf{\bibinfo{volume}{33}}, \bibinfo{pages}{1942--1953} (\bibinfo{year}{1997}). \bibitem{Li_PRL_2014} \bibinfo{author}{Li, Z.} \emph{et~al.} \newblock \bibinfo{title}{Single-shot visualization of evolving laser wakefields using an all-optical streak camera}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{113}}, \bibinfo{pages}{085001} (\bibinfo{year}{2014}). \bibitem{Savert_2014} \bibinfo{author}{{S{\"a}vert}, A.} \emph{et~al.} \newblock \bibinfo{title}{{Direct imaging of the dynamics of a laser-plasma accelerator operating in the bubble-regime}}. \newblock \emph{\bibinfo{journal}{e-Print arXiv:1402.3052}} (\bibinfo{year}{2014}). \bibitem{ICAN} \bibinfo{author}{Mourou, G.}, \bibinfo{author}{Brocklesby, B.}, \bibinfo{author}{Tajima, T.} \& \bibinfo{author}{Limpert, J.} \newblock \bibinfo{title}{The future is fibre accelerators}. \newblock \emph{\bibinfo{journal}{Nat Photon}} \textbf{\bibinfo{volume}{7}}, \bibinfo{pages}{258--261} (\bibinfo{year}{2013}). \bibitem{Glinec_EPL_08} \bibinfo{author}{Glinec, Y.} \emph{et~al.} \newblock \bibinfo{title}{Direct observation of betatron oscillations in a laser-plasma electron accelerator}. \newblock \emph{\bibinfo{journal}{EPL (Europhysics Letters)}} \textbf{\bibinfo{volume}{81}}, \bibinfo{pages}{64001} (\bibinfo{year}{2008}). \bibitem{Mangles_APL_2009} \bibinfo{author}{Mangles, S. P.~D.} \emph{et~al.} \newblock \bibinfo{title}{Controlling the spectrum of x-rays generated in a laser-plasma accelerator by tailoring the laser wavefront}. \newblock \emph{\bibinfo{journal}{Applied Physics Letters}} \textbf{\bibinfo{volume}{95}}, \bibinfo{pages}{181106} (\bibinfo{year}{2009}). \bibitem{Popp_PRL_2010} \bibinfo{author}{Popp, A.} \emph{et~al.} \newblock \bibinfo{title}{All-optical steering of laser-wakefield-accelerated electron beams}. \newblock \emph{\bibinfo{journal}{Phys. Rev. Lett.}} \textbf{\bibinfo{volume}{105}}, \bibinfo{pages}{215001} (\bibinfo{year}{2010}). \bibitem{Schnell_NatComm_2013} \bibinfo{author}{Schnell, M.} \emph{et~al.} \newblock \bibinfo{title}{Optical control of hard X-ray polarization by electron injection in a laser wakefield accelerator}. \newblock \emph{\bibinfo{journal}{Nat. Commun.}} \textbf{\bibinfo{volume}{4}}, \bibinfo{pages}{2421} (\bibinfo{year}{2013}). \bibitem{Hou:08} \bibinfo{author}{Hou, B.}, \bibinfo{author}{Easter, J.}, \bibinfo{author}{Mordovanakis, A.}, \bibinfo{author}{Krushelnick, K.} \& \bibinfo{author}{Nees, J.~A.} \newblock \bibinfo{title}{Vacuum-free x-ray source based on ultrashort laser irradiation of solids}. \newblock \emph{\bibinfo{journal}{Opt. Express}} \textbf{\bibinfo{volume}{16}}, \bibinfo{pages}{17695--17705} (\bibinfo{year}{2008}). \end{thebibliography} \begin{addendum} \item This work was supported by DARPA (Contract No. N66001-11-1-4208), the NSF (Grant No. 0935197), NSF CAREER (Grant No 1054164), the AFOSR Young Investigator Program (Grant No. FA9550-12-1-0310) and MCubed at the University of Michigan. The authors acknowledge the OSIRIS Consortium, consisting of University of California, Los Angeles (UCLA) and Instituto Superior T\'ecnico (IST) (Lisbon, Portugal), for the use of the \textsc{osiris} 2.0 framework. This research was supported in part through computational resources and services provided by Advanced Research Computing at the University of Michigan, Ann Arbor. \item [Author contributions] Z.H, B.H., J.A.N. designed and carried out the experiment. Z.H. performed all data analysis and the simulations. V.L. and B.H. developed the LabView program. K.M.K and A.G.R.T. directed and guided the project. Z.H. and A.G.R.T. wrote the paper. All authors discussed the results and contributed to the manuscript. \item[Competing Interests] The authors declare that they have no competing financial interests. \item[Correspondence] Correspondence and requests for materials should be addressed to Z.-H. He. (email: zhhe@umich.edu). \end{addendum} |
1501.04120 | Title: Project of the underwater system for chemical threat detection
Abstract: In this article we describe a novel method for the detection of explosives
and other hazardous substances in the marine environment using neutron
activation. Unlike the other considered methods based on this technique we
propose to use guides forneutron and gamma quanta which speeds up and
simplifies identification. Moreover, it may provide a determination of the
density distribution of a dangerous substance. First preliminary results of
Monte Carlo simulations dedicated for design of a device exploiting this method
are also presented.
Body: \title{Project of the underwater system for chemical threat detection} \author{M.~Silarski$^{a}$, D.~Hunik$^{a}$, P.~Moskal$^{a}$, M.~Smolis$^{a}$, S.~Tadeja$^{a}$} \affil{$^{a}$Institute of Physics, Jagiellonian University \\ {\L}ojasiewicza 11, 30-348 Krak\'ow, Poland} \maketitle {PACS: 28.41.Ak, 89.20.Dd, 28.20.Cz, 28.20.Fc} \begin{abstract} In this article we describe a novel method for the detection of explosives and other hazardous substances in the marine environment using neutron activation. Unlike the other considered methods based on this technique we propose to use guides for neutron and gamma quanta which speeds up and simplifies identification. Moreover, it may provide a determination of the density distribution of a dangerous substance. First preliminary results of Monte Carlo simulations dedicated for design of a device exploiting this method are also presented. \end{abstract} \section{Introduction} During both World Wars the Baltic Sea was a scene of intense naval warfare and became a ``container'' for almost whole Nazi chemical munition arsenal. In accordance with the provisions of the Potsdam Conference, Germany was demilitarized but only a small part of their arsenal was neutralized on land. Technologies for safe and effective disposal of chemical weapons were not known shortly after the war. Thus, till 1948 about 250000 tons of munition, including up to 65000 tons of chemical agents, were sunk in Baltic Sea waters by both Germans and Allies. The main known contaminated areas are Little Belt, Bornholm Deep (east of Bornholm) and the south-western part of the Gotland Deep~. Apart from known underwater stockyard there is unknown amount of dangerous war remnants spread over the whole Baltic, especially along maritime convoys paths and in vicinity of coasts. It is not clear how dangerous are these underwater arsenals. At the bottom of the sea (in approx. 5-7$^o$C) the chemical agents take a form of oily liquids hardly soluble in water. Thus, the sunk ammunition does not release hazardous substances. It becomes dangerous however, if the rusted tanks and shells are raised from the bottom of the sea. Chemical munitions, containing mostly mustard gas, was fished several times by fishermen on the Baltic Sea over the last fifty years. Moreover, already in 1952 and 1955 the contamination was found at the polish coast causing serious injuries to people. It was estimated that if only $1/6$ of the sunk chemical agents was released into Baltic the life in the sea and at its shores would be entirely ruined for the next 100 years~. High economic and environmental costs have been preventing so far any activities aiming at extraction of these hazardous substances, but it is clear that we are about to face a very serious problem in the Baltic Sea. Appropriate actions for preventing the ecological catastrophe demand a precise knowledge of location and amount of sunk munitions.\\ Presently used methods for underwater munition detection is based on sonars which show only a shape of underwater objects, like e.g. sunk ships or depth charges. To estimate the amount of dangerous substances and to determine the exact location of sunk munition it is still necessary that people are diving and searching the bottom of the sea. This operation is always very dangerous for divers since the corrosion state of the shells is usually not known. Moreover, these methods are very expensive and slow, thus they cannot be used in practice to search big sea areas. The above mentioned disadvantages can be to large extend overcome by using devices based on Neutron Activation Analysis techniques (NAA) which will be discussed in more details in next sections of this article. \section{Underwater detection of hazardous substances with neutron beams} Most of the commonly used explosives or drugs are organic materials. Therefore, they are composed mostly of oxygen, carbon, hydrogen and nitrogen. War gases contain also sulfur, chlorine, phosphorus and fluorine. Thus, these substances can be unambiguously identified by the determination of the ratio between number of C, H, N, O, S, P and F atoms in a molecule, which can be done noninvasively applying Neutron Activation Analysis techniques. The suspected item can be irradiated with a flux of neutrons produced using compact generators based e.g. on deuteron-tritium fusion, where deuterons are accelerated to the energy of 0.1 MeV and hit a solid target containing tritium. As a result of the fusion an alpha particle is created together with the neutron, which is emitted nearly isotropically with a well defined energy equal to about 14.1 MeV~. Such energy is sufficient to excite all nuclei composing organic substances and the resulting $\gamma$ quanta from the de-excitation of nuclei are then detected by e.g. a germanium detector providing a very good energy resolution. Counting the number of gamma quanta emitted by nuclei from the examined item provides information about its stoichiometry. Devices using NAA to detect explosives on the ground were already designed and are produced, e.g. in USA~ and Poland~. In the aquatic environment however we encounter serious problems since neutrons are strongly attenuated by water. Moreover, as in the case of ground detectors, an isotropic generation of neutrons induces a large environmental background, in this case from oxygen. This noise can be significantly reduced by the requirement of the coincident detection of the alpha particle, which allows for the neutron tagging~. The attenuation of neutrons can be compensated by decreasing the distance between generator and examined item~. There are also solutions based on low energy neutrons which are moderated in water before reaching the tested object. The detector is then counting the gamma quanta from thermal neutron capture and secondary neutrons originating from the irradiated object. The identification is done by searching for anomalies in the observed spectra of gamma quanta and neutrons~. However, these methods do not allow to detect explosives buried deeply in the bottom of the sea. Moreover, the device has to approach the suspected item very close and the strong attenuation of neutrons and gamma quanta significantly increases the exposure time and make the interpretation of results difficult. Therefore, we propose to build a detector which uses NAA technique and special guides for neutrons and emitted gamma rays~. The device allows for detection of dangerous substances hidden deep in the bottom of the sea with significantly reduced background and provides determination of the density distribution of the dangerous substance in the tested object. \section{Concept of using the neutron guides} Scheme of the proposed device is presented in Fig~. Neutron generator collides deuterium ions with a Tritium target producing a neutron and an $\alpha$ particle. Because of the much higher energy released in this reaction compared to the energy of deuterium, both particles are produced almost isotropically and move back-to-back. The $\alpha$ particle is detected by a system of position sensitive detectors, e.g. silicon pads or scintillation hodoscope, disposed on the walls of the generator. Neutrons emitted towards the subject of interrogation fly inside a guide built out of a stainless steal pipe filled with low pressure air or some other gas having low cross section for neutron interaction. Neutrons after leaving a guide may be scattered inelastically on atomic nuclei in the tested substance. The nuclei deexcite and emit gamma quanta with energy specific to the element. Part of the emitted $\gamma$ quanta fly towards a dedicated detector within an analogous guide which decrease their absorption and scattering with respect to gamma quanta flying in water. The $\gamma$ ray detector could be again a position sensitive detector measuring the energy, time and impact point of impinging particles. If the diameter-to-length ratio of both guides is small the depth at which $\gamma$ quanta excite nuclei can be determined by measuring the time $\Delta t$ elapsed from the $\alpha$ particle registration until the time of the $\gamma$ quantum registration. Generated neutrons travel with known velocity $v_n$ the distance $L_n$ from the Tritium target to the point of interaction. Similarly the gamma emitted from the tested object fly over a distance $L_{\gamma}$ with a speed of light $c$. Thus, $\Delta t$ can be expressed as: \begin{equation} \Delta t = L_n/v_n + L_{\gamma}/c - L_{\alpha}/v_{\alpha}, \end{equation} If we know the positions of the generator target and $\gamma$ ray detector and lengths of both guides the distance $x$ covered by neutrons from the end of the guide to the point of interaction can be calculated as: \begin{equation} x = \left(\Delta t + \frac{L_{\alpha}}{v_{\alpha}} - \frac{l_{n}}{v_{n}} -\frac{l_{\gamma}}{c} \right) \frac{c v_n \mathrm{cos\phi}}{c \mathrm{cos}\phi + v_{n}}, \end{equation} where $l_{n}$ and $l_{\gamma}$ are the length of guides for neutrons and $\gamma$ quanta, respectively. Additional information on the depth of interaction is given by changing the relative position of the guides and the angle $\phi$ between them. As it is demonstrated in Fig.~ a) and b) changing the distance between the guides provides registration of gamma quanta emitted from different parts of the object. This allows one to determine the density distribution of elements building the suspected object. In order to remove background resulting from interaction of neutrons emitted in other directions only signals registered by the $\gamma$ quanta detector in coincidence with signals from $\alpha$ particle detectors mounted in-line with the neutron guide are saved, while the other coincidences are discarded. Taking into account cross section for neutron inelastic scattering with different nuclides and the detection efficiency of $\gamma$ quanta we can reconstruct the number of atoms of each element building the object and compare them with the known stoichiometry of hazardous substances stored in the database. The whole detection unit can be mounted on a small submarine steered remotely from a ship. \section{Prototype design: simulations} In order to optimize the dimensions and relative positions of detectors and guides we have developed dedicated open source software package written in the C++ programming language and based on the Monte Carlo simulation methods. Our goal is to create a fast and user-friendly tool using novel methods of geometry definition and particle tracking based on hypergraphs~. The simulation framework is written using the C++11 standard and Open MPI library~ supporting parallel computing and it is destined for Unix-like operating systems. The application needs to be configured with input file defining scene description (location and shape of all objects included in simulation, as well as substances building them), and neutron source parameters (location, number of generated neutrons and their energies). The parameters of neutron interaction with selected nuclei, e.g. total cross sections, neutron and $\gamma$ quanta angular distributions and multiplicities, are parametrized as a function of neutron energy using data from the ENDF database~. Similarly, gamma quanta energies were taken from the Evaluated Nuclear Structure Data Files (ENSDF)~. This information is stored in an SQLite database and recalled during the simulation only for elements specified by user. So far we have implemented only main processes induced by 14.1~MeV neutrons which we are interested in, i.e. elastic and inelastic scattering and radiative capture, but in the near future the simulations will be supplemented with other processes, e.g. fission. At present neglected processes are taken into account effectively as one process after which neutron is no longer tracked. Cross section for this process is calculated in a way that the sum of all processes induced by neutron is equal to the total cross section for a given nuclei with which the reaction occurred. Neutrons are tracked according to the algorithm presented in Fig.~ until they reach the scene boundary or their energy goes below the lowest included in the database~\footnote{ All results presented in this article were obtained however neglecting neutrons with energy E~$<$~10~keV.}, i.e. $10^{-5}$~eV~. The reaction place is randomly generated with exponential probability distribution taking into account total cross section values from the database and concentrations of all the isotopes building the substance~\footnote{ Currently we take into consideration only the most abundant isotope for each element.}. User can choose to calculate concentrations for a single chemical compound or a mixture of substances for a specified density. Selection of an isotope with which the interaction took place is also random and it is done again based on neutron total cross sections and known stoichiometry of the material in which the reaction is simulated. Next the reaction type is drawn according to cross section values for each process from the list specified by user in the input file. The direction of neutron after the reaction is generated using angular distributions parametrized with Legendre polynomials or using uniform distribution if there is no relevant data in the database. It is first determined in the center of mass system and then the four-momentum is transformed back to the laboratory coordinate system. In case of inelastic scattering neutron energy in the laboratory is calculated taking into account the nuclei excitation energy. In the current version of simulations we take into account up to tenth excited nuclei level. Directions of gamma quanta coming from the nuclei deexcitation are currently generated uniformly in the laboratory frame. \\ As a starting point for design of the device for underwater threats detection we have defined a simple setup with point-like source generating 14.1~MeV neutrons uniformly in space. The scheme of the simulated setup with superimposed points of of neutrons interaction are shown in Figs.~ and~. We have considered two neutron guides filled with air under normal conditions: cuboid with dimensions 40~x~40~x~100~cm$^3$ (Figs.~a and~a) and truncated pyramid with height equal to 100~cm and bases with dimensions of 5~cm and 20~cm (Figs.~b and~b). The interrogated object with dimensions 194~x~255~x~50~cm$^3$ lies on the bottom of a sea and contains mustard gas. As one can see in Fig.~ the distribution of the path length of neutrons in water is characterized by a mean free path of about 9.5~cm. At the same time the flux flying through the guide filled with air reaches the container with mustard gas and excites its nuclei. Comparing Fig.~a and Fig.~b one can see also that both shapes of the neutron guide give effectively the same flux irradiating the gas container, but for the trapezoidal shape a better spatial separation between regions where neutrons interact in water and in the interrogated object is clearly visible. The optimization of shapes and configuration of the neutron and gamma quanta guides will be a subject of future investigations. The energy spectra of gamma quanta from deexcitation of nuclei in water and interrogated object are presented in Fig.~. In Fig.~a one can see strong lines from the excitation of Oxygen with energies about 3~MeV, 6~MeV and 7~MeV, and characteristic line of neutron capture by Hydrogen (about 2.2~MeV). The spectrum for the container with mustard gas in Fig.~b contains a big peak for gamma quanta emmited by Carbon $^{12}$C ($E_{\gamma} =$~4.43~MeV) and a structure at low energies and a peak at $E_{\gamma} \sim$~6~MeV comming from Chlorine $^{37}$Cl. Hydrogen and Sulfur $^{32}$S compose one line at $E_{\gamma} =$~2.2~MeV, which shows that identification of this element will be difficult. \section{Summary and outlook} Methods of chemical thread detection based on neutron activation have a huge potential and may open a new frontier in homeland security. In the aquatic environment application of this method encounters serious problems since neutrons are strongly attenuated. In order to suppress this attenuation and to decrease background from gamma radiation induced in the water we propose to use guides for neutrons and gamma quanta which speeds up and simplifies identification. Moreover, it may provide a determination of the density distribution of a dangerous substance. For designing of a device exploiting this idea we have been developing a fast and user-friendly simulation package using novel methods of geometry definition and particle tracking based on hypergraphs. Although we are in a very early stage of the development the first results indicate that indeed the guides will increase the performance of underwater threats detection with fast neutrons. \section*{Acknowledgments} This work was supported by the Polish Ministry of Science and Higher Education through grant No. 7150/E-338/M/2014 for the support of young researchers and PhD students of the Department of Physics, Astronomy and Applied Computer Science of the Jagiellonian University. \begin{thebibliography}{} \bibitem{czasmorza} T. Kasperek, {\em Czas Morza} {\bf 1}, 15 (2001). \bibitem{moskalAnn} P. Moskal, Annales UMCS, Physica \textbf{66}, 71 (2012). \bibitem{maglich} B. C. Maglich, AIP Conf. Proc. \textbf{796}, 431 (2005). \bibitem{SWAN} {\L}. Ka{\'z}mierczak et al., Acta Physica Polonica A, this issue. \bibitem{mpdActab} M.~Silarski, Acta Phys.\ Polon.\ Supp.\ {\bf 6}, 1061 (2013). \bibitem{uncoss} C. Eleon, B. Perot, C. Carasco, D. Sudac, J. Obhodas, V. Valkovic, Nucl. Instrum. Methods. Phys. Res. A {\bf 629}, 220 (2011). \bibitem{pat} D. Lambertus, T. Schneider-Pungs, K. Buckup, Patent application WO2012089584 A1. \bibitem{patent} M. Silarski, P.~Moskal, Patent application PL409388. \bibitem{infuj} E. Grabska, A. {\L}achwa, G. {\'S}lusarczyk, Adv. Eng. Inform. \textbf{26}, 681 (2012). \bibitem{mpi} J. Q. Michael, \textit{Parallel Programming in C with MPI and OpenMP}, McGraw-Hill Inc., 2004, ISBN 0-07-058201-7. \bibitem{endf} M.B. Chadwick et al., Nucl. Data Sheets \textbf{112}(12), 2887 (2011). \bibitem{ensdf} M. R. Bhat, Nuclear Data for Science and Technology (revised as of April 2014) edited by S. M. Qaim (Springer-Verlag, Berlin, Germany, 1992) \end{thebibliography} |
1501.04122 | Title: Numerical simulation of super-square patterns in Faraday waves
Abstract: We report the first simulations of the Faraday instability using the full
three-dimensional Navier-Stokes equations in domains much larger than the
characteristic wavelength of the pattern. We use a massively parallel code
based on a hybrid Front-Tracking/Level-set algorithm for Lagrangian tracking of
arbitrarily deformable phase interfaces. Simulations performed in rectangular
and cylindrical domains yield complex patterns. In particular, a
superlattice-like pattern similar to those of [Douady & Fauve, Europhys. Lett.
6, 221-226 (1988); Douady, J. Fluid Mech. 221, 383-409 (1990)] is observed. The
pattern consists of the superposition of two square superlattices. We
conjecture that such patterns are widespread if the square container is large
compared to the critical wavelength. In the cylinder, pentagonal cells near the
outer wall allow a square-wave pattern to be accommodated in the center.
Body: \NeedsTeXFormat{LaTeX2e} \maketitle \begin{abstract} We report the first simulations of the Faraday instability using the full three-dimensional Navier-Stokes equations in domains much larger than the characteristic wavelength of the pattern. We use a massively parallel code based on a hybrid Front-Tracking/Level-set algorithm for Lagrangian tracking of arbitrarily deformable phase interfaces. Simulations performed in square and cylindrical domains yield complex patterns. In particular, a superlattice-like pattern similar to those of [Douady \& Fauve, {\em Europhys. Lett.}~{\bf 6}, 221-226 (1988); Douady, {\em J. Fluid Mech.}~{\bf 221}, 383-409 (1990)] is observed. The pattern consists of the superposition of two square superlattices. We conjecture that such patterns are widespread if the square container is large compared to the critical wavelength. In the cylinder, pentagonal cells near the outer wall allow a square-wave pattern to be accommodated in the center. \end{abstract} \section{Introduction} In the problem, the interface between two superposed fluid layers, subjected to periodic vertical vibration of sufficient amplitude, forms sustained standing wave patterns, called Faraday waves. Many universal dynamical-systems phenomena were first discovered through the Faraday experiment. In particular, a number of experimental investigations have focused on the variety of patterns formed in square \cite[e.g.][]{Douady-epl-1988,Douady-jfm-1990,Simonelli-jfm-1989} and cylindrical \cite[e.g.][]{Ciliberto-jfm-1985,Das-jfm-2008,Batson-jfm-2013,Rajchenbach-prl-2013} containers, which have led to important theoretical developments \cite[e.g.][]{Meron-pra-1987,Silber-phl-1989,Crawford-physd-1990,Crawford-physd-1991,Gomes-nonl-1994} in the fields of nonlinear physics and pattern formation. Recently, performed the first three-dimensional direct numerical simulation of Faraday waves in horizontally periodic domains containing a few basic cells. Here, we describe numerical simulations of Faraday waves in much larger domains which have yielded more complicated patterns. For this purpose, we have used the free-surface code developed by for simulations of two-phase flows on massively parallel computer architectures. In a square domain (i.e.~with square horizontal cross-section), we have obtained patterns with two different length scales, very similar to those reported in experiments by and . Patterns of this type were later termed superlattices by and , who observed a large variety of such patterns in their experiments, but by using a temporal forcing with two frequencies. The spectral analysis of our patterns, produced by single-frequency forcing, shows that they result primarily from the superposition of two superlattices constructed from very similar wavelengths. We have also simulated the same conditions in a cylindrical container. The resulting interface contains squares and pentagons. The paper is organized as follows. First, we briefly present the problem and explain the methods involved in the code. We then present the numerical linear stability analysis that is validated by comparison with the theoretical treatment of . We demonstrate the robustness of the obtained superlattice-like pattern by testing several boundary conditions. This pattern is described and analysed spectrally. Then, we show the results of simulations with the same conditions but in a cylindrical domain. \section{Problem formulation, governing equations and numerical scheme} The governing equations for an incompressible two-phase flow can be expressed by a single field formulation: \begin{equation} \displaystyle{\rho \left( \frac{\partial \textbf{u}}{\partial t} +\textbf{u}\cdot\nabla \textbf{u}\right) = -\nabla P + \rho \textbf{G} + \nabla \cdot \mu\left(\nabla\textbf{u} +\nabla\textbf{u}^T \right) + \textbf{F}}, \qquad \displaystyle{\nabla\cdot\textbf{u}=0 } \end{equation} where $\textbf{u}$ is the velocity, $P$ is the pressure, $\rho$ is the density, $\mu$ is the dynamic viscosity, $\textbf{G}$ represents the homogeneous volume forces, and $\textbf{F}$ is the local surface tension force at the interface. Here, ${\bf G}$ contains the gravitational term supplemented by a time-dependent force accounting for the vibrations of the domain: \begin{equation} {\bf G}=-\left(g+\gamma\cos(\omega t)\right)\textbf{e}_z \end{equation} where $g$ is the gravitational acceleration, $\textbf{e}_z$ is the upward vertical unit vector, $\gamma$ is the amplitude of the inertial forcing and $\omega=2\pi/T$ its frequency. The code essentially consists of two modules: one, which solves the three-dimensional incompressible Navier-Stokes equations, and the other, which treats the free surface using a parallel Lagrangian Tracking method and is only active if the flow is a two-phase flow. Material properties such as density or viscosity are defined in the entire domain: \begin{equation} \left.\begin{array}{c} \rho(\textbf{x},t) = \rho_{1} +\left( \rho_{2} -\rho_{1}\right)I(\textbf{x},t)\\ \mu(\textbf{x},t) = \mu_{1} +\left( \mu_{2} -\mu_{1}\right)I(\textbf{x},t). \end{array}\right. \end{equation} The indicator function, $I$, is a numerical Heaviside function, ideally zero in one phase and one in the other phase. $I$ is resolved with a sharp but smooth transition across 3 to 4 grid cells and is generated using a vector distance function computed directly from the tracked interface (). The fluid variables $\textbf{u}$ and $P$ are calculated by a projection method \cite[][]{Chorin-mc-1968}. The temporal scheme is first order, with implicit time integration used for the viscous terms. For spatial discretization we use the staggered-mesh marker-in-cell (MAC) method \cite[][]{Harlow-pof-1965} on a uniform finite-difference grid with second-order essentially non-oscillatory (ENO) advection \cite[][]{Shu-jcp-1989}. The pressure and distance function are located at cell centers while the $x$, $y$ and $z$ components of velocity are located at the faces. All spatial derivatives are approximated by standard second-order centered differences. The treatment of the free surface uses a hybrid Front-Tracking/Level-Set technique which defines the interface both by a discontinuous density field on the Eulerian grid as well as by triangles on the Lagrangian interface mesh. The surface tension $\textbf{F}$ is implemented by the hybrid formulation \begin{equation} \textbf{F} = \sigma \kappa_{H} \nabla I , \qquad \kappa_{H}=\frac{\textbf{F}_{L}\cdot\textbf{N}}{\textbf{N}\cdot\textbf{N}}\end{equation} where $\sigma$ is the surface tension coefficient assumed to be constant and $\kappa_{H}$ is twice the mean interface curvature field calculated on the Eulerian grid, with \begin{equation} \textbf{F}_{_{L}} = \int_{\Gamma(t)} \kappa_f \textbf{n}_{f} \delta_f \left( \textbf{x} - \textbf{x}_{f} \right) \mathrm{d}s , \qquad \textbf{N} = \int_{\Gamma(t)} \textbf{n}_{f} \delta_f \left( \textbf{x} - \textbf{x}_{f} \right) \mathrm{d}s \end{equation} Here, $\textbf{x}_f$ is a parameterization of the time-dependent interface, $\Gamma(t)$, and $\delta_f \left( \textbf{x} -\textbf{x}_{f} \right) $ is a Dirac distribution that is non-zero only where $\textbf{x}=\textbf{x}_f$; $\textbf{n}_f$ stands for the unit normal vector to the interface and $ \mathrm{d}s$ is the length of an interface element; $\kappa_f$ is twice the mean interface curvature obtained on the Lagrangian interface. The geometric information, unit normal, $\textbf{n}_f$, and interface element length, $ \mathrm{d}s$ in $\textbf{N}$ are computed directly from the Lagrangian interface and then distributed onto the Eulerian grid using the discrete delta function and the immersed boundary method of . A detailed description of the procedure for calculating $\textbf{F}$, $\textbf{N}$ and $I$ can be found in . The Lagrangian interface is advected by integrating $\mathrm{d}\textbf{x}_f/\mathrm{d}t = \textbf{V}$ with a second-order Runge-Kutta method where the interface velocity, $\textbf{V}$, is interpolated from the Eulerian velocity. The time step $\Delta t$ is chosen at each iteration in order to satisfy a criterion based on \begin{equation} \left\{\Delta t_{\rm CFL}, \Delta t_{\rm int}, \Delta t_{\rm vis}, \Delta t_{\rm cap}\right\} \end{equation} which ensures stability of the calculations. These bounds are defined by: \begin{equation} \begin{array}{ll} \displaystyle{\Delta t_{\rm CFL} \equiv \min_j \left( \min_{\rm domain} \left(\frac{\Delta x_j}{u_j}\right) \right)} & \displaystyle{\Delta t_{\rm int} \equiv \min_j \left( \min_{\Gamma(t)} \left( \frac{\Delta x_j}{\|{\bf V}\|}\right) \right)}\\ \displaystyle{\Delta t_{\rm vis} \equiv \min\left(\frac{\rho_{2}}{\mu_{2}},\frac{\rho_{1}}{\mu_{1}}\right)\frac{\Delta x_{\min}^2}{6}}& \displaystyle{\Delta t_{\rm cap} \equiv \frac{1}{2}\bigg(\frac{(\rho_{1} + \rho_{2}) \Delta x_{\min}^3}{\pi \sigma} \bigg)^{1/2}} \end{array} \end{equation} where $\Delta x_{\min} = \min_j \left(\Delta x_j \right)$. With the periodic volume force term $\textbf{G}$ of (), a supplementary upper bound $\Delta t_{\omega}=2\pi/(50\omega)$ is required. Fortran 2003 allows the definition of a set of dynamically allocated derived types and generic procedures associated with the grids, scalar and vector fields, operators as well as the various solvers used in the Navier-Stokes and Lagrangian tracking modules. The parallelization of the code is based on algebraic domain decomposition, where the velocity field is solved by a parallel generalized minimum residual (GMRES) method for the implicit viscous terms and the pressure by a parallel multigrid method motivated by the algorithm of . Communication across process threads is handled by message passing interface (MPI) procedures. Finally, the code also contains a module for the definition of immersed solid objects and their interaction with the flow, which we have used to simulate Faraday waves in a cylindrical container. A Navier-slip dynamic contact line model is implemented where hysteresis is accounted for by fixing the contact angle limits to prescribed advancing or receding angles as in . Further details are available in . \section{Parameters and linear stability} \begin{figure*} \caption{ a) Evolution of the total kinetic energy for $\omega/2\pi=30$ Hz and various accelerations in a small domain. The fact that the peaks for a given acceleration lie along a line shows that the energy growth is very close to exponential. b) Evolution of the various bounds on the time step during the course of the simulation. The smallest, $\Delta t_{\rm cap}$, is used in the simulations. } \end{figure*} \begin{table} \begin{center} \begin{tabular}{ccccc} $\omega/2\pi $& $\lambda_c$ & Floquet & DNS & relative error\\ Hz & mm & $\gamma_c/g$ & $\gamma_c/g$ & Floquet/DNS\\ \hline 30 & 11.3 & 0.733 & 0.713 & 2.73 \ 60 & 5.9 & 2.65 & 2.61 & 1.50 \ 90 & 4.3 & 5.32 & 5.3 & 0.38 \ \end{tabular} \end{center} \caption{Comparison of Floquet and DNS instability thresholds for various frequencies.} \end{table} We consider a layer of silicone oil of thickness $h=14.5$ mm with density $\rho_{1} = 965$ kg m$^{-3}$ and dynamic viscosity $\mu_{1} = 0.02$ kg m$^{-1}$s$^{-1}$. This layer is overlaid with $29.5$ mm of air of density $\rho_{2}= 1.205$ kg m$^{-3}$ and dynamic viscosity $\mu_{2}=1.82\times10^{-5}$ kg m$^{-1}$s$^{-1}$. The surface tension at the interface between the oil and the air is $\sigma=0.02$ kg s$^{-2}$. This choice of parameters was originally motivated by experiments on Faraday waves in a rotating cylindrical container~\cite[][]{CPC}, which will be the subject of a future investigation. Before simulating the patterns, we computed the critical acceleration $\gamma_c$ from direct numerical simulations (DNS) and compared it with that of the Floquet linear stability analysis of . These simulations are conducted in a small periodic domain of length $\lambda_c$ and width $\lambda_c/2$, where the critical wavelength is $\lambda_c = 2\pi/k_c= 11.3~{\rm mm}$ with $k_c$ the critical wavenumber. We compute the growth rate of the total kinetic energy of the system (figure (a)) after the growth becomes exponential. Near the threshold $\gamma_c$, the growth rate varies linearly with $\gamma$, so that $\gamma_c$ can be deduced by linear interpolation. Table~ compares the Floquet predictions of $\gamma_c$ to the results from DNS for three vibration frequencies. The discrepancy is under 3\ gravitational to capillary forces is measured by the Bond number, whose value here is $g\Delta\rho/(\sigma k_c^2)=1.5$, indicating that both forces are at work. Our large-scale simulations are carried out with a vibration frequency of $\omega/2\pi=30$ Hz, i.e. a period of $T\approx0.033$ s, and an amplitude of $\gamma= g = 1.36\gamma_c$. Figure (b) tracks the timestep bounds () during the simulation. It can be seen that $\Delta t_{\rm cap}$ is the limiting timestep throughout the simulation and is on the order of $5 \times 10^{-5}~{\rm s}$. The domain is of size $L \times L \times L/3$, where $L=132~{\rm mm} = 11.7~\lambda_c$. This domain is subdivided into $8\times8\times4=256$ subdomains whose traces can be seen in figure (a), each of which contains $64^3$ gridpoints, leading to an overall resolution of $512 \times 512 \times 256$ gridpoints on a Cartesian mesh. The critical wavelength is thus resolved by $44=512\lambda_c/L$ grid cells. Each subdomain is assigned to a process thread. Our initial condition has zero velocity and random noise on the interface of order $10^{-2}$ mm. The pattern emerges after approximately 40 forcing periods, increases exponentially, and saturates at about 90 forcing periods, requiring about 15 days of computation time on 256 threads of an IBM x3750-M4. The cylindrical container has been modeled by an impermeable solid object as shown in figure (b). \section{Results and discussion } \subsection{Visualisations} \begin{figure*} \caption{ Snapshots of the top view of the interface colored by the vertical velocity for {\it (a)} no-slip, {\it (b)} free-slip, and {\it (c)} periodic boundary conditions.} \end{figure*} \begin{figure*} \begin{center} \end{center} \caption{Interface profiles at times separated by intervals of $T/2$, where $T$ is the forcing period and $2T$ is the subharmonic response period. In (a) and (c), the small squares consist of wells surrounded by ridges with peaks at each corner, while in (b), low central peaks are surrounded by four higher peaks. In (a), a weak diagonal ``bridge'' connects the large squares on the bottom left and top right; in (c), the bridge links the top left and bottom right.} \caption{Velocity profiles on the plane $y=3L/4$, corresponding to each of the three instants shown in figure . The vectors are colored according to their vertical component.} \end{figure*} The top and bottom of the domain are taken to be rigid; we impose no-slip boundary conditions on the velocity field. Three types of boundary conditions have been tested on the lateral walls, {\it (a)} no-slip, {\it (b)} free-slip, and {\it (c)} periodic. Although our parameters were originally chosen for a different purpose, we find that the patterns that we observe are strikingly similar to those observed in earlier experiments by , figure 1(b) and , figures 10(b), 11(b) and 15(b), with different parameters: water, with $\rho_1 =1000\;{\rm kg}\;{\rm m}^{-3}$, $\mu_1=9\times 10^{-4}\;{\rm kg}\;{\rm m}^{-1}\;{\rm s}^{-1}$, and $\sigma = 0.03\;{\rm kg}\;{\rm s}^{-2}$, in a container of dimensions $80\;{\rm mm}\times 80\;{\rm mm}\times 5\;{\rm mm}$ oscillated with $f=37\;{\rm Hz}$ or $f=70\;{\rm Hz}$. For all three boundary conditions, we observe superlattice-like patterns shown in figure . The interface in the $x$-$y$ plane consists of four large square sub-blocks, each composed of smaller squares whose sides have approximate length $\lambda_c$. In each case, the overall pattern has symmetry $D_2$, i.e., it is invariant under reflections through the two diagonals. Each of the four sub-blocks is in phase opposition with its two adjacent neighbors. The patterns in (a) and (c) have periodicity length $L$ while (b) has a period of $2L$. Although one would expect the periodic case (c) to be the simplest to analyze, we observe that the blocks in the free-slip case (b) are more homogeneous than those in cases (a) and (c); we will therefore focus on this case. Figure displays the temporal evolution of the interface at saturation. (A movie of this evolution is available as supplementary material to this article.) After one period of forcing oscillation $T$, troughs are replaced by crests (see figure (a) and (c)) as a consequence of the subharmonic behavior of the interface. This subharmonic behavior is also displayed by the large sub-blocks, since the appearance of each block is transformed into that of its two adjacent neighbors after time $T$ and then returns to its initial shape after time $2T$ has passed. The pattern is thus invariant under the combined operations of rotation by $\pi/2$ and time-translation by $T$, a spatio-temporal symmetry. Figure shows the interface and the velocity on a vertical slice at the same three instants. The contrasting behavior of the velocity in the left and right halves of the slice is displayed. The interface of figure (a) has a maximum (minimum) at the left (right) boundary and vice versa for (c), illustrating again that the field is not periodic in a domain of length $L$. \subsection{Fourier spectra} \begin{figure*} \caption{Spatial spectrum of a snapshot of the interface in the doubled domain. Circles $\mathcal{C}_1$ (dark blue) and $\mathcal{C}_3$ (dark brown) have slightly different radii. $\mathcal{C}_1$ contains the fundamental mode $\hat\zeta_{23,1}$ and its images under reflection and rotation (dark blue dots surrounded by circles). $\mathcal{C}_3$ contains the highest third-harmonic mode $\hat\zeta_{23,3}$ and its images (dark brown dots surrounded by squares). Only these modes and, to a lesser extent, the weaker second harmonic $\hat\zeta_{46,0}$ (orange) have significant amplitude. a) Restriction to $[-25,25]^2$ highlights the square symmetry of the pattern. b) Restriction to the first quadrant $[0,50]^2$ includes $\hat{\zeta}_{46,0}$.} \end{figure*} For free-slip boundary conditions, the pattern is periodic when it is doubled horizontally by reflection at the boundaries. The doubled pattern is invariant under the symmetry group $D_4$ of the square, consisting of reflections through the $x$ and $y$ coordinate axes and rotations of multiples of $\pi/2$. We decompose the doubled interface height profile $\zeta(x,y,t)$ into spatial Fourier modes, and then into spatio-temporal modes as follows: \begin{equation} \zeta(x,y,t)=\sum_{l,m=-\infty}^{\infty} \hat{\zeta}_{l,m}(t)e^{i\textbf{k}_{l,m}\cdot\textbf{x}}=\sum_{l,m=-\infty}^{\infty} \left[\sum_{n=-\infty}^{\infty}\hat{\hat{\zeta}}_{l,m,n}e^{in\Omega t/2}\right] e^{i\textbf{k}_{l,m}\cdot\textbf{x}} \end{equation} where $\textbf{k}_{l,m}\equiv\left(\frac{\pi l}{L},\frac{\pi m}{L}\right)$, $\textbf{x}=\left(x,y\right)$ and $k_{l,m}=|\textbf{k}_{l,m}|$. The square symmetry of the doubled interface implies that the eight modes in an octet $\hat{\zeta}_{\pm l,\pm m}$ and $\hat{\zeta}_{\pm m,\pm l}$ all have the same amplitude. Figure shows two such octets, with representative elements $\hat\zeta_{23,1}$ and $\hat\zeta_{23,3}$, located on circles which we will call $\mathcal{C}_1$ and $\mathcal{C}_3$, respectively, i.e. with $l=23$ and $m=1$ or $m=3$. The circle $\mathcal{C}_1$ has radius $k_c$, the critical wavenumber. The dominant wavevector octet of our pattern, represented by $\textbf{k}_{23,1}$, is in accordance with the experimental observations of , who scanned over the forcing frequency in order to vary the selected wavenumbers and patterns. They noted that the range of existence was greatest for pairs $(l,m)$ with a small ratio $m/l$ and was maximum for $m=1$, as in our case, i.e. for wavevectors which are almost parallel to the walls. The modes on $\mathcal{C}_3$ can be seen as the result of cubic interactions between modes of $\mathcal{C}_1$: \begin{equation} \begin{array}{c} {\bf k}_{23,3}={\bf k}_{23,1}+{\bf k}_{23,1}+{\bf k}_{-23,1} \end{array} \end{equation} Modes on $\mathcal{C}_3$ are fed continually and damped slowly, if at all, since $|k_{23,3}-k_c|/k_c = 0.75\ Generalizing () to any $m \ll l$ shows that such cubic interactions lead to $\textbf{k}_{l,3m}$, with $|k_{l,3m}-k_c|/k_c \sim 4(m/l)^2$. Due to its proximity to the critical circle, this mode is weakly damped and should often be present when Faraday waves are excited in a square domain which is much larger than the critical wavelength. A second harmonic of $\mathcal{C}_1$, arising from the quadratic interaction \begin{equation} {\bf k}_{46,0} = {\bf k}_{23,1} + {\bf k}_{23,-1} \end{equation} is seen in figure (b) along with its image under $\pi/2$ rotation. These modes occur in a quartet rather than an octet because of their reflection symmetry in the coordinate axes and are of smaller amplitude than the modes on $\mathcal{C}_1$ or $\mathcal{C}_3$. The experimental techniques of did not give them access to amplitudes of Fourier modes other than the dominant one. According to the terminology of , the four components $\hat{\zeta}_{\pm l,\pm m}$ comprise one pure mode, while $\hat{\zeta}_{\pm m,\pm l}$ comprise another pure mode. observed mixed modes (which they called symmetric) for small $m/l$, and pure modes (which they called dissymmetric) for larger $m/l$. Pure modes may be combined to form two types of mixed modes, the in-phase mode in which all of the components have the same sign (as in our figure ), and the out-of-phase mode in which the components of the two pure modes are of opposite sign. showed that these two types of mixed modes are equivalent if $l$ and $m$ have opposite parities. showed that this was also true for $l$ and $m$ with the same parity (as in our case) by invoking the concept of hidden symmetries. In a square container with Neumann boundary conditions, the doubled pattern formed by reflecting at each boundary, as we have done above, satisfies periodic boundary conditions, whose inherited (hidden) translation invariances are incorporated into the normal form. For each of $\mathcal{C}_1$ or $\mathcal{C}_3$, the resulting pattern has a superlattice structure, in which squares of different sizes coexist, as shown in figure (a) and (b). Although shown in domains of length $L$, the periodicity length is $2L$. The smallest squares have length scale $2\pi/k_{23,m}$ for $m=1,3$. The patterns formed by modes on $\mathcal{C}_3$ have an additional intermediate length scale $2L/3$. Close examination of figure (b) shows that adjacent medium squares are almost, but not exactly identical. Figure (c) shows a superposition of the two patterns of figure (a) and (b), weighted by the amplitudes of their respective Fourier components. During most of the time, the interface resembles this superposition; compare with figure . When the amplitudes of modes on $\mathcal{C}_1$ and $\mathcal{C}_3$ approach zero, which occurs twice during one subharmonic period (see figure (a) and figure (b)), it is necessary to take modes like $\hat{\zeta}_{46,0}$ into account. \begin{figure*} \caption{Patterns constructed from Fourier modes (a) in $\mathcal{C}_1$, (b) in $\mathcal{C}_3$, and (c) a superposition of the two. Lighter (darker) zones correspond to the peaks (troughs) of the interface. Superlattices, consisting of smaller and larger squares, are observed in all of the plots, Smallest boxes are approximately of size $\lambda_c$. b) Medium squares of size $2L/3$ are present (long-dashed box). c) Reconstructed pattern, made by combining patterns in (a) and (b) weighted by the corresponding Fourier coefficients $\hat{\zeta}_{23,1}$ and $\hat{\zeta}_{23,3}$. Compare with figure~(c). } \end{figure*} The temporal evolution of $\hat\zeta_{23,1}$, $\hat\zeta_{23,3}$ and $\hat\zeta_{46,0}$ is displayed in figure (a) and the spatio-temporal spectrum is shown in figure (b). The mode amplitudes $\hat\zeta_{23,1}$ and $\hat\zeta_{23,3}$ are related by a multiplicative constant and contain only subharmonic temporal components (i.e. $n$ odd in ()). The temporal component $n=1$ corresponding to $\Omega=\omega/2$ dominates the temporal evolution. Both $\hat\zeta_{23,1}$ and $\hat\zeta_{23,3}$ evolve in phase opposition, crossing zero at the same time and yielding patterns such as that in figure (b). In contrast, $\hat\zeta_{46,0}$ is harmonic and dominated by two temporal components $n=0$ and $n=2$, as expected from the quadratic resonance of $\hat\zeta_{23,1}$ with itself or with $\hat\zeta_{23,3}$. Each spatial mode has a finite $n=0$ mean component; the offset of $\hat\zeta_{46,0}$ is especially noticeable. \begin{figure*} \caption{a) Temporal evolution of the three main spatial modes. Modes $\hat\zeta_{23,1}$ (blue curve and circles) and $\hat\zeta_{23,3}$ (red curve and $+$ symbols) are subharmonic: their period of oscillation is twice that of the forcing. Mode $\hat\zeta_{46,0}$ (black curve and $\times$ symbols) is harmonic. b) Temporal spectrum of the main modes. $\hat\zeta_{23,1}$, $\hat\zeta_{23,3}$ contain only odd temporal harmonics while $\hat\zeta_{46,0}$ has mainly even harmonics. Each spatial mode has a finite $n=0$ component and hence a non-zero temporal mean.} \end{figure*} \subsection{Simulation in a cylindrical container} We performed simulations under the same conditions, but in a cylindrical container of diameter $D=L$ and height $h=L/3$ (see figure (b)). The lateral boundary conditions are free-slip and the advancing and receding contact angles are fixed at $100^{\circ}$ and $60^{\circ}$ respectively. Figure shows snapshots of the interface every $T/2$. (A movie of this evolution is available as supplementary material to this article.) The pattern has the spatio-temporal symmetry of invariance under combined time-evolution by $T/2$ and rotation by $\pi/2$, while each instantaneous pattern is reflection-symmetric about both diagonal axes. At the center, we observe squares whose characteristic wavelength remains close to $\lambda_c$ of Table for $f=30\;{\rm Hz}$. Closer to the boundary, five-sided cells can be seen as the pattern organizes itself to fit in the circular domain. Aside from this, the pattern is relatively insensitive to the circular shape of the domain because $L/\lambda_c = 11.7 \gg 1$, in contrast with small-radius experiments and theory. In most of these \cite[e.g.][]{Ciliberto-jfm-1985,Crawford-physd-1990,Das-jfm-2008,Batson-jfm-2013}, the interface takes the form of Bessel functions, including axisymmetric modes, while report intriguing patterns of the form of stars and pentagons. \begin{figure*} \begin{center} \end{center} \caption{Interface profile in the cylinder at time intervals of $T/2$. The central part of the domain is still occupied by squares, but five-sided cells are present near the boundaries, allowing the pattern to fit inside a circle.} \end{figure*} \section{Conclusion} We have presented the first three-dimensional numerical results of the Faraday instability in domains much larger than the critical wavelength in both square and cylindrical containers. In the square domain, the interface displays patterns that are very similar to those found in the work of and . By means of a spatial spectral decomposition, we have described these patterns quantitatively and have found that they have a complex superlattice-like structure due to resonances between modes of very similar wavelengths. We conjecture that this pattern arises from two effects: (i) In a square container which is large compared to the critical wavelength, Faraday wave patterns tend to be mixed modes whose wavevectors are almost parallel to the boundaries of the box, as stated by . (ii) In this case, cubic nonlinearities generate another wavelength which is very close to the dominant one. The combination of the two effects leads to the superlattice pattern that we observe in our simulations and we infer that such patterns should be observed in a wide variety of systems. The present work demonstrates that it is now possible to numerically compute all the fields of Faraday waves in domains sufficiently large to produce complex patterns such as superlattices, quasicrystalline patterns or oscillons. The hydrodynamic mechanisms responsible for the formation of these patterns can then be explored numerically. \subsubsection*{Acknowledgements} We thank S.~Douady, S.~Fauve, and G.~Pucci for helpful discussions. This work was performed using high performance computing resources provided by the Institut du Developpement et des Ressources en Informatique Scientifique (IDRIS) of the Centre National de la Recherche Scientifique (CNRS), coordinated by GENCI (Grand \'Equipement National de Calcul Intensif). N.~P\'erinet was supported by a grant from CONICYT - FONDECYT/postdoctoral research project 3140522. L.S.T. acknowledges support from the Agence Nationale de la Recherche (ANR) for the TRANSFLOW project. This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and future planning (NRF-2014R1A2A1A11051346). \bibliographystyle{jfm} \begin{thebibliography}{99} \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi \bibitem[Arbell \& Fineberg(2002)]{Arbell-pre-2002} \textsc{Arbell, H. \& Fineberg, J.} 2002 Pattern formation in two-frequency forced parametric waves. \emph{Phys.~Rev.~E} \textbf{65}, 036224. \bibitem[Batson, Zoueshtiagh \& Narayanan(2013)]{Batson-jfm-2013} \textsc{Batson, W., Zoueshtiagh, F. \& Narayanan, R.} 2013 The Faraday threshold in small cylinders and the sidewall non-ideality. \emph{J.~Fluid Mech.}~\textbf{729}, 496--523. \bibitem[Chorin(1968)]{Chorin-mc-1968} \textsc{Chorin, A.~J.} 1968 Numerical simulation of the Navier-Stokes equations. \emph{Math.~Comput.}~\textbf{22}, 745--762. \bibitem[Ciliberto \& Gollub(1985)]{Ciliberto-jfm-1985} \textsc{Ciliberto, S. \& Gollub, J.~P.} 1985 Chaotic mode competition in parametrically forced surface waves \emph{J.~Fluid Mech.}~\textbf{158}, 381--398. \bibitem[Cl\'ement, Pucci \& Couder(unpublished)]{CPC} \textsc{Cl\'ement, F., Pucci, G. \& Couder, Y.} Instabilit\'e de Faraday en rotation. unpublished. \bibitem[Crawford, Knobloch \& Riecke(1990)]{Crawford-physd-1990} \textsc{Crawford, J.~D., Knobloch, E., Riecke, H.} 1990 Period-doubling mode interactions with circular symmetry \emph{Physica D}~\textbf{44}, 340--396. \bibitem[Crawford(1991)]{Crawford-physd-1991} \textsc{Crawford, J.~D.} 1991 Normal forms for driven surface waves: Boundary conditions, symmetry, and genericity \emph{Physica D}~\textbf{52}, 429--457. \bibitem[Das \& Hopfinger(2008)]{Das-jfm-2008} \textsc{Das, S.~P. \& Hopfinger, E.~J.} 2008 Parametrically forced gravity waves in a circular cylinder and finite-time singularity \emph{J.~Fluid Mech.}~\textbf{599}, 205--228. \bibitem[Douady \& Fauve(1988)]{Douady-epl-1988} \textsc{Douady, S. \& Fauve, S.} 1988 Pattern selection in Faraday instability. \emph{Europhys.~Lett.}~\textbf{6}, 221--226. \bibitem[Douady(1990)]{Douady-jfm-1990} \textsc{Douady, S.} 1990 Experimental study of the Faraday instability \emph{J.~Fluid Mech.}~\textbf{221}, 383-409. \bibitem[Faraday(1831)]{Faraday-ptrsl-1831} \textsc{Faraday, M.} 1831 On a peculiar class of acoustical figures; and on certain forms assumed by groups of particles upon vibrating elastic surfaces. \emph{ Philos.~Trans.~R.~Soc.~London} \textbf{121}, 299--340. \bibitem[Gomes \& Stewart(1994)]{Gomes-nonl-1994} \textsc{Gomes, M.G.M. \& Stewart, I.} 1994 Steady PDEs on generalized rectangles: a change of genericity in mode interactions. \emph{Nonlinearity} \textbf{7}, 253-272. \bibitem[Harlow \& Welch(1965)]{Harlow-pof-1965} \textsc{Harlow, F.~H. \& Welch, J.~E.} 1965 Numerical calculation of time dependent viscous incompressible flow of fluid with free surface. \emph{Phys.~Fluids.}~\textbf{8}, 2182. \bibitem[Kudrolli, Pier \& Gollub(1998)]{Kudrolli-physd-1998} \textsc{Kudrolli, A., Pier, B. \& Gollub, J.P.} 1998 Superlattice patterns in surface waves. \emph{Physica D}~{\bf 123}, 99--111. See arXiv:chao-dyn/9803016 for more legible versions of figures. \bibitem[Kumar \& Tuckerman(1994)]{Kumar-jfm-1994} \textsc{Kumar, K. \& Tuckerman, L.~S.} 1994 Parametric instability of the interface between two fluids. \emph{J.~Fluid Mech.}~\textbf{279}, 49--68. \bibitem[Kwak \& Lee(2004)]{Kwak-InterScience-2004} \textsc{Kwak, D.~Y. \& Lee, J.~S.} 2004 Multigrid algorithm for the cell-centered finite-difference method II: Discontinuous coefficient case. \emph{Numer.~Meth.~Part.~Differ.~Equ.}~\textbf{20}, 723-–741. \bibitem[Meron(1987)]{Meron-pra-1987} \textsc{Meron, E.} 1987 Parametric excitation of multimode dissipative systems. \emph{Phys.~Rev.~A}~\textbf{35}, 4892--4895. \bibitem[P\'erinet, Juric \& Tuckerman(2009)]{Perinet-jfm-2009} \textsc{P\'erinet, N., Juric, D. \& Tuckerman, L.~S.} 2009 Numerical simulation of Faraday waves. \emph{J.~Fluid Mech.}~\textbf{635}, 1--26. \bibitem[P\'erinet, Juric \& Tuckerman(2012)]{Perinet-prl-2012} \textsc{P\'erinet, N., Juric, D. \& Tuckerman, L.~S.} 2012 Alternating hexagonal and striped patterns in Faraday waves. \emph{Phys.~Rev.~Lett.}~\textbf{109}, 164501. \bibitem[Peskin(1977)]{Peskin-jcp-1977} \textsc{Peskin, C.~S.} 1977 Numerical analysis of blood flow in the heart. \emph{J.~Comput. Phys.}~\textbf{25}, 220--252. \bibitem[Rajchenbach, Clamond \& Leroux(2013)]{Rajchenbach-prl-2013} \textsc{Rajchenbach, J., Clamond, D., Leroux, A.} 2013 Observation of star-shaped surface gravity waves. \emph{Phys.~Rev.~Lett.} \textbf{110}, 094502. \bibitem[Shin(2007)]{Shin-jcp-2007} \textsc{Shin, S.} 2007 Computation of the curvature field in numerical simulation of multiphase flow. \emph{J.~Comput.~Phys.}~\textbf{222}, 872--878. \bibitem[Shin \& Juric(2009a)]{Shin-ijnmf-2009} \textsc{Shin, S. \& Juric, D.} 2009 A hybrid interface method for three-dimensional multiphase flows based on front-tracking and level set techniques. \emph{Int.~J.~Num.~Meth.~Fluids}~\textbf{60}, 753--778. \bibitem[Shin \& Juric(2009b)]{Shin-jmst-2009} \textsc{Shin, S. \& Juric, D.} 2009 Simulation of droplet impact on a solid surface using the level contour reconstruction method. \emph{J.~Mech.~Sci.~Technol.}~\textbf{23}, 2434--2443. \bibitem[Shin, Chergui \& Juric(2014)]{Shin-cf-2014} \textsc{Shin, S. Chergui, J. \& Juric, D.} 2014 A solver for massively parallel direct numerical simulation of three-dimensional multiphase flows. \emph{Comput.~Fluids}, submitted. arXiv:1410.8568[physics.flu-dyn]. \bibitem[Shu \& Osher(1989)]{Shu-jcp-1989} \textsc{Shu, C. W. \& Osher, S.} 1989 Efficient implementation of essentially non-oscillatory shock capturing schemes, II. \emph{J.~Comput.~Phys.}~\textbf{83}, 32--78. \bibitem[Silber \& Knobloch(1989)]{Silber-phl-1989} \textsc{Silber, M. \& Knobloch, E.} 1989 Parametrically excited surface waves in square geometry \emph{Phys.~Lett.~A} \textbf{137}, 349--354. \bibitem[Simonelli \& Gollub(1989)]{Simonelli-jfm-1989} \textsc{Simonelli, F. \& Gollub, J.P.} 1989 Surface wave mode interactions: effects of symmetry and degeneracy \emph{J.~Fluid Mech.}~\textbf{199}, 471--494. \end{thebibliography} |
1501.04125 | Title: GUT-Inspired Supersymmetric Model for h\rightarrow \gamma \gamma and
Muon g-2
Abstract: We study a GUT-inspired supersymmetric model with non-universal gaugino
masses that can explain the observed muon g-2 anomaly while simultaneously
accommodating an enhancement or suppression in the h \rightarrow\gamma\gamma
decay channel. In order to accommodate these observations and m_h \simeq
125-126 GeV, the model requires a spectrum consisting of relatively light
sleptons whereas the colored sparticles are heavy. The predicted stau mass
range corresponding to R_{\gamma \gamma}\ge 1.1 is 100 {\rm \ GeV} \lesssim
m_{\tilde{\tau}} \lesssim 200 {\rm \ GeV}. The constraint on the slepton
masses, particularly on the smuons, arising from considerations of muon g-2 is
somewhat milder. The slepton masses in this case are predicted to lie in the
few hundred GeV range. The colored sparticles turn out to be considerably
heavier with m_{\tilde{g}} \gtrsim 4.5 {\rm \ TeV} and m_{\tilde{t}_1} \gtrsim
3.5 {\rm \ TeV}, which makes it challenging for these to be observed at the 14
TeV LHC.
Body: \begin{center} {\Large\bf GUT-Inspired Supersymmetric Model \\ for $h\rightarrow \gamma \gamma$ and Muon $g-2$ } \vspace{1cm} {\large M. Adeel Ajaib$^{a,}$, Ilia Gogoladze$^{b,}$\footnote{E-mail: ilia@bartol.udel.edu\\ \hspace*{0.5cm} On leave of absence from: Andronikashvili Institute of Physics, 0177 Tbilisi, Georgia.}, and Qaisar Shafi$^{b,}$\footnote{ E-mail: shafi@bartol.udel.edu} } \vspace{.5cm} {\baselineskip 20pt \it $^a$Department of Physics and Astronomy, Ursinus College, Collegeville, PA 19426\\ $^b$Bartol Research Institute, Department of Physics and Astronomy, \\ University of Delaware, Newark, DE 19716, USA } \vspace{.5cm} \vspace{1.5cm} {\bf Abstract} \end{center} We study a GUT-inspired {supersymmetric} model with non-universal gaugino masses that can explain the observed {muon }$g-2$ anomaly while simultaneously accommodating an enhancement {or} suppression in the $h\rightarrow\gamma\gamma$ {decay} channel. In order to accommodate these observations {and $m_h \simeq 125-126$ GeV}, the model {requires} a spectrum consisting of {relatively} light sleptons whereas the colored sparticles are heavy. The predicted stau mass range corresponding to $R_{\gamma \gamma}\ge 1.1$ is $100 {\rm \ GeV} \lesssim m_{\tilde{\tau}} \lesssim 200 {\rm \ GeV}$. { The constraint on the slepton masses, particularly on the smuons, arising from considerations of muon $g-2$ is somewhat milder. The slepton masses in this case are predicted to lie in the few hundred GeV range.} The colored sparticles {turn out to be considerably} heavier with $ m_{\tilde{g}} \gtrsim 4.5 {\rm \ TeV} $ and $ m_{\tilde{t}_1} \gtrsim 3.5 {\rm \ TeV} $, which makes it challenging for these to be observed at {the} 14 TeV LHC. \newpage \renewcommand{\thefootnote}{\arabic{footnote}} \setcounter{footnote}{0} \section{Introduction} {The} ATLAS and CMS {experiments at the LHC} have independently reported the discovery of a Standard Model (SM)--like Higgs boson of mass $m_h \simeq 125-126$ GeV using the combined 7~TeV and 8~TeV data. This discovery is compatible with low ({TeV}) scale supersymmetry . At the same time, after the first LHC {run} we have the following lower bounds on the {gluino and squark} masses~ \begin{equation} m_{\tilde{g}} \gtrsim 1.4~{\rm TeV}~ ({\rm for}~ m_{\tilde{g}}\sim m_{\tilde{q}})~~~ {\rm and}~~~ m_{\tilde{g}}\gtrsim 0.9~{\rm TeV}~ ({\rm for}~ m_{\tilde{g}}\ll m_{\tilde{q}}). \end{equation} {In some well motivated SUSY models the gluino is {the} NLSP in which case $m_{\tilde g} \gtrsim 400$ GeV . } These bounds combined with the bound of 125 GeV on the lightest CP even Higgs boson mass place stringent constraints on the slepton and gaugino (bino or wino) {mass} spectrum in {several} well studied scenarios such as constrained MSSM (cMSSM) , NUHM1 and NUHM2 . {In particular}, as we shall show later, in the above mentioned models, the first two generation sleptons are predicted to be more than 1 TeV in order to accommodate the light CP even Higgs with 125 GeV mass. The stau leptons can still be relatively light due to {a relatively large} trilinear soft supersymmetry breaking (SSB) A-term. There are several motivations to study models that allow for the sleptons be as light as $\sim $ 100 GeV. For instance, the SM prediction for the anomalous magnetic moment of the muon, $a_{\mu}=(g-2)_{\mu}/2$ (muon $g-2$) , {shows} a discrepancy with the experimental results : \begin{eqnarray} \Delta a_{\mu}\equiv a_{\mu}({\rm exp})-a_{\mu}({\rm SM})= (28.6 \pm 8.0) \times 10^{-10}. \end{eqnarray} If supersymmetry is to offer a solution to this discrepancy, the smuon and gaugino (bino or wino) SSB masses should be ${\cal O}(100)$ GeV or so . Thus, it is hard to simultaneously explain the observed Higgs boson mass and resolve the muon $g-2$ anomaly if we consider CMSSM, NUHM1 or NUHM2, since in all these cases, the slepton masses are {larger} than 1 TeV. Recently, there have been several attempts to reconcile this {apparent} tension between muon $g-2$ and the Higgs boson mass within the MSSM framework by assuming {non-universal SSB mass terms for the gauginos or the sfermions at the GUT scale}. {Indeed, a} simultaneous explanation of $m_h$ and muon $g-2$ is possible {in the presence of} $t-b-\tau$ Yukawa coupling unification condition . It has been shown that constraints from FCNC processes are very mild and easily satisfied for the case {in which} the third generation sfermion masses are split from those of the first two generations. However, {if } the muon $g-2$ anomaly and the Higgs boson mass are simultaneously explained with non-universal gaugino and/or sfermion masses, the correct relic abundance of neutralino dark matter is typically not obtained . Consistency with the observed dark matter abundance would further constrain the SUSY parameter space. The Higgs decay channel $h\rightarrow \gamma\gamma$ {in recent times} attracted a {fair }amount of attention because of the {apparent} deviation compared to the SM {prediction}. Currently, the deviation from the SM prediction has significantly reduced but has not completely disappeared. For example, the ATLAS collaboration reported $\mu_{\gamma\gamma} = 1.17\pm 0.27$ , where $\mu_{\gamma\gamma}=\frac{\sigma (pp\rightarrow h\rightarrow \gamma\gamma)}{\sigma (pp\rightarrow h\rightarrow \gamma\gamma)^{SM}}$. The CMS collaboration reported a best-fit signal strength in their main analysis $\mu_{\gamma\gamma} = 1.14^{+ 0.26}_{- 0.23}$ . On the other hand, a cut-based analysis by CMS produced $\mu_{\gamma\gamma}=1.29^{+ 0.29}_{- 0.26}$, which is a slightly different value. This enhancement or suppression in the $h\rightarrow\gamma\gamma$ channel with respect to the SM {may provide a clue} for physics beyond the SM if it is confirmed in the second LHC run. It is known that in order to accommodate an enhancement or suppression in the $h\rightarrow \gamma\gamma$ decay channel in the framework of MSSM, the stau is the one of the best candidates, and {its mass has to be} around 200 GeV or so. It is problematic to accommodate an enhancement or suppression in the $h\rightarrow \gamma \gamma $ decay channel in the framework of CMSSM, NUHM1 or NUHM2 models. { In this paper we present a GUT inspired model which explains the observed $g-2$ anomaly while simultaneously accommodating an enhancement or suppression in the $h\rightarrow\gamma\gamma$ channel. }\ The paper is organized as follows: In Section we describe the phenomenological constraints and the scanning procedure we implement in our analysis. In Section we {provide }motivations for the model used in this paper by briefly reviewing the status of the muon $g-2$ anomaly and $h\rightarrow\gamma\gamma$ in CMSSM and NUHM2. Our results for the $h\rightarrow\gamma\gamma$ channel in {the proposed} model are presented in Section and for the muon $g-2$ anomaly in Section . {Our conclusions are outlined} in Section . \section{Phenomenological Constraints and Scanning Procedure } {We employ Isajet~7.84 interfaced with Micromegas 2.4 and FeynHiggs 2.10.0 to perform random scans over the parameter space.} In Isajet, the weak scale values of gauge and third generation Yukawa couplings are evolved to $M_{\rm GUT}$ via the MSSM renormalization group equations (RGEs) in the $\overline{DR}$ regularization scheme. We do not strictly enforce the unification condition $g_3=g_1=g_2$ at $M_{\rm GUT}$, since a few percent deviation from unification can be assigned to unknown GUT-scale threshold corrections~. With the boundary conditions given at $M_{\rm GUT}$, the SSB parameters, along with the gauge and third family Yukawa couplings, are evolved back to the weak scale $M_{\rm Z}$. In evaluating the Yukawa couplings the SUSY threshold corrections~ are taken into account at a common scale $M_S= \sqrt{m_{\tilde t_L}m_{\tilde t_R}}$. The entire parameter set is iteratively run between $M_{\rm Z}$ and $M_{\rm GUT}$ using the full 2-loop RGEs until a stable solution is obtained. To better account for the leading-log corrections, one-loop step-beta functions are adopted for the gauge and Yukawa couplings, and the SSB scalar mass parameters $m_i$ are extracted from RGEs at appropriate scales $m_i=m_i(m_i)$.The RGE-improved 1-loop effective potential is minimized at an optimized scale $M_S$, which effectively accounts for the leading 2-loop corrections. Full 1-loop radiative corrections are incorporated for all sparticle masses. We implement the following random scanning procedure: A uniform and logarithmic distribution of random points is first generated in the given parameter space. The function RNORMX is then employed to generate a Gaussian distribution around each point in the parameter space. The data points collected all satisfy the requirement of radiative electroweak symmetry breaking (REWSB), with the neutralino in each case being the LSP. { We use Micromegas to calculate the relic density and $BR(b \rightarrow s \gamma)$. The diphoton ratio $R_{\gamma \gamma}$ is calculated using FeynHiggs.} After collecting the data, we impose the mass bounds on all the particles and use the IsaTools package~ to implement the various phenomenological constraints. We successively apply the following experimental constraints on the data that we acquire from ISAJET~7.84: \begin{table}[h!]\centering \begin{tabular}{rlc} $123~{\rm GeV} \leq m_h \leq 127~{\rm GeV}$~~&& \\ $ 0.8 \times 10^{-9} \ \leq \ BR(B_s \rightarrow \mu^+ \mu^-) $&$ \leq\, 6.2 \times 10^{-9} \; (2\sigma)$ & \\ $2.99 \times 10^{-4} \ \leq \ BR(b \rightarrow s \gamma) $&$ \ \leq\, 3.87 \times 10^{-4} \; (2\sigma)$ & \\ $0.15 \leq \frac{BR(B_u\rightarrow \tau \nu_{\tau})_{\rm MSSM}}{BR(B_u\rightarrow \tau \nu_{\tau})_{\rm SM}}$&$ \leq\, 2.41 \; (3\sigma)$. & \end{tabular} \end{table} \section{Slepton Masses in CMSSM and NUHM2 } Before discussing the scenarios {where we} address the muon $g-2$ anomaly and the decay rate $h\rightarrow \gamma \gamma$, we {first} present the {relationship} between {the} light CP even Higgs boson and slepton masses in {two} well studied models, namely CMSSM and NUHM2. While it is true that radiative corrections to the light CP even Higgs boson mass from {the} first two family sleptons are negligible, in the following section {we show} that relations among SSB mass terms {from} GUT scale boundary conditions in CMSSM and NUHM2 models {yield a} strong correlation {between them}. We do not consider the NUHM1 model since it is an intermediate step between CMSSM and NUHM2 in terms of the independent SSB parameters. Therefore, {the} light CP even Higgs boson mass dependence on slepton masses in NUHM1 {can be inferred, more or less, from the} CMSSM and NUHM2 models. {We} have performed random scans in the fundamental parameter space of CMSSM and NUHM2 with ranges of the parameters given as follows: \begin{flalign} ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~&0 \leq m_{16} \leq 5\, \rm{TeV}& \nonumber \\ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~&0 \leq M_{1/2} \leq 3\, \rm{TeV} &\nonumber \\ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~&-3 \leq A_{0}/m_{3} \leq 3 &\nonumber \\ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~&35 \leq \tan\beta \leq 55 ;& \nonumber \end{flalign} \vspace{-1cm} \begin{flalign} &{\rm For~ CMSSM:}~~~~~~~~~~~~~~~~~~ m_{16}=M_{H_u} = M_{H_d}& \nonumber \\ &{\rm For~~ NUMH2:}~~~~~~~~~ \ \ \ 0 \leq M_{H_u} \neq M_{H_d} \leq 5\, \rm{TeV}& \end{flalign} \noindent Here $m_{16}$ is the universal SSB mass parameter for sfermions, and $M_{1/2}$ denotes the universal SSB gaugino masses. $A_{0}$ is the SSB trilinear scalar interaction coupling, $\tan\beta$ is the ratio of the MSSM Higgs vacuum expectation values (VEVs), and $M_{H_u}$, $M_{H_d}$ stand for the SSB mass terms for the MSSM up and down Higgs doublets. { Since the masses of the light CP even Higgs boson and sleptons do not change significantly for $\tan\beta<35$, we used data from our former analysis for $35 \leq \tan\beta \leq 55$ to generate Figure }. In Figure we display our results in the $m_h - m_{\tilde{l}}$ plane for CMSSM (left panel) and NUHM2 (right panel). Here $m_{\tilde{l}}$ stands for {the} left handed slepton masses {for the first two families}. We {observe} that in the CMSSM and NUHM2 models there is a {fairly} strong correlation between the Higgs boson mass ($m_h$) and the first two generation slepton masses ($m_{\tilde{l}}$). Note that the bounds for {the} right handed slepton masses are very similar {and are therefore not displayed}. {The} {\it gray} points are consistent with REWSB and neutralino LSP, {and the} {\it green} points form a subset of the {\it gray} {points} and satisfy the sparticle and Higgs mass bounds, {as well as} all other constraints described in Section . We see from Figure that for both the CMSSM and NUHM2 models, {compatibility with the measurement} $123~{\rm GeV} \leq m_h \leq 127~{\rm GeV}$ requires that the slepton masses lie above 1 TeV. The salient features of the results in Figure 1 can be understood by noting that in order for the stop quark mass to be more than 1 TeV (which is necessary to achieve $m_h \approx 125$ GeV), with universal SSB parameters $M_{1/2}$ and $m_0$, the first and second generation squark masses {acquire masses} in the multi-TeV range, and the corresponding smuon masses lie around the TeV scale. On the other hand, as mentioned above, in order to have an enhancement in muon $g-2$ and in the decay rate of $h\rightarrow \gamma \gamma$, the sleptons need to be much lighter than 1 TeV. Overall, we learn from Figure 1 that in the CMSSM, NUHM1 and NUHM2 scenarios, it is not possible to have enhancement in muon $g-2$ and the decay rate of $h\rightarrow \gamma \gamma$ {relative to the Standard Model}. This conclusion motivates us to explore {alternative} scenarios which can simultaneously accommodate an enhancement or suppresion of $h\rightarrow \gamma \gamma$ and an enhancement in muon $g-2$. \section{$h\rightarrow \gamma \gamma$ Decay and Particle Spectra} One of the most promising Higgs boson decay channels is the $\gamma \gamma$ final state which, at leading order, proceeds through a loop containing charged particles, {including} the charged Higgs, sfermions and charginos.{ In the SM, the leading contribution to $h \rightarrow \gamma \gamma$ decay comes from the W boson loop, the top loop being the next dominant one.} The decay width is given by (see and references therein) \begin{eqnarray} \Gamma(h\rightarrow\gamma\gamma)&=&\frac{G_{F}\alpha^{2}m_{h}^{3}}{128\sqrt{2}\pi} \left| N_{c}\, Q_{t}^{2}\, g_{htt}\, A_{1/2}^{h}(\tau_{t})+ g_{hWW}\, A_{1}^{h}(\tau_{W}) + {\cal A}_{ \text{\tiny{SUSY}} }^{\gamma\gamma}\right|^2, \end{eqnarray} where $g_{hWW}$ is the coupling of $h$ to the $W$ boson. The supersymmetric contribution $ {\cal A}_{ \text{\tiny{SUSY}} }^{\gamma\gamma} $ is given by \begin{eqnarray} {\cal A}_{ \text{\tiny{SUSY}} }^{\gamma\gamma} &=& g_{hH^{+}H^{-}}\, \frac{m_{W}^{2}}{m^{2}_{H^{\pm}}} \, A_{0}^{h}(\tau_{H^{\pm}}) + \sum_f N_c Q_f^2\, g_{h\tilde{f}\tilde{f}}\, \frac{m_Z^2}{m^2_{\tilde{f}}}\, A_0^h(\tau_{\tilde{f}}) + \nonumber \\ && \quad \quad \quad \sum_i g_{h\chi_i^+\chi_i^-}\, \frac{m_W}{m_{\chi_i}}\, A_{\frac12}^h(\tau_{\chi_i}), \end{eqnarray} where $g_{hXX}$ is the coupling of $h$ to the particle $X$ ($= H^{\pm}, \tilde{f}, \chi^{\pm}_i$). The stop and sbottom loop factors have similar contributions as the gluon fusion case. In this case, however, the stau can also contribute to enhance the decay width without changing the gluon fusion cross section. The chargino contribution to the decay width is known to be less than 10\ In the MSSM framework it was shown that only a light stau can give significant enhancement/suppresion in the {process} $gg \rightarrow h \rightarrow \gamma \gamma$, while keeping the lightest CP-even Higgs boson mass in the interval $123~{\rm GeV} \leq m_h \leq 127~{\rm GeV}$. In this paper we discuss the {scenario with non-universal and opposite sign gaugino {masses at $M_{\rm GUT}$, {while} the sfermion masses at $M_{GUT}$ assumed to be universal}. This is a follow up of the work presented in ref. , where it was shown that the muon $g-2$ anomaly can be explained in this model, but the decay rate for $h \rightarrow \gamma \gamma$ was not analyzed. It was shown in ref. that the sleptons can be as light as 100 GeV in this model. This observation motivated us to investigate the decay rate for $h \rightarrow \gamma \gamma$ and study the parameter space {which yields} enhancement or suppression for this {process}. We { perform random scans for} following ranges of the parameters: \begin{eqnarray} 0 \leq m_{16} \leq 3\, \rm{TeV} \nonumber \\ 0 \leq M_{1} \leq 5\, \rm{TeV} \nonumber \\ 0 \leq M_{2} \leq 5\, \rm{TeV} \nonumber \\ -5 \leq M_{3} \leq 0\, \rm{TeV} \nonumber \\ -3 \leq A_{0}/m_{16} \leq 3 \nonumber \\ 2 \leq \tan\beta \leq 60 \nonumber \\ 0 \leq m_{10} \leq 5\, \rm{TeV} \nonumber \\ \mu > 0. \end{eqnarray} Here $M_{1}$, $M_{2}$, and $M_{3}$ denote the SSB gaugino masses for $U(1)_{Y}$, $SU(2)_{L}$ and $SU(3)_{c}$ respectively. We choose different sign for gauginos which was again motivated from the work presented in ref , where it was shown that an opposite sign non-universal gaugino mass case is more preferable from the muon $g-2$ point of view than the same sign non-universal gaugino case. The main message of Section is that {with} universal SSB mass terms for the gaugino and sfermion sectors, it is impossible to have significant SUSY contributions to the decay $h \rightarrow \gamma \gamma $ and muon $g-2$. On the other hand, as shown in ref. , non-universal gaugino masses allow for {sufficiently} light sleptons while keeping the colored sparticles in the multi TeV region. Because of this observation, we investigate the extent to which non-universality is allowed in the gaugino sector to enhance or suppress the decay channel $h \rightarrow \gamma \gamma $. The color coding in Figure is given as follows, {\it Gray} points are consistent with REWSB and neutralino LSP. {\it Green} points form a subset of the {\it gray} points and satisfy the sparticle and Higgs mass bounds, as well as all other constraints described in Section . {\it Brown} points belong to a subset of {\it green} points and satisfy {the constraint} $0.001 \leq \Omega h^2 \leq 1$ on the LSP neutralino relic abundance. We have chosen to display our results for a wider range of $\Omega h^2$, keeping in mind that one can always find points compatible with the current WMAP range for relic abundance with dedicated scans within the brown regions. The results from the $R_{\gamma \gamma} - M_1/M_2$, $R_{\gamma \gamma} - M_1/M_3$ and $R_{\gamma \gamma} - M_2/M_3$ planes show that a significant deviation from universality of gaugino masses in order to have sizable SUSY contribution to $h \rightarrow \gamma \gamma$ decay {is necessary}. For instance, the ratio $M_1/M_3$ needs to be more than 5, while $M_2/M_3> 3$ and $M_1/M_2 > 2$. Not only do we observe a strict prediction of gaugino mass ratios, but also a precise prediction of their values. {In particular, from the } $R_{\gamma \gamma} - M_1$ panel {we can see} that it is difficult to have {an} enhancement of $h\rightarrow \gamma \gamma$ if $M_1 \gtrsim 300$ GeV. At the same time, the upper bound on $M_2$ is {less stringent} and enhancement {of} $h\rightarrow \gamma \gamma$ occurs even with $M_2$ around 1 TeV. { We observe from the $R_{\gamma \gamma} - M_3$ panel that the parameter $M_{3} \gtrsim 2.5$ TeV. The reason for such a large value of $M_3$ is the following. Since we assume universality in sfermion masses and seek solutions {with} sleptons {not heavier} than a few hundred GeV, $m_{16}$ is required to be around a few hundred GeV. Moreover, {with a} tau slepton {mass of} around a hundred GeV, in order to avoid breaking {the} charge symmetry, $A_{\tau}$ needs to be around a hundred GeV. This places a constraint on $A_t$ because we assume a universal trilinear SSB $A_0$ term. Consequently, a relatively small value of $A_t$ is obtained at low scale. On the other hand, it was shown in that the stop mass needs to be more than 3 TeV {if} $A_t$ is not the dominant contributor to the radiative corrections of the light CP-even Higgs boson mass. In order to obtain such a heavy stop quark, when $m_{16}$ is of order hundred GeV, a {fairly} large $M_3$ is required. This tendency can be observed from the {following} semi-analytic {expressions} for stop quark {masses} \begin{eqnarray} m_{Q_{t}}^{2} & \approx &5.41 M_{3}^{2}+0.392 M_{2}^{2}+0.64 m_{16}^{2} +0.115 M_{3}A_{t_{0}} -0.072 M_{3}M_{2}+ \ldots, \nonumber \\ m_{U_{t}}^{2} & \approx &4.52 M_{3}^2-0.188 M_{2}^2+0.273 m_{16}^2-0.066 A_{t_0}^2 -0.145 M_{3}M_{2} + \ldots. \end{eqnarray} It is clear from Eq. () that {if} $m_{16}$, $M_2$ and $A_t$ are of the order {of} hundred GeV or so, the way to obtain several TeV stop quark masses is to also have $M_3$ around several TeV. } In Figure we display our results for the fundamental parameters in the $R_{\gamma \gamma} - m_{16}$, $R_{\gamma \gamma} - \mu$, $R_{\gamma \gamma} - m_{10}$ and $R_{\gamma \gamma} - \tan\beta$ planes. We can see that an enhancement in the $h\rightarrow \gamma \gamma$ channel constrains the parameters in this model. The fundamental parameter $m_{16}$ is restricted to a narrow range, $200 {\rm \ GeV} \lesssim m_{16} \lesssim 600 {\rm \ GeV}$, for $R_{\gamma \gamma}\ge 1.1$. Similarly, the range for the other parameters for a corresponding enhancement in the $h\rightarrow \gamma \gamma$ channel are: $2.5 {\rm \ TeV} \lesssim \mu \lesssim 5.5 {\rm \ TeV}$, $m_{10} \lesssim 2 {\rm \ TeV}$, $10 \lesssim \tan\beta \lesssim 20$. Figure shows our results for the sparticle masses in the $R_{\gamma \gamma} - m_{\tilde{\tau}}$, $R_{\gamma \gamma} - m_{\tilde{\chi}^0_1}$, $R_{\gamma \gamma} - m_{\tilde{g}}$ and $R_{\gamma \gamma} - m_{\tilde{t}_1}$ planes. For $R_{\gamma \gamma}\ge 1.1$, the stau and the neutralino are both {relatively} light with mass ranges $100 {\rm \ GeV} \lesssim m_{\tilde{\tau}} \lesssim 200 {\rm \ GeV} $ and $50 {\rm \ GeV} \lesssim m_{\tilde{\chi}_1^0} \lesssim 200 {\rm \ GeV}$. From the lower panels of Figure we can see that the colored sparticles corresponding to $R_{\gamma \gamma}\ge 1.1$ are heavy with $ m_{\tilde{g}} \gtrsim 4.5 {\rm \ TeV} $ and $ m_{\tilde{t}_1} \gtrsim 3.5 {\rm \ TeV} $. { The reason for such heavy stop and gluino masses has been discussed above.} Testing squarks and gluinos with this mass would be challenging at 14 TeV LHC. \section{The Muon Anomalous Magnetic Moment} The leading contribution from low scale supersymmetry to the muon anomalous magnetic moment depends on the following parameters: \begin{equation} M_1, \, M_2, \, \mu,\, \tan\beta, m_{\tilde{\mu}_{L}}, \, m_{\tilde{\mu}_{R}}, \end{equation} Since we assume a universal the trilinear SSB term $A_0$, it follows that $A_\mu < \mu \tan\beta$ and we therefore do not consider the trilinear SSB-term contribution here. The colored particles do not directly provide significant contribution to the muon $g-2$ calculation but are still constrained from the bound on the light CP even Higgs boson mass and the muon $g-2$ calculation. Figure shows our results in the $\Delta a_\mu \times 10^{10} - m_{\tilde{g}}$ and $\Delta a_\mu \times 10^{10} - m_{\tilde{t}_1}$ planes. We can see that the above mentioned constraint yields a {stringent} lower bound for the gluino and stop {masses}. Both these sparticles have to be heavier than 3 TeV, which makes it very hard {to see them} in the second LHC run. However, there is hope that these sparticles can be observed at {a future} 100 TeV collider. \begin{table} \centering \begin{tabular}{|p{2.6cm}|p{2.6cm}p{2.6cm}p{2.6cm}p{2.6cm}|} \hline \hline & Point 1 & Point 2 & Point 3 & Point 4 \\ \hline $m_{16}$ &$ 351 $&$ 438 $&$ 561 $&$ 392 $\\ $m_{10}$ &$ 451 $&$ 32 $&$ 478 $&$ 2.7 $\\ $A_0/m_{16}$ &$ -2.6 $&$ 1.6 $&$ 0.3 $&$ 2.6 $\\ $\tan\beta$ &$ 12 $&$ 26 $&$ 42 $&$ 43 $\\ $M_{1}$ &$ 88 $&$ 5 $&$ 344 $&$ 749 $\\ $M_{2} $ &$ 714 $&$ 1051 $&$ 124 $&$ 722 $\\ $M_{3} $ &$ -4913 $&$ -4550 $&$ -4420 $&$ -4524 $\\ \hline $\mu$ &$ 471 $&$ 747 $&$ 680 $&$ 811 $\\ \hline $m_h$ &$ 123 $&$ 124 $&$ 124 $&$ 125 $\\ $m_H$ &$ 4847 $&$ 4074 $&$ 1607 $&$ 591 $\\ $m_A$ &$ 4815 $&$ 4047 $&$ 1596 $&$ 587 $\\ $m_{H^{\pm}}$ &$ 4847 $&$ 4075 $&$ 1610 $&$ 600 $\\ \hline $m_{\tilde{\chi}^0_{1,2}}$ &$ 75, 720 $&$ 33, 999 $&$ 184, 198 $&$ 367, 712 $\\ $m_{\tilde{\chi}^0_{3,4}}$ &$ 4774, 4774 $&$ 4570, 4570 $&$ 4512, 4513 $&$ 4628, 4628 $\\ $m_{\tilde{\chi}^{\pm}_{1,2}}$ &$ 726, 4729 $&$ 1004, 4528 $&$ 199, 4470 $&$ 714, 4585 $\\ $m_{\tilde{g}}$ &$ 9709 $&$ 9029 $&$ 8857 $&$ 9008 $\\ \hline $m_{ \tilde{u}_{L,R}}$ &$ 8247, 8262 $&$ 7703, 7697 $&$ 7553, 7582 $&$ 7674, 7689 $\\ $m_{\tilde{t}_{1,2}}$ &$ 7217, 7800 $&$ 6664, 7146 $&$ 6573, 6707 $&$ 6625, 6754 $\\ \hline $m_{ \tilde{d}_{L,R}}$ &$ 8247, 8269 $&$ 7704, 7703 $&$ 7553, 7588 $&$ 7675, 7692 $\\ $m_{\tilde{b}_{1,2}}$ &$ 7756, 8208 $&$ 7107, 7432 $&$ 6613, 6750 $&$ 6608, 6750 $\\ \hline $m_{\tilde{\nu}_{1}}$ &$ 410 $&$ 727 $&$ 446 $&$ 543 $\\ $m_{\tilde{\nu}_{3}}$ &$ 419 $&$ 728 $&$ 677 $&$ 813 $\\ \hline $m_{ \tilde{e}_{L,R}}$ &$ 561, 150 $&$ 787, 354 $&$ 469, 549 $&$ 552, 449 $\\ $m_{\tilde{\tau}_{1,2}}$ &$ 176, 519 $&$ 160, 786 $&$ 368, 892 $&$ 526, 979 $\\ \hline $\Delta(g-2)_{\mu}$ &$ 29.5\times 10^{-10} $&$ 4.67\times 10^{-10} $&$ 29.3\times 10^{-10} $&$ 29.4\times 10^{-10} $\\ $R_{\gamma \gamma}$ &$ 1.2 $&$ 1.1 $&$ 1 $&$ 0.74 $\\ \hline \end{tabular} \caption{{Four} benchmark points from our analysis. The first point has $R_{\gamma \gamma}=1.2$ with the central value of $g-2$. The second point has $(g-2)_\mu$ {consistent with} the SM {and} $R_{\gamma \gamma}=1.1$. The third point also has the central value of $(g-2)_{\mu}$ but with no enhancement in the diphoton channel. {The fourth point shows a suppression in the diphoton channel with $R_{\gamma \gamma}=0.74$}. For all these points the sleptons are {relatively} light {but} the colored sparticles are considerably {heavier}.} \end{table} In Figure we display {results in the} $R_{\gamma \gamma} -\Delta a_\mu $ plane. The plot shows that there is a considerable region of the parameter space that allows for simultaneous enhancement in the decay channel $h \rightarrow \gamma\gamma$ and muon $g-2$. In this model we can have the correct neutralino dark matter relic abundance through the slepton coannihilation channel. It is also interesting that this model connects the parameter space relevant for two different experiments. If the enhancement in the $h \rightarrow \gamma\gamma$ decay channel is excluded than a considerable region of the parameter space that explains the $g-2$ anomaly will also be excluded. Finally, in Table we display {four} benchmark points from our analysis. All these points satisfy the constraints described in section 2. The first and third point {yield a muon $g-2$} around the central {measured} value with $R_{\gamma \gamma} = 1.2$ and $R_{\gamma \gamma} = 1.0$, respectively. The second point has $(g - 2 )_\mu$ {consistent with} the SM and $R_{\gamma \gamma} = 1.1$. {The fourth point shows a suppression in the diphoton channel with $R_{\gamma \gamma}=0.74$}. For all these points the sleptons are light {while} the colored sparticles are considerably heavier. \section{ Conclusion } We studied a supersymmetric model with non-universal gaugino masses at $M_{GUT}$ that accommodates enhancement or suppression in the $h\rightarrow \gamma \gamma$ channel while simultaneously explaining the muon $g-2$ anomaly. The parameter space {we obtain} is consistent with the current bounds on the sparticle masses and constraints from B-physics. {The desired} neutralino dark matter relic abundance is achieved in this model through slepton coannihilation channel. We find that the parameter space with $R_{\gamma \gamma}\ge 1.1$ predicts {relatively} light sleptons with a stau mass range of $100 {\rm \ GeV} \lesssim m_{\tilde{\tau}} \lesssim 200 {\rm \ GeV} $. The colored sparticles corresponding to $R_{\gamma \gamma}\ge 1.1$ are heavy with $ m_{\tilde{g}} \gtrsim 4.5 {\rm \ TeV} $ and $ m_{\tilde{t}_1} \gtrsim 3.5 {\rm \ TeV} $, which makes it very challenging to {observe them in} the second run of LHC. \section*{Acknowledgments} This work is supported in part by the DOE Grant No. DE-FG02-12ER41808. This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by the National Science Foundation grant number OCI-1053575. I.G. acknowledges support from the Rustaveli National Science Foundation No. 03/79. \begin{thebibliography}{99} \bibitem{:2012gk} G.~Aad {\it et al.} [ATLAS Collaboration], Phys.\ Lett.\ B {\bf 716}, 1 (2012). \bibitem{:2012gu} S.~Chatrchyan {\it et al.} [CMS Collaboration], Phys.\ Lett.\ B {\bf 716}, 30 (2012). \bibitem{Martin:1997ns} See, for instance, S.~P.~Martin, arXiv:hep-ph/9709356 [hep-ph] and references therein. \bibitem{Aad:2012fqa} G.~Aad {\it et al.} [ATLAS Collaboration], Phys.\ Rev.\ D {\bf 87}, 012008 (2013). \bibitem{Chatrchyan:2012jx} S.~Chatrchyan {\it et al.} [CMS Collaboration], JHEP {\bf 1210}, 018 (2012). \bibitem{Gogoladze:2009bn} I.~Gogoladze, R.~Khalid and Q.~Shafi, Phys.\ Rev.\ D {\bf 80}, 095016 (2009); M.~A.~Ajaib, T.~Li and Q.~Shafi, Phys.\ Lett.\ B {\bf 705}, 87 (2011); S.~Raza, Q.~Shafi and C.~S.~�n, arXiv:1412.7672 [hep-ph]. \bibitem{Kane:1993td} G.~L.~Kane, C.~F.~Kolda, L.~Roszkowski and J.~D.~Wells, Phys.\ Rev.\ D {\bf 49}, 6173 (1994). \bibitem{Baer:2004fu} H.~Baer, A.~Mustafayev, S.~Profumo, A.~Belyaev and X.~Tata, Phys.\ Rev.\ D {\bf 71}, 095008 (2005). \bibitem{nuhm2} J. Ellis, K. Olive and Y. Santoso, \plb{539}{2002}{107}; J.~Ellis, T.~Falk, K.~Olive and Y.~Santoso, \npb{652}{2003}{259}; H.~Baer, A.~Mustafayev, S.~Profumo, A.~Belyaev and X.~Tata, \jhep{0507}{2005}{065}. \bibitem{Davier:2010nc} M.~Davier, A.~Hoecker, B.~Malaescu and Z.~Zhang, Eur.\ Phys.\ J.\ C {\bf 71}, 1515 (2011) [Erratum-ibid.\ C {\bf 72}, 1874 (2012)]; K.~Hagiwara, R.~Liao, A.~D.~Martin, D.~Nomura and T.~Teubner, J.\ Phys.\ G {\bf 38}, 085003 (2011). \bibitem{Bennett:2006fi} G.~W.~Bennett {\it et al.} [Muon (g-2) Collaboration], Phys.\ Rev.\ D {\bf 73}, 072003 (2006); Phys.\ Rev.\ D {\bf 80}, 052008 (2009). \bibitem{Moroi:1995yh} T.~Moroi, Phys.\ Rev.\ D {\bf 53}, 6565 (1996) [Erratum-ibid.\ D {\bf 56}, 4424 (1997)]; S.~P.~Martin and J.~D.~Wells, Phys.\ Rev.\ D {\bf 64}, 035003 (2001). G.~F.~Giudice, P.~Paradisi, A.~Strumia and A.~Strumia, JHEP {\bf 1210}, 186 (2012) \bibitem{Gogoladze:2014cha} I.~Gogoladze, F.~Nasir, Q.~Shafi and C.~S.~Un, Phys.\ Rev.\ D {\bf 90}, 035008 (2014). \bibitem{Akula:2013ioa} S.~Mohanty, S.~Rao and D.~P.~Roy, JHEP {\bf 1309}, 027 (2013); S.~Akula and P.~Nath, Phys.\ Rev.\ D {\bf 87}, 115022 (2013); J.~Chakrabortty, S.~Mohanty and S.~Rao, arXiv:1310.3620 [hep-ph]. \bibitem{Babu:2014sga} K.~S.~Babu, I.~Gogoladze, S.~Raza and Q.~Shafi, Phys.\ Rev.\ D {\bf 90}, 056001 (2014). \bibitem{Ibe:2013oha} M.~Ibe, T.~T.~Yanagida and N.~Yokozaki, JHEP {\bf 1308}, 067 (2013). \bibitem{Ajaib:2014ana} M.~A.~Ajaib, I.~Gogoladze, Q.~Shafi and C.~S.~Un, JHEP {\bf 1405}, 079 (2014). \bibitem{yukawaUn} B. Ananthanarayan, G. Lazarides and Q. Shafi, Phys. Rev. D {\bf 44}, 1613 (1991); Phys.\ Lett.\ B {\bf 300}, 245 (1993); Q.~Shafi and B.~Ananthanarayan, Proceedings of the Summer School in High Energy Physics and Cosmology 1991, edited by E. Gava, K. Narain, S. Randjbar-Daemi, E. Sezgin, and Q. Shafi, ICTP Series in Theoretical Physics Vol. 8 (World Scientific, River Edge, NJ, 1992). \bibitem{Baer:2004xx} H.~Baer, A.~Belyaev, T.~Krupovnickas and A.~Mustafayev, JHEP {\bf 0406}, 044 (2004); K.~S.~Babu, I.~Gogoladze, Q.~Shafi and C.~S.~Un, Phys.\ Rev.\ D {\bf 90}, 116002 (2014). \bibitem{Carena:2012xa} M.~Carena, I.~Low and C.~E.~M.~Wagner, JHEP {\bf 1208}, 060 (2012) K.~Schmidt-Hoberg, F.~Staub and M.~W.~Winkler, JHEP {\bf 1301} (2013) 124; M.~Carena, S.~Gori, N.~R.~Shah and C.~E.~M.~Wagner, JHEP {\bf 1203} (2012) 014; M.~A.~Ajaib, I.~Gogoladze and Q.~Shafi, Phys.\ Rev.\ D {\bf 86}, 095028 (2012) N.~Maru and N.~Okada, Phys.\ Rev.\ D {\bf 87}, no. 9, 095019 (2013); J.~Guo, Z.~Kang, J.~Li and T.~Li, arXiv:1308.3075 [hep-ph]. M.~Hemeda, S.~Khalil and S.~Moretti, Phys.\ Rev.\ D {\bf 89}, no. 1, 011701 (2014). A.~Chakraborty, B.~Das, J.~L.~Diaz-Cruz, D.~K.~Ghosh, S.~Moretti and P.~Poulose, Phys.\ Rev.\ D {\bf 90}, no. 5, 055005 (2014); S.~Chakraborty, A.~Datta and S.~Roy, arXiv:1411.1525 [hep-ph]. \bibitem{ATLAS-mugg-1} G.~Aad {\it et al.} [ ATLAS Collaboration], ``Measurement of Higgs boson production in the diphoton decay channel in $pp$ collisions at center-of-mass energies of 7 and 8 TeV with the ATLAS detector,'' arXiv:1408.7084 [hep-ex]. \bibitem{CMS-mugg-2} V.~Khachatryan {\it et al.} [CMS Collaboration], ``Observation of the diphoton decay of the Higgs boson and measurement of its properties,'' arXiv:1407.0558 [hep-ex]. \bibitem{ISAJET} F.~E.~Paige, S.~D.~Protopopescu, H.~Baer and X.~Tata, hep-ph/0312045. \bibitem{Belanger:2008sj} G.~Belanger, F.~Boudjema, A.~Pukhov and A.~Semenov, Comput.\ Phys.\ Commun.\ {\bf 180}, 747 (2009). \bibitem{feynhiggs} M.~Frank, T.~Hahn, S.~Heinemeyer, W.~Hollik, H.~Rzehak and G.~Weiglein, JHEP {\bf 0702}, 047 (2007); G.~Degrassi, S.~Heinemeyer, W.~Hollik, P.~Slavich and G.~Weiglein, Eur.\ Phys.\ J.\ C {\bf 28}, 133 (2003); S.~Heinemeyer, W.~Hollik and G.~Weiglein, Eur.\ Phys.\ J.\ C {\bf 9}, 343 (1999); S.~Heinemeyer, W.~Hollik and G.~Weiglein, Comput.\ Phys.\ Commun.\ {\bf 124}, 76 (2000). \bibitem{Hisano:1992jj} J.~Hisano, H.~Murayama , and T.~Yanagida, { Nucl. Phys.} {\bf B402} (1993) 46. Y.~Yamada, { Z. Phys.} {\bf C60} (1993) 83; J.~L.~Chkareuli and I.~G.~Gogoladze, Phys.\ Rev.\ D {\bf 58}, 055011 (1998). \bibitem{Pierce:1996zz} D.~M. Pierce, J.~A. Bagger, K.~T. Matchev, and R.-j. Zhang, { Nucl. Phys.} {\bf B491} (1997) 3. \bibitem{Leva} J.L. Leva, Math. Softw. 18 (1992) 449; J.L. Leva, Math. Softw. 18 (1992) 454. \bibitem{Nakamura:2010zzi} J.~Beringer {\it et al.} [Particle Data Group Collaboration], Phys.\ Rev.\ D {\bf 86}, 010001 (2012). \bibitem{Baer:2002fv} H.~Baer, C.~Balazs, and A.~Belyaev, { JHEP} {\bf 03} (2002) 042; H.~Baer, C.~Balazs, J.~Ferrandis, and X.~Tata { Phys. Rev.} {\bf D64} (2001) 035004. \bibitem{:2007kv} T.~Aaltonen {\it et al.} [CDF Collaboration], Phys.\ Rev.\ Lett.\ {\bf 100}, 101802 (2008). \bibitem{Barberio:2008fa} E.~Barberio {\it et al.} [Heavy Flavor Averaging Group], arXiv:0808.1297 [hep-ex]. \bibitem{Ajaib:2012vc} M.~A.~Ajaib, I.~Gogoladze, F.~Nasir and Q.~Shafi, Phys.\ Lett.\ B {\bf 713}, 462 (2012). \bibitem{Djouadi:2005gj} A.~Djouadi, Phys.\ Rept.\ {\bf 459}, 1 (2008). \bibitem{gunion90} J.~F.~Gunion, H.~E.~Haber, G.~L.~Kane and S.~Dawson, {\it``The Higgs Hunter's Guide''}, Addison-Wesley, Reading (USA), 1990. \bibitem{Gogoladze:2009bd} I.~Gogoladze, M.~U.~Rehman and Q.~Shafi, Phys.\ Rev.\ D {\bf 80}, 105002 (2009). \end{thebibliography} |
1501.04127 | Title: Is a description deeper than the quantum one possible?
Abstract: Recently, it has been argued that quantum mechanics is a complete theory, and
that different quantum states do necessarily correspond to different elements
of reality, under the assumptions that quantum mechanics is correct and that
measurement settings can be freely chosen. In this work, we prove that this
result is a consequence of an unnecessarily strong mathematical expression of
the free choice assumption, which embodies more conditions than explicitly
stated. The issues of the completeness of quantum mechanics, and of the
interpretation of the state vector, are by no means resolved. Taking this
perspective, we describe how the recently introduced class of crypto-nonlocal
hidden variables theories can be used to characterize the maximal possible
departure from quantum mechanics, when the system consists of a pair of qubits.
Body: \title[Is a description deeper than the quantum one possible?]{Is a description deeper than the quantum one possible?} \author{GianCarlo Ghirardi$^1$ and Raffaele Romano$^2$} \address{$^1$ Department of Physics, University of Trieste, the Abdus Salam ICTP, Trieste, Italy} \ead{ghirardi@ictp.it} \address{$^2$ Department of Mathematics, Iowa State University, Ames IA, USA} \ead{rromano@iastate.edu} \begin{abstract} \noindent Recently, it has been argued that quantum mechanics is a complete theory, and that different quantum states do necessarily correspond to different elements of reality, under the assumptions that quantum mechanics is correct and that measurement settings can be freely chosen. In this work, we prove that this result is a consequence of an unnecessarily strong mathematical expression of the free choice assumption, which embodies more conditions than explicitly stated. The issues of the completeness of quantum mechanics, and of the interpretation of the state vector, are by no means resolved. Taking this perspective, we describe how the recently introduced class of crypto-nonlocal hidden variables theories can be used to characterize the maximal possible departure from quantum mechanics, when the system consists of a pair of qubits. \end{abstract} \pacs{03.65.Ta, 03.65.Ud} \maketitle \section{Introduction} Despite the unprecedented success of quantum theory (and its field-theoretical relativistic generalization) in explaining any experimental evidence of the microscopic world, the interpretation of the formalism represents a long-standing problem. This is partially due to the counter-intuitive features of the description of the micro-world, when compared to the concepts derived from classical physics, the prominent examples being given by probabilism, indeterminism and non-locality. Nonetheless, the really unpleasant feature of the quantum formalism is the fact that it appears more like a set of operational prescriptions to fit the experimental data, rather than a coherent description of reality. The theory relies on two different kinds of evolution depending on the rather vague notion of measurement, is not able to account for the behavior of the classical world in the limit of macroscopic objects, and, consequently, its range of validity is not well defined. The famous incompleteness argument by Einstein, Podolski and Rosen has raised the questions of whether quantum mechanics is a complete theory or not, and how to interpret the state description required by the theory in terms of the quantum state vector $\psi$. In particular, it is unclear whether $\psi$ represents a state of reality or rather a state of knowledge, as suggested by its updating following a measurement procedure. According to these lines, several {\it ontological models} of quantum mechanics have been introduced, that is, theories which are predictively equivalent to quantum mechanics, but providing a possibly richer description of the microscopic reality through the so-called {\it ontic state}, the most accurate specification of the physical state of the system, at least in principle. In these theories, the state vector $\psi$ might embody only partial information on the ontic state, since it is associated to a distribution $\rho_{\psi} (\lambda)$ on the space of the ontic variable $\lambda$, with $\rho_{\psi} (\lambda) \geq 0$ and \begin{equation} \int \rho_{\psi} (\lambda) d \lambda = 1 \qquad {\rm for \, all} \, \psi. \end{equation} The ontic state $\lambda$ accounts for the elements of reality of the underlying theory since it provides the complete description of the state of the system, but, in principle, it might be not fully accessible. This is the reason why these models were formerly characterized as {\it hidden variables theories}. Notice that the ontological status of $\psi$ can be understood by studying the distribution $\rho_{\psi} (\lambda)$, which, in the so-called {\it $\psi$-epistemic models}, has overlapping supports, while in the so-called {\it $\psi$-ontic models}, it has disjoint supports~ for different $\psi$. Accordingly, in a $\psi$-ontic theory different state vectors necessarily correspond to different ontic states, whereas in a $\psi$-epistemic theory they could correspond to the same ontic state. One could associate to $\psi$ well defined elements of reality only in the first case. In a recent work it has been argued that, if quantum mechanics is correct and the experimenters can freely choose their own settings, no theory can outperform its predictive power, that is, the microscopic world can only be described in terms of probabilistic laws, and the probabilities are definitely those provided by the quantum formalism~. In other words, quantum mechanics really is complete. Moreover, as a corollary of this result, it has been proven that $\psi$-epistemic models necessarily contrast with quantum mechanics~. Accordingly, it has been concluded that $\psi$ does not represents a state of knowledge but rather a state of reality. This result already appeared in the recent literature~, although limited by the assumption that factorized quantum states correspond to factorized states of the underlying theory, as highlighted by an explicit model~. In our opinion, the general scenario envisaged in~ is not the right one, since the mathematical expression of the free choice assumption, denoted by $FR$, is unnecessarily strong. While a general criticism to these works from a more epistemological perspective has been presented in Ref.~, here we prove that $FR$ embodies more than the free choice assumption. Therefore, we provide evidence that the argument put forward in~ is not conclusive for stating the completeness of quantum mechanics, nor in determining the ontological status of the quantum state vector. This result triggers the following question: are there theories which are compatible with quantum mechanics, but potentially distinguishable from it? We argue that this question has a positive answer, and the theories fulfilling this requirement are exactly the recently introduced crypto-nonlocal hidden variables models. We discuss what is the maximal departure from quantum expectations that these models can provide in the simple case of a pair of two-level systems. \section{Free will and the ontological status of quantum mechanics} As already anticipated, Ref.~ derives completeness of quantum mechanics and the one-to-one correspondence between quantum state vectors and elements of reality from the assumptions $QM$ of the validity of the predictions of quantum mechanics, and a request $FR$ which, according to the authors, expresses the freedom of choosing the measurement settings. One considers two space-like separated observers performing local measurements on the two parties of an entangled state $\psi$. The measurement settings are given by vectors $A$ and $B$, the outcomes are denoted by $X$ and $Y$. Following~, we assume that additional information on the ontic state $\lambda$ is available and can be obtained through a measurement with setting $C$ and output $Z$. In the following, we do not exclude the case $Z = \lambda$, which means that the ontic state is fully accessible. We consider all these quantities as random variables. The $FR$ assumption is the condition that \begin{quote} {\it [...] the input $A$ can be chosen to be uncorrelated with all the space-time random variables whose coordinates lie outside the future light-cone of its coordinates}~, \end{quote} and the same requirement holds also for $B$ and $C$. The authors of~ have expressed this assumption by imposing the following constraints on the conditional probabilities: \begin{equation} P_{A|BCYZ} = P_A, \quad P_{B|ACXZ} = P_B, \quad P_{C|ABXY} = P_C, \end{equation} which are all needed to derive the main results of~. However, we notice that the free will condition is consistent, among others, with a condition weaker than $FR$, which makes reference exclusively to the fact that the two observers can independently choose which observables to measure: \begin{equation} P_{A|B\lambda} = P_A, \quad P_{B|A\lambda} = P_B, \end{equation} where $\lambda$ is the aforementioned ontic state. This condition, denoted by $FW$ in the following, produces the relevant factorization $P_{AB\lambda} = P_A P_B P_{\lambda}$. Meaningfully, $FW$ is unrelated with the physically important assumption that the two observers cannot communicate superluminally, denoted as $NS$, and expressed by \begin{equation} P_{X|AB} = P_{X|A}, \quad P_{Y|AB} = P_{Y|B}. \end{equation} This is reasonable: one could imagine artificial models in which free will and superluminal signalling coexist. Notice that $FR \Rightarrow NS$, supporting the idea that $FR$ embodies more than the free choice assumption. We observe that in~ it is pointed out that the information supplementing $\psi$ \begin{quote} {\it [...] must be static, that is, its behavior cannot depend on where or when it is observed}. \end{quote} Otherwise said, the region of events corresponding to the acquisition of this information can be chosen to be space-like with respect to the events associated to $A$ and $B$. In~, this statement is presented as a simple remark and it does not constitute a new assumption; nonetheless, here we choose to denote it by $ST$, and to express it as $P_{CZ|ABXY} = P_{CZ}$. It turns out that \begin{equation} FW \wedge NS \wedge ST \Rightarrow FR. \end{equation} In fact, from $ST$ it follows that $P_{ABY|CZ} = P_{ABY}$; moreover we have \begin{equation} P_{ABY|CZ} = P_{A|BYCZ} P_{BY|CZ} = P_{A|BYCZ} P_{BY} \end{equation} by using again $ST$, but also \begin{equation} P_{ABY} = P_{AB} P_{Y|AB} = P_A P_B P_{Y|B} = P_A P_{BY} \end{equation} from $NS$ and $FW$. By comparing () and () we find that $P_{A|BYCZ} = P_A$, and a similar argument proves that $P_{B|AXCZ} = P_B$. Finally, $P_{C|AXBY} = P_C$ is a direct implication of $ST$. This result clearly shows that $FR$ corresponds to more than the free will of the observers. Notice that, if we assume $Z = \lambda$, we can prove also the converse of implication (), meaning that, if the ontic state would be completely accessible, $FR$ would be equivalent to the conjunction of the conditions $FW$, $NS$ and $ST$. For a detailed account of the role of the accessibility of $\lambda$ in this analysis, see~. Therefore, violation of $FR$ does not necessarily imply lack of free will as long as $ST$ or $NS$ are violated. This means that ontological models fully consistent with quantum mechanics, with the free will assumption (expressed through $FW$) and without superluminal communication are indeed possible, as long as the supplementary information on the ontic state is not static. Moreover, these models could be made of the $\psi$-epistemic type, for instance, by following the lines described in~. We want to comment about our condition $FW$. We do not consider it as the ultimate expression of the free will assumption, but only a meaningful substitute to $FR$, which enables us to raise our criticism to the $FR$ assumption. As $FR$, also $FW$ relies on conditional probabilities, and, in our opinion, this is not the most appropriate way to express the free will. Our target here is only to prove that the conclusions of~ are not appropriate, rather than providing an accurate mathematical expression of the free will assumption. Notice that the general approach to free will has been expressed by J.S. Bell as \begin{quote} {\it for me this means that the values of such variables have implications only in their future light cones}~, \end{quote} and, in our opinion, neither $FR$ nor $FW$ are able to properly express this fact, since lack of correlations is stronger than lack of implications. \section{Beyond quantum mechanics: the role of crypto-nonlocal hidden variables models} Following our reasoning, we conclude that the issue about completeness of quantum mechanics is still open, and similarly there are not conclusive conclusions concerning the ontological status of the vector $\psi$, which represents the state of the system in quantum mechanics. We now focus on the following question: could there be a theory, predictively equivalent to quantum mechanics, but experimentally distinguishable from it? In other words, would it be possible that a more refined knowledge of $\lambda$ could produce different outcomes (e.g. different statistics) from quantum mechanics, consistently with the fact that, if the information on the state reduces to that encoded in $\psi$, these outcomes are exactly those of quantum mechanics? Of course, one has to further constrain the theory in order to avoid physical inconsistencies. More explicitly, any information on $\lambda$ cannot be used to implement superluminal communication between distant parties. It turns out that this requirement is exactly addressed by the class of crypto-nonlocal hidden variable models, recently introduced by Leggett in a different context (investigation of non-locality and entanglement of correlated photons)~. In the simplified case where the ontic state, jointly with the settings $a$ and $b$ (the actual values of the random variables $A$ and $B$) determine the outcomes $x$ and $y$ (the actual values of the random variables $X$ and $Y$)~\footnote{In general, only the probabilities of these outcomes are determined.}, these models can be described as follows. We express $\lambda$ through two variables $(\mu, \tau)$, $\mu$ denoting the unaccessible part of the ontic state, and $\tau$ the accessible one. Now, we can write $\rho_{\psi} (\lambda) = \rho_{\psi, \tau} (\mu) \rho_{\psi} (\tau)$, and impose that knowledge of $\tau$ does not allow superluminal communication, that is \begin{eqnarray} \int x (a, b, \lambda) \rho_{\psi, \tau}(\mu) d \mu &=& f_{\psi}(a, \tau), \nonumber \\ \int y (a, b, \lambda) \rho_{\psi, \tau}(\mu) d \mu &=& g_{\psi}(b, \tau), \end{eqnarray} which are the so-called {\it non-signalling conditions}. The quantities $f_{\psi}(a, \tau)$ and $g_{\psi}(b , \tau)$ are the local averages of the theory at the intermediate level, that is, when the state of the system is described by $\tau$. Non-locality, which is apparent by the functional dependence $x(a,b,\lambda)$ and $y(a,b,\lambda)$, has been canceled out. As required, when we additionally average over $\tau$ we recover the quantum expectations, \begin{eqnarray} \int f_{\psi}(a, \tau) \rho_{\psi}(\tau) d \tau &=& \langle x(a) \rangle_{\psi}, \nonumber \\ \int g_{\psi}(b, \tau) \rho_{\psi}(\tau) d \tau &=& \langle y(b) \rangle_{\psi}. \end{eqnarray} The theory is experimentally distinguishable from quantum mechanics as long as $f_{\psi}(a, \tau) \ne \langle x(a) \rangle_{\psi}$ and/or $g_{\psi}(b, \tau) \ne \langle y(b) \rangle_{\psi}$. An example of this scheme has been recently described in~ for a pair of two qubits. It is a generalization of the famous Bell's model for the singlet state of a pair of two-level systems, valid for an arbitrary state $\psi$ written as \begin{equation} \vert \psi \rangle = \sin{\frac{\theta}{2}} \vert 00 \rangle + \cos{\frac{\theta}{2}} \vert 11 \rangle, \end{equation} with $\theta \in [0, \pi/2]$. If $\theta = 0$, $\vert \psi \rangle$ is a separable state state; if $\theta = \pi/2$ it is a maximally entangled state. The ontic state is given by the pair $\lambda = (\psi, \tilde{\lambda})$, where $\tilde{\lambda}$ is a unit vector in the 3-dimensional real space. By construction, the model is $\psi$-ontic, because different vectors $\psi$ are necessarily associated to different ontic states $\lambda$~\footnote{Nonetheless, a $\psi$-epistemic model can be obtained by suitably modifying this scheme.}. For specific values of the local settings, $A = a$ and $B = b$ ($a$ and $b$ are real, unit vectors), the local observables are given by $\sigma \cdot a$ and $\sigma \cdot b$, where $\sigma = (\sigma_x, \sigma_y,\sigma_z)$ is the vector of Pauli matrices. In particular, $\sigma_z$ is defined so that $\vert 0 \rangle$ and $\vert 1 \rangle$ are its $+1$ and $-1$ eigenvectors respectively. Without loss of generality we assume that $a$ and $b$ lie in the plane orthogonal to the direction of propagation of the entangled particles, assumed to depart from a common source. The local measurement outcomes are given by \begin{equation} x (a, b, \tilde{\lambda}) = \left\{ \begin{array}{ll} +1, & \hbox{\rm if $\hat{a} \cdot \tilde{\lambda} \geq\cos{\xi}$,} \\ -1, & \hbox{\rm if $\hat{a} \cdot \tilde{\lambda} < \cos{\xi}$,} \end{array} \right. \end{equation} and \begin{equation} y (b, \tilde{\lambda}) = \left\{ \begin{array}{ll} +1, & \hbox{\rm if $b \cdot \tilde{\lambda} \geq \cos{\chi}$,} \\ -1, & \hbox{\rm if $b \cdot \tilde{\lambda} < \cos{\chi}$.} \end{array} \right. \end{equation} In the previous relations, $\hat{a} = \hat{a} (a, b)$ is in the plane of $a$ and $b$, as detailed in~; moreover, $\cos{\xi} = - \langle x(a) \rangle_{\psi}$, and $\cos{\chi} = - \langle y(b) \rangle_{\psi}$. With the additional assumption that $\tilde{\lambda}$ is uniformly distributed on the unit sphere, it is possible to prove that this model is predictively equivalent to quantum mechanics (see~ for the details). Moreover, it belongs to the crypto-nonlocal family. We identify $(\mu, \tau)$ with the spherical coordinates of $\tilde{\lambda}$: $\mu$ is the azimuthal angle and $\tau$ the polar angle, and the north pole is identified by the direction of the incoming particle. It is easy to prove that $\rho_{\psi, \tau} (\mu) = 1 / 2 \pi$, and, by construction, integration over $\mu$ cancels non-locality in local averages. We find that~ \begin{equation} f_{\psi} (a, \tau) = \frac{1}{\pi} \, \cos^{-1}{\left(\frac{2 \langle x(a) \rangle_{\psi}^2}{\sin^2{\tau}} - 1 \right)} - 1, \end{equation} if $\vert \tau - \frac{\pi}{2} \vert \leq \xi$ and $f_{\psi} (a, \tau) = - 1$ otherwise, and a similar relation (with $\xi$ replaced by $\chi$) for $g_{\psi} (b, \tau)$. In general, $f_{\psi} (a, \tau) \ne \langle x(a) \rangle_{\psi}$ and $g_{\psi} (b, \tau) \ne \langle y(b) \rangle_{\psi}$. Therefore, despite the model is absolutely artificial, it provides evidence that models compatible with quantum mechanics, but in principle distinguishable from it, are indeed possible, without violating the free will assumption. In~, as a measure of the maximal departure from quantum expectations that crypto-nonlocal hidden variables models models can provide, we have used the variance of the variable $f_{\psi} (a, \tau)$ over the distribution $\rho_{\psi}(\tau)$: \begin{equation} \delta_{\psi} (a) = \int \Big( f_{\psi} (a, \tau) - \langle x (a) \rangle_{\psi} \Big)^2 \, \rho_{\psi}(\tau) d \tau. \end{equation} When the system consist of a pair of qubits, we have expressed the upper bound for this quantity for generic models as \begin{equation} \delta_{\psi} (a) \leq \cos{\theta} - \langle x (a) \rangle_{\psi}^2. \end{equation} In Fig.~ we show the dependence of this constraint by entanglement, and plot the corresponding curve for the specific model described in~. \section{Final remarks and conclusions} In this contribution we have criticized the form of the free will assumption which has been recently adopted to derive some striking results, in particular that (i) quantum mechanics is a complete theory, and (ii) the quantum state vector $\psi$ is necessarily in one-to-one correspondence with the elements of reality of the theory. To strengthen our argument, we have provided a different definition of free will, which makes clear that the former free will assumption actually encodes more than the observers' free choice (that is, the condition that the theory is non-signalling, and the staticity of the information supplementing $\psi$). We believe that both approaches to free will do not represent the correct necessary and sufficient condition for this assumption, and we conjecture that this condition cannot be simply expressed through simple expressions involving conditional probabilities. Therefore, ontological models which are compatible but possibly experimentally distinguishable from quantum mechanics are possible. Since these models should necessarily be non-signalling, we have suggested that they are given by the class of crypto-nonlocal hidden variables theories, and we have provided a simple example. Finally, we have described an upper bound for the local averages of any deterministic ontological theory for quantum mechanics in the case a pair of qubits. These results suggest that crypto-nonlocal hidden variables theories represent a relevant tool in the study of non-locality. \ack This research is partially supported by the ARO MURI grant W911NF-11-1-0268. \section*{References} \begin{thebibliography}{99} \bibitem{harrigan} N. Harrigan and R.W. Spekkens, Found. Phys. 40, 125 (2010) \bibitem{colbeck2} R. Colbeck and R. Renner, Nature Comm. 2, 411 (2011) \bibitem{colbeck} R. Colbeck and R. Renner, Phys. Rev. Lett. 108, 150402 (2012) \bibitem{pusey} M.F. Pusey, J. Barrett and T. Rudolph, Nature Phys. 8, 476 (2012) \bibitem{lewis} P.G. Lewis, D. Jennings, J. Barrett and T. Rudolph, Phys. Rev. Lett. 109, 150404 (2012) \bibitem{ghirardi} G.C. Ghirardi and R. Romano, Foundations of Physics: Volume 43, Issue 7 881 (2013) \bibitem{bell} J.S. Bell, {\it Free variables and local causality}, in {\it Speakable and unspeakable in quantum mechanics}, Cambridge University Press (1987) \bibitem{leggett} A.J. Leggett, Found. Phys. 33, 1469 (2003) \bibitem{ghirardi2} G.C. Ghirardi and R. Romano, Phys. Rev. A 85, 042102 (2012) \bibitem{ghirardi3} G.C. Ghirardi and R. Romano, Phys. Rev. Lett. 110, 170404 (2013) \end{thebibliography} |
1501.04131 | Title: Structure Learning and Statistical Estimation in Distribution Networks -
Part I
Abstract: Traditionally power distribution networks are either not observable or only
partially observable. This complicates development and implementation of new
smart grid technologies, such as those related to demand response, outage
detection and management, and improved load-monitoring. In this two part paper,
inspired by proliferation of metering technology, we discuss estimation
problems in structurally loopy but operationally radial distribution grids from
measurements, e.g. voltage data, which are either already available or can be
made available with a relatively minor investment. In Part I, the objective is
to learn the operational layout of the grid. Part II of this paper presents
algorithms that estimate load statistics or line parameters in addition to
learning the grid structure. Further, Part II discusses the problem of
structure estimation for systems with incomplete measurement sets. Our newly
suggested algorithms apply to a wide range of realistic scenarios. The
algorithms are also computationally efficient -- polynomial in time -- which is
proven theoretically and illustrated computationally on a number of test cases.
The technique developed can be applied to detect line failures in real time as
well as to understand the scope of possible adversarial attacks on the grid.
Body: \title{Structure Learning and Statistical Estimation in Distribution Networks - Part I} \author{\authorblockN{Deepjyoti~Deka*, Scott~Backhaus\dag, and Michael~Chertkov\ddag\\} \authorblockA{*Corresponding Author. Electrical \& Computer Engineering, University of Texas at Austin\\ \dag MPA Division, Los Alamos National Lab\\ \ddag Theory Division and the Center for Nonlinear Systems, Los Alamos National Lab} \thanks{D. Deka is with the Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX 78712. Email: deepjyotideka@utexas.edu} \thanks{S. Backhaus is with the MPA Division of LANL, Los Alamos, NM 87544. Email: *backhaus@lanl.gov} \thanks{M. Chertkov is with the Theory Division and the Center for Nonlinear Systems of LANL, Los Alamos, NM 87544. Email: *chertkov@lanl.gov}} \maketitle \begin{abstract} Traditionally power distribution networks are either not observable or only partially observable. This complicates development and implementation of new smart grid technologies, such as those related to demand response, outage detection and management, and improved load-monitoring. In this two part paper, inspired by proliferation of metering technology, we discuss estimation problems in structurally loopy but operationally radial distribution grids from measurements, e.g. voltage data, which are either already available or can be made available with a relatively minor investment. In Part I, the objective is to learn the operational layout of the grid. Part II of this paper presents algorithms that estimate load statistics or line parameters in addition to learning the grid structure. Further, Part II discusses the problem of structure estimation for systems with incomplete measurement sets. Our newly suggested algorithms apply to a wide range of realistic scenarios. The algorithms are also computationally efficient -- polynomial in time -- which is proven theoretically and illustrated computationally on a number of test cases. The technique developed can be applied to detect line failures in real time as well as to understand the scope of possible adversarial attacks on the grid. \end{abstract} \begin{IEEEkeywords} Power Distribution Networks, Power Flows, Struture/graph Learning, Voltage measurements, Transmission Lines. \end{IEEEkeywords} \section{Introduction} The power grid is composed of a network of transmission and distribution lines that enable the transfer of electrical power from generators to loads. The design, operation and control of these networks is typically hierarchical with a major division occurring between the transmission network of high voltage lines connecting sub-stations and power plants, and the distribution network of medium and low voltage lines that connect the transmission sub-stations to the end-users. Here, we focus on distribution networks. The design of distribution networks may appear to be loopy or meshed, however for practical engineering concerns, the vast majority of distribution grids are operated as "radial" networks, i.e. as a set of non-overlapping trees. Switches in the network are used to achieve one radial configuration out of many possibilities. Each tree in the network has a substation at the root and customers positioned at the other nodes. Switching from one tree-like operational configuration to another is typically caused by system upsets, e.g. faults and outages, and may occur few times a day or even an hour. The radial configuration distinguishes distribution networks from transmission networks that generally have multiple loops energized all the time to guarantee continuous delivery of power to every node, even in case of occasional line faults and outages. Radial configurations and one-way flow of power have led to much less monitoring, observability, and state estimation in distribution as compared to meshed transmission networks . The recent proliferation of smart grid technology, including smart meters that measure electricity consumption at the node level, is creating a new opportunities to extract information important to grid operators and planners. Such efforts are also getting additional attention in view of mounting concerns over data security and protection of user privacy . In this paper (Part I), we seek to \textbf{\textit{develop low-complexity algorithms to learn the current operational structure in `radial' distribution networks using only nodal measurements}}. Nodal measurements may include voltage magnitudes, voltage phase (potentially), and power injections and are typically available at smart meters, pole-mount or pad-mount transformers, and distribution phasor measurement units (PMU). Accurate structural estimation impacts many important applications including failure identification , outage management, and recovery following major and minor disruptions (e.g. hurricanes to individual lightning strikes), grid reconfiguration for power flow optimization and generation scheduling, and quantifying the need for additional meter placement. From an adversarial viewpoint, our work can be viewed as low-intrusion learning by a rogue agent interested in estimating the grid structure for a data attack . In the subsequent part (Part II), we will look at developing algorithms that are able to estimate the statistics of power consumption at the grid nodes or estimate the parameters of operational lines in addition to determining the grid's radial structure. Further, we will analyze learning the operational grid structure with missing data, where observations from a subset of nodes are not available. \subsection{Related Work} Our work falls in the broad category of `graph learning' problems that have been approached from different directions. For general graphs and graphical models , maximum-likelihood structure estimation has been researched in several papers by utilizing prior information such as sparsity of the parameter space , size of the graph neighborhood , etc. Techniques employed include both traditional convex optimization as well as greedy learning . For power grids, structure estimation techniques discussed in the literature can be classified based on the type of measurements available as well as assumptions made regarding grid structure and user behavior. In , a maximum likelihood estimator (MLE) with regularizers for low-rank and sparsity is used to recover the grid structure using locational marginal prices (LMPs). In , a model using bus phase angles as a Markov random field for the DC power flow builds a dependency graph based approach to detect faults in grids. In work specific to radial distribution grids, provides a structure identification algorithm that uses signs within the inverse covariance matrix (or concentration matrix) of voltage measurements to generate a minimum spanning tree. In , topology identification with limited measurements in a distribution grid with Gaussian loads is used to design a machine learning (ML) estimate with approximate schemes. Our work uses ordering of second moments, not a ML approach, to reconstruct a radial grid sequentially from the leaves to the root, making it distinct from previous work. Our algorithm design is based on a linear coupled approximation for lossless AC power flow that is idealized but practical for analyzing distribution grids where the line and voltage characteristics limit the accuracy of traditional approximations. Unlike related work, our topology learning algorithm is agnostic to the load profile distributions or variability in line impedances and requires only a less restrictive assumption on the correlation of load profiles. The rest of the paper is organized as follows. We start the next section with a description of the distribution grid and summary of the learning problems discussed in the manuscript. Section and Appendix describe the linear coupled (LC) power flow model and its special case, the DC-resistive power flow model. Statistical trends in observed nodal measurements are discussed in Section . Next, we use the derived results to design algorithms for learning the distribution grid structure using the power flow models in Section . Simulation results elucidating the performance of our algorithms on test distribution grids are presented in Section . Finally, Section concludes and suggests future directions that will be explored in a subsequent work. \section{Technical Preliminaries} The structure of a radial distribution network has important features that motivates our algorithm development. We discuss the radial structure in detail here and introduce the notation used in this paper. We then formulate the learning problem tackled in this paper in terms of its input data and deliverables and discuss the underlying motivation. \subsection{Structure of Radial Distribution Network} We consider a meshed distribution network which is operated as a union of non-intersecting `radial' trees, i.e. a spanning forest, by configuring switches as shown in Fig.~. There are exponentially many (in the number of switches) possible configurations of spanning forests. The grid-graph with all the switches closed is denoted ${\cal G}=({\cal V},{\cal E})$, where ${\cal V}$ is the set of nodes of the graph and ${\cal E}$ is the set of undirected edges of the graph. We denote nodes with single Roman letter subscripts $a$ and undirected edges with pairs of Roman letter subscripts $(ab)$. The operational grid is a forest denoted by ${\cal F}$ which spans all the nodes in ${\cal V}$. Specifically, ${\cal F}$ is a special subgraph of ${\cal G}$ (${\cal F}\subset {\cal G}$) such that \begin{itemize} \item ${\cal F}$ is a union of $K$ non-overlapping trees covering all the nodes of the graph \item Each tree contains exactly one of the $K$ bases (substations), ${\cal F}=\cup_{k=1,\cdots,K}{\cal T}_k$. \end{itemize} The distribution system ${\cal F}$ is a \textit{\textbf{`base-constrained spanning forest'}} with operational edges ${\cal E}^{\cal F}$ where ${\cal E}^{\cal F}\subset {\cal E}$. Table~ provides other relevant notations (nomenclature) used through out this manuscript to denote various nodal and edge features of the grid $\cal G$ and the operational forest $\cal F$. \begin{table}[ht] \caption{Notation Table} \begin{center} \begin{tabular}{|l|l|} \hline ${\cal F}\in{\cal G}$ & \begin{tabular}{|l} a particular forest configuration\\ of the physical distribution network ${\cal G}$ \end{tabular}\\ $\cal V$ & vertex set of ${\cal G}$\\ $N$ & $\#$ of nodes other than sub-stations in $\cal V$ \\ $K$ & number of sub-stations in the network \\ $\cal E$ & edge set of ${\cal G}$\\ ${\cal E}^{\cal F}$ & set of edges operational within $\cal F$ \\ ${\cal T}_k\in{\cal F}$ & \begin{tabular}{|l} tree within ${\cal F}$ containing\\ the $k^{th}$ sub-station\end{tabular} \\ $M_k$ & reduced incidence matrix of the tree ${\cal T}_k$ \\ ${\cal E}^{{\cal T}_k}$ & set of edges in ${\cal T}_k$ \\ ${\cal V}^{{\cal T}_k}$ & set of nodes in ${\cal T}_k$\\ path between $a, b\in {\cal V}^{{\cal T}_k}$ & \begin{tabular}{|l} subset of edges from ${\cal E}^{{\cal T}_k}$ s.t. each\\ node with edge in the subset, except\\ $a$ and $b$, contributes exactly two edges\end{tabular} \\ ${\cal E}_a^{{\cal T}_k}$ & path from $a$ to slack bus in ${{\cal T}_k}$ \\ $b$ is a descendant of $a$ & $a$ contributes ${\cal E}_b^{{\cal T}_k}$ \\ $b$ is the parent of $a$ & $(ab)\in {\cal E}_a^{{\cal T}_k}$ and $b$ contributes ${\cal E}_a^{{\cal T}_k}$ \\ $\theta_a, v_a$ & voltage phase and magnitude resp. at bus $a$ \\ $\theta, v$ & \begin{tabular}{|l} Vector of non-substation voltage\\ phases and magnitudes resp. \end{tabular}\\ $\varepsilon_a$ & $=1-v_a$, voltage deviation at node $a$\\ $\varepsilon$ & $=1-v$, voltage deviation \\ $p_a, q_a$ & \begin{tabular}{|l} resp. active and reactive power\\ injection/consumption ($+/-$) at bus $a$\end{tabular} \\ $p, q$ & \begin{tabular}{|l} Vector of non-substation \\ power injections/consumptions \end{tabular}\\ $\beta_{ab},g_{ab},r_{ab},x_{ab}$ & \begin{tabular}{|l} susceptance, conductance, resistance,\\ reactance resp. of edge $(ab)$ \end{tabular}\\ $\beta,g,r,x$ & \begin{tabular}{|l} diagonal matrix of line susceptances,\\ conductances, resistances, reactances resp.\end{tabular} \\ $\Sigma_y$ & matrix of second moments for variable $y$\\ $\Omega_y$ & matrix of covariances for variable $y$\\ $\mu_y$ & vector of means for variable $y$ \\ $H_y$ & \begin{tabular}{|l} reduced weighted Laplacian matrix\\ with edge weights in $y$ \end{tabular}\\ $D^{{\cal T}_k}_a$ & Descendants (including itself) of $a$ in ${\cal T}_k$\\ $\cal M$ & set of unobserved nodes\\ \hline \end{tabular} \end{center} \end{table} \subsection{Problem Formulation and Contribution} We consider large distribution grids where the utility (observer) is unsure of the grid configuration because of insufficient or inaccurate switching data, perhaps caused by a recent system upset. Alternatively, we could take the point of view of a third party observer, who may be an aggregator or adversary, trying to extract the current forest configuration from available nodal measurements. We assume the current spanning forest configuration is kept intact sufficiently long for load profiles at grid nodes to attain a steady distribution (longer than the fluctuations but shorter than changes in the mean load). We assume that the observer has access to nodal measurements, but not edge measurements -- an assumption consistent with the recent expansion of smart grid monitoring devices. Smart meters generally provide nodal voltages and power injections at fine spatial resolution, i.e. at the individual customer level, but they do not provide any edge flow data. Additional instrumentation is emerging for pole-mount or pad-mount transformers , however, these new devices still only provide nodal voltages and aggregated customer power injections. Some edge flow data is available to utilities, however, this is generally at a few select locations in the distribution grid, e.g. at the substation/root node, voltage regulators, reclosers, or other major utility equipment. These select locations may also have nodal and edge data from another emerging technology, i.e. distribution grid PMUs . However, we continue to restrict our input data to nodal values, which is consistent with the new, ubiquitous sensing provided by smart meters. The nodal devices provide the observer with temporal samples of the nodal voltage magnitudes. The observer seeks to use these samples to learn the current configuration of switches that determine the `base-constrained spanning forest'. To supplement the voltage magnitude samples, the observer has historical information about statistics of the nodal consumption. \section{Power Flow Models and Statistical Correlations} Our approach to the structure learning problem relies on linearized PF models on radial spanning forests that enable efficient reconstruction of the grid structure via a second-moment analysis. The most general of the two, termed the Linear Coupling (LC) model, ignores losses of active and reactive powers and consistently assumes small voltage magnitude and phase drops between connected nodes. For tree-like distribution grids, the LC-PF model becomes equivalent to the LinDistFlow PF model in . The second model considered in the paper, coined the DC-resistive model, corresponds to the special resistance dominating case of the LC-PF model. These PF models are described in more detail in Appendix . \subsection{Linear Coupled Power Flow (LC-PF) model} As noted in Eqs.~(,) in Appendix , the LC-PF model is derived from the general AC power flow model by assuming small voltage magnitude deviations and phase differences between neighboring buses in the grid. It is convenient to restate the linear equations in LC-PF model in matrix form as: \begin{eqnarray} && p= H_g\varepsilon+H_{\beta}\theta, ~~ q= H_{\beta}\varepsilon-H_g\theta \end{eqnarray} where $p,q,\varepsilon$ and $\theta$ are defined in Table . $H_g$ and $H_{\beta}$ are the weighted graph Laplacian matrices associated with forest ${\cal F}$ such that \begin{eqnarray} {\huge H}_g(a,b)&=\begin{cases}\sum_{c:(a,c) \in {\cal E}^{\cal F}}g_{ac} & \quad\text{if~} b = a\\ -g_{ab} & \quad\text{if~} (ab) \in {\cal E}^{\cal F}\\ 0 & \quad\text{otherwise}\end{cases} \end{eqnarray} $H_{\beta}$ has a similar structure with $g$-weights replaced by $\beta$-weights. The weighted graph Laplacians can be stated in terms of the directed incidence matrix $M$ as \begin{eqnarray} H_g = M^Tg^{\cal F}M, \quad H_{\beta} = M^T{\beta}^{\cal F}M \end{eqnarray} Here, $g^{\cal F}$ and ${\beta}^{\cal F}$ are diagonal matrices representing, respectively, line conductances and susceptances for edges within ${\cal F}$. $M$ is the edge to node directed incidence matrix of $\cal F$. See Fig. for an example. Every row $m_{ab}$ in $M$ is equal to $\pm(e_a^T -e_b^T)$ and represents the directed edge $(ab)$, where the direction of an edge is chosen arbitrarily. $e_a \in \mathbb{R}^{N+k}$ is the standard basis vector associated with the vertex $a$, with $1$ at the $a^{th}$ position and zero everywhere else. We can combine Eqs.~() and express the complex power flows as: \begin{eqnarray} p+\hat{i}q = M^T(g^{\cal F}+ \hat{i}{\beta}^{\cal F})M(\varepsilon - \hat{i}\theta) \end{eqnarray} Both $H_\beta$ and $H_g$ are weighted graph Laplacians and are degenerate --- showing $K$ zero-eigenvalues associated with the freedom in fixing phase and voltage deviation (from nominal) at any node within each tree of the forest. It is natural to fix phases and voltages at the sub-stations making these `slack buses' $a_k$ for trees ${\cal T}_k$ of the (operational) forest, ${\cal F}$ such that $\theta_{a_k}=\varepsilon_{a_k}=0, $ for any $k$, $1 \leq k \leq K$. Formally, elimination of the set of $K$ sub-stations corresponds to elimination of $K$ components from all the vectors contributing Eqs.~(), and reduction of $K$ rows and $K$ columns from the weighted Laplacian matrices. All the eigenvalues of the resulting reduced graph Laplacian matrices are thus strictly positive. Without loss of generality, we will use the same notation for the original and reduced dimension variables $\theta, \varepsilon, p$ and $q$ and also refer to Eqs.~(,) as applied to the reduced vectors of dimension $N\times 1$. We will also keep notations, $H_{\beta}$ and $H_g$ for the reduced graph Laplacian matrices, and $M$ for the reduced incidence matrix respectively. The reduced $M$ has a block diagonal structure: $M=\mbox{diag}(M_1,M_2,\cdots, M_K)$, where, $M_k$ is the invertible reduced incidence matrix of tree ${\cal T}_k$ in $\cal F$. Thus, $M$ and correspondingly $H_\beta$ and $H_g$ are full rank, invertible and block-diagonal matrices. Inverting the linear non-degenerate Eqs.~() we arrive at \begin{eqnarray} \theta=& M^{-1}x^{\cal F}{M^{-1}}^Tp -M^{-1}r^{\cal F}{M^{-1}}^Tq = H^{-1}_{1/x}p - H^{-1}_{1/r}q\\ \varepsilon=& M^{-1}r^{\cal F}{M^{-1}}^Tp +M^{-1}x^{\cal F}{M^{-1}}^Tq = H^{-1}_{1/r}p + H^{-1}_{1/x}q \end{eqnarray} where $H_{1/r} \doteq M^T{r^{\cal F}}^{-1}M$ and $H_{1/x} \doteq M^T{x^{\cal F}}^{-1}M$. $r^{\cal F}$ and $x^{\cal F}$ are diagonal matrices representing, respectively, line resistances and reactances within the forest ${\cal F}$. Their relation to $g^{\cal F}$ and ${\beta}^{\cal F}$ are expressed in Eqs.~(). \subsection{Relations between second moments} The real and reactive nodal power injections $p$ and $q$ in Eqs.~(,) fluctuate because of exogenous processes, and their second moments are related by: \begin{align} \mathbb{E}[\theta\theta^T] =& H^{-1}_{1/x}\mathbb{E}[pp^T]H^{-1}_{1/x} + H^{-1}_{1/r}\mathbb{E}[qq^T]H^{-1}_{1/r}\nonumber\\ &~- H^{-1}_{1/x}\mathbb{E}[pq^T]H^{-1}_{1/r}- H^{-1}_{1/r}\mathbb{E}[qp^T]H^{-1}_{1/x}\nonumber\\ \Rightarrow~ \Sigma_{\theta} =& H^{-1}_{1/x}\Sigma_pH^{-1}_{1/x} + H^{-1}_{1/r}\Sigma_qH^{-1}_{1/r} \nonumber\\ &~- H^{-1}_{1/x}\Sigma_{pq}H^{-1}_{1/r}- \left[H^{-1}_{1/x}\Sigma_{pq}H^{-1}_{1/r}\right]^T \\ \text{Similarly,}~~\Sigma_{\varepsilon} =& H^{-1}_{1/r}\Sigma_pH^{-1}_{1/r} + H^{-1}_{1/x}\Sigma_qH^{-1}_{1/x}\nonumber\\ &~+ H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}+\left[H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}\right]^T\\ \Sigma_{\theta\varepsilon} =& H^{-1}_{1/x}\Sigma_{p}H^{-1}_{1/r} - H^{-1}_{1/r}\Sigma_qH^{-1}_{1/x}\nonumber\\ &~+ H^{-1}_{1/x}\Sigma_{pq}H^{-1}_{1/x} - H^{-1}_{1/r}\Sigma_{qp}H^{-1}_{1/r} \end{align} These formulas are the basis for reconstruction/learning analysis in the rest of the paper. \subsection{DC-resistive model} The DC-resistive PF model (see Appendix ) is an extremal case of the LC-PF model realized when line reactance can be ignored in comparison with resistance ($x/r\to 0$). The relation between the statistics of active powers and voltage second-order moments deviations reduces to \begin{eqnarray} \mathbb{E}[pp^T] = H_g\mathbb{E}[\varepsilon\varepsilon^T]H_g \Rightarrow~ \Sigma_{\varepsilon} =H^{-1}_g\Sigma_pH^{-1}_g \end{eqnarray} \section{Trends in Second Moments over Tree Networks} We now derive key results related to the second moments in voltage magnitudes that arise from the properties of the forest $\cal F$. We denote the unique path from node $a$ to the slack bus in tree ${\cal T}_k$ by ${\cal E}_a^{{\cal T}_k}$. From , the inverse of the reduced incidence matrix of a tree has the following special structure: \begin{align} \squeezeup {\huge M}_k^{-1}(a,r)=\begin{cases}1 & \text{if edge $r\in {\cal E}_a^{{\cal T}_k}$ is directed}\\ &\text{along path from $a$ to slack bus},\\ -1 & \text{if edge $r\in {\cal E}_a^{{\cal T}_k}$ is directed}\\ &\text{against path from $a$ to slack bus}, \\ 0 & \text{if edge~} r \not\in {\cal E}_a^{{\cal T}_k} \end{cases} \squeezeup \end{align} Here, the direction of edge $r = (cd)$ is specified by its representative row $m_{cd}$ in the directed incidence matrix. For example, if $m_{cd} = e_c^T - e_d^T$, the edge is directed from node $c$ to node $d$, whereas for $m_{cd} = e_d^T - e_c^T$, the direction is from node $d$ to node $c$. An immediate corollary of () is that $M^{-1}(a,r) = 0$ if edge $r$ and node $a$ lie on separate trees within the forest $\cal F$, a fact consistent with the block diagonal structure of $M$. Using () in $H_g^{-1} = M^{-1}{g^{\cal F}}^{-1}{M^{-1}}^T$, we derive for forest $\cal F$ \begin{align} H_g^{-1}(a,b)&= 0 \text{~if $a,b$ are on different trees ${\cal T}_{k}$ and}\\ H_g^{-1}(a,b)&= \sum_{r \in {\cal E}^{{\cal T}_k}} M^{-1}(a,r){g^{{\cal F}}}^{-1}(r,r)M^{-1}(b,r) \text{~if~} a,b \in {\cal T}_k\nonumber\\ &= \sum_{(cd) \in {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_b^{{\cal T}_k}} \frac{1}{g_{cd}} \text{~if~} a,b \in {\cal T}_k \end{align} Thus, $H_g^{-1}(a,b)$ is equal to \textit{the sum of the inverse conductances of lines that are common to the paths from both nodes to the slack bus.} If no such line exists, the corresponding entry in $H_g^{-1}$ is $0$. See Fig.~ for illustration. Similar results hold for other measurement matrices like \begin{eqnarray} H^{-1}_{1/r}(a,b) &&=\begin{cases} \sum_{(cd) \in {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_b^{{\cal T}_k}} r_{cd} \text{~~if nodes~} a,b \in {\cal T}_k\\ 0 ~~ \text{otherwise,} \end{cases} \end{eqnarray} Let $D^{{\cal T}_k}_a$ denote the set of descendants of node $a$ within the tree ${\cal T}_k$. We call $b$ a descendent of $a$, if $a$ lies on the (unique) path from $b$ to the slack bus of ${\cal T}_k$, also including $a$ itself in the set of its descendants. We call $b$ the parent of $a$ within ${\cal T}_k$ (there can only be one) if $(ab)\in{\cal T}_k$ and $a$ is an immediate descendant of $b$ as illustrated in Fig. . The following statement holds. \begin{lemma} For two nodes, $a$ and its parent $b$, in tree ${\cal T}_k$ \begin{align} {\huge H}_g^{-1}(a,c)-{\huge H}_g^{-1}(b,c)&&=\begin{cases}\frac{1}{g_{ab}} & \quad\text{if node $c \in D^{{\cal T}_k}_a$}\\ 0 & \quad\text{otherwise,} \end{cases} \end{align} \end{lemma} \begin{proof} For any node $c$ which belongs to a tree not containing nodes $a$ and $b$, $H_g^{-1}(a,c)-H_g^{-1}(b,c) = 0$ according to (). Now, focus on nodes contained, together with $a$ and $b$, within ${\cal T}_k$. Since $b$ is $a$'s parent, ${\cal E}_a^{{\cal T}_k} = {\cal E}_b^{{\cal T}_k}\bigcup \{(ab)\}$ and we derive (also validating on the illustrative example in Fig.~) that for any node $c$ in tree ${\cal T}_k$, \begin{align*} {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k} &= {\cal E}_b^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k} \quad &&\text{if node $c \not\in D^{{\cal T}_k}_a$}\\ {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k} &= [{\cal E}_b^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k}] \bigcup \{(ab)\} \quad &&\text{if node $c \in D^{{\cal T}_k}_a$} \end{align*} resulting in Eq.~(). \end{proof} We now prove our main results regarding trends in second moments of deviations in voltage magnitudes ($\varepsilon$) along any tree in the network. The results are conditioned on the following assumption regarding correlations in power injections at the non-substation buses. \textbf{Assumption $1$}: For any two buses $a$ and $b$ drawing power from the same distribution sub-station, $\Sigma_p(a,b)>0, \Sigma_q(a,b)>0, \Sigma_{pq}(a,b)>0$. Note that this assumption holds, in particular, if the overall node balance is such that each non-substation node $a$ always consumes strictly more than it produces in active and reactive powers, i.e. $p_a <0, q_a <0$. This assumption is certainly true in any distribution grid with small and/or moderate penetration of renewables . However, the assumption is also reasonable for a system with significant penetration of generation which is still dominated in average by the consumption. \begin{theorem} If node $a \neq b$ is a descendant of node $b$ within tree ${\cal T}_k$, then for the DC-resistive model, $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(b,b)$. \end{theorem} \begin{proof} We first show that for any node $a$ and its parent $b$, $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(b,b)$. Consider $\Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b)$. From Eq.~(), we derive \begin{align} \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) =& \sum_{c,d}H_g^{-1}(a,c)\Sigma_p(c,d)\nonumber\\ &~~\left(H_g^{-1}(a,d)-H_g^{-1}(b,d)\right) \\ (\text{using Lemma })=& \sum_{c,d}H_g^{-1}(a,c)\Sigma_p(c,d)\frac{1}{g_{ab}}\textbf{1}(d \in D^{{\cal T}_k}_a) \nonumber\\ \Rightarrow \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) >& 0 ~(\text{using Assumption $1$})\\ \text{Also,}\Sigma_{\varepsilon}(a,b) - \Sigma_{\varepsilon}(b,b) &= \smashoperator[lr]{\sum_{c \in D^{{\cal T}_k}_a,d}}H_g^{-1}(b,d)\Sigma_p(c,d)\frac{1}{g_{ab}} > 0 \end{align} Combining Eqs.~() and () we derive $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(a,b) > \Sigma_{\varepsilon}(b,b)$. Since node $a$ is a descendant of node $b$, there is a path $a, c_1,...c_r,b$, such that each node in the path is a parent of its predecessor. Then, we derive $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(c_1,c_1) >...> \Sigma_{\varepsilon}(c_r,c_r) > \Sigma_{\varepsilon}(b,b)$. \end{proof} The following theorem is the LC-PF version of Theorem . \begin{theorem} If node $a \neq b$ is a descendant of node $b$ on tree ${\cal T}_k$ in forest $\cal F$, then $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(b,b)$ in the LC-PF model. \end{theorem} \begin{proof} Consider Eq.~(). Notice that the right hand side has four constituent terms ($H^{-1}_{1/r}\Sigma_pH^{-1}_{1/r}$, $H^{-1}_{1/x}\Sigma_qH^{-1}_{1/x}$, $H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}$ and $H^{-1}_{1/x}\Sigma_{qp}H^{-1}_{1/r}$). We denote each of these terms by $\Sigma^j_{\varepsilon}$ where $j \in \{1,2,3,4\}$. For each individual term, applying Assumption $1$ and the analysis in Theorem , we find that $\Sigma^j_{\varepsilon}(a,a) > \Sigma^j_{\varepsilon}(b,b)$ if node $a$ is a descendant of node $b$, other than itself. Thus, the statement also holds for the sum. \end{proof} We now focus on evaluating the term $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2]$, which is the expected value of the squared difference between two node voltage deviations ($\varepsilon$). For any two nodes $a$ and $b$, the DC-resistive model yields: \begin{align} \mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] &= \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) - \Sigma_{\varepsilon}(b,a) + \Sigma_{\varepsilon}(b,b)\nonumber\\ (\text{using ()}) &= \sum_{c,d}H_g^{-1}(a,c)\Sigma_p(c,d)\left(H_g^{-1}(a,d)- H_g^{-1}(b,d)\right)\nonumber\\ &~- \sum_{c,d}H_g^{-1}(b,c)\Sigma_p(c,d)\left(H_g^{-1}(a,d)- H_g^{-1}(b,d)\right)\nonumber\\ &= \sum_{c,d}\left(H_g^{-1}(a,c)- H_g^{-1}(b,c)\right)\Sigma_p(c,d)\nonumber\\ &\quad\quad~\left(H_g^{-1}(a,d)- H_g^{-1}(b,d)\right) \end{align} Our next Lemma follows directly by applying Lemma to Eq.~(). \begin{lemma} For two nodes, $a$ and its parent $b$ belonging to tree ${\cal T}_k$, in the DC-resistive model, we derive $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] = \sum_{c,d \in D^{{\cal T}_k}_a} \frac{1}{g_{ab}^2}\Sigma_p(c,d)$ \end{lemma} For the LC-PF model, we evaluate the expression $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2]$ as well. In this case, for two nodes, $a$ and its parent $b$, that lie on tree ${\cal T}_k$, we arrive at \squeezeup \begin{eqnarray} \mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] &&= \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) - \Sigma_{\varepsilon}(b,a) + \Sigma_{\varepsilon}(b,b)\nonumber \end{eqnarray} where $\Sigma_{\varepsilon}$ is given by Eq.~(). Let $\Sigma^1_{\varepsilon}$ and $\Sigma^2_{\varepsilon}$ represent the symmetric terms $H^{-1}_{1/r}\Sigma_pH^{-1}_{1/r}$ and $H^{-1}_{1/x}\Sigma_qH^{-1}_{1/x}$ respectively in $\Sigma_{\varepsilon}$. Extending the result of Lemma , we derive \begin{align} \Sigma^1_{\varepsilon}(a,a) -\Sigma^1_{\varepsilon}(a,b) - \Sigma^1_{\varepsilon}(b,a) + \Sigma^1_{\varepsilon}(b,b) = \smashoperator[lr]{\sum_{c,d \in D^{{\cal T}_k}_a}} r_{ab}^2\Sigma_p(c,d)\\ \Sigma^2_{\varepsilon}(a,a) -\Sigma^2_{\varepsilon}(a,b) - \Sigma^2_{\varepsilon}(b,a) + \Sigma^2_{\varepsilon}(b,b) = \smashoperator[r]{\sum_{c,d \in D^{{\cal T}_k}_a}} x_{ab}^2\Sigma_q(c,d) \end{align} Similarly, let $\Sigma^3_{\varepsilon}= H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}$ and $\Sigma^4_{\varepsilon}= {\Sigma^3}^T_{\varepsilon}= \left[H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}\right]^T$, the non-symmetric terms in $\Sigma_{\varepsilon}$. Using Lemma with Eq.~(), we get \begin{align} &\Sigma^3_{\varepsilon}(a,a) -\Sigma^3_{\varepsilon}(a,b) - \Sigma^3_{\varepsilon}(b,a) + \Sigma^3_{\varepsilon}(b,b) =\smashoperator[r]{\sum_{c,d\in D^{{\cal T}_k}_a}}r_{ab}\Sigma_{pq}(c,d)x_{ab}\nonumber\\ &=\Sigma^4_{\varepsilon}(a,a) -\Sigma^4_{\varepsilon}(a,b) - \Sigma^4_{\varepsilon}(b,a) + \Sigma^4_{\varepsilon}(b,b) \end{align} Combining Eqs.~(,,) we arrive at the following Lemma. \begin{lemma} In the LC-PF model, $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] = \sum_{c,d \in D^{{\cal T}_k}_a} r_{ab}^2\Sigma_p(c,d)+x_{ab}^2\Sigma_q(c,d)+2r_{ab}x_{ab}\Sigma_{pq}(c,d)$, holds for a node $a$ and its parent $b$ belonging to the (operational) tree ${\cal T}_k$. \end{lemma} \section{Learning Structure of Base-Constrained Spanning Forest} Here, we propose Algorithm $1$ to learn the structure of the distribution network using properties of voltage deviations for the LC-PF model. The polynomial time algorithm, based on the Theorems proved in the previous section, requires positivity of the correlation between nodal power injections (Assumption $1$). The Algorithm is also agnostic to the probability distribution of active and reactive power injections. To reconstruct the Base-Constrained Spanning Forest (${\cal F}=\cup_{k=1,\cdots,K}{\cal T}_k$), Algorithm $1$ takes as input `$m$' measurements of nodal voltage deviations. These measurements are used to create the empirical voltage deviation second moment matrix $\Sigma_{\varepsilon}$. The voltage deviations $\varepsilon^j_a$ at the substation nodes are assumed to be zero which implies the elements of the row $\Sigma_{\varepsilon}(a,:)$=0 for each of the $K$ substation nodes. The observer has prior information (or estimates using power injection measurements) for the true second moment matrix of power injections $\Sigma_p$ for non-substation nodes. \begin{algorithm} \caption{Base Constrained Spanning Forest Learning: LC-PF Model} \textbf{Input:} True $\Sigma_p, \Sigma_q$ and $\Sigma_{pq}$, $m$ voltage deviation observations $\varepsilon^j, 1\leq j \leq m$, all line resistances $r$ and reactances $x$\\ \begin{algorithmic}[1] \State Compute $\Sigma_{\varepsilon}(a,a) = \sum_{j = 1}^m\varepsilon^j_a\varepsilon^j_a/m$ for all nodes $a$. \State Undiscovered Set $U \gets \{1,2,...,N+K\}$, Leaf Set $L \gets \phi$, Descendant Sets $D_a \gets \{a\} \forall$ nodes $a$. \While {($U \neq \phi)$} \State $b^* \gets \max_{b \in U} \Sigma_{\varepsilon}(b,b)$ \ForAll{$a \in L$} \If {$\sum_{j=1}^m(\varepsilon^j_a-\varepsilon^j_{b^*})^2/m = \sum_{c,d \in D_a} r_{ab}^2\Sigma_p(c,d)+x_{ab}^2\Sigma_q(c,d)+2r_{ab}x_{ab}\Sigma_{pq}(c,d)$} \State Draw edge between nodes $a$ and $b^*$ \State $D_{b^*} \gets D_{b^*} \bigcup D_a$ \State $L \gets L - \{a\}$ \EndIf \EndFor \State $L \gets L \bigcup \{b^*\}$ \EndWhile \end{algorithmic} \end{algorithm} \textbf{Algorithm 1 Overview:} We reconstruct each tree within the distribution grid forest sequentially moving from the leaves to the root nodes. At every stage, $U$ represents the set of undiscovered nodes that are not part of the current reconstructed tree while $L$ represents the set of `current leaves' (nodes that are in the current reconstructed tree but with undiscovered parents). At each iteration, Step selects the node $b^*$ from set $U$ with the largest second moment of voltage deviation. Next Step adds edges between node $b^*$ and nodes in set $L$ of the growing tree using Lemma . In the ideal case when infinitely many voltage magnitude samples are collected, second-order moments of the power injections satisfy the relation in Lemma . However, we have a finite number of samples. Thus the presence of an edge is determined in Algorithm $1$ by checking if the relative difference between the reals on the left and right sides of the condition in Step is less than a predefined tolerance, $\tau$: \begin{align} 1 - \Biggl|\frac{\sum_{j=1}^m(\varepsilon^j_a-\varepsilon^j_{b^*})^2/m}{\sum_{c,d \in D_a} r_{ab}^2\Sigma_p(c,d)+x_{ab}^2\Sigma_q(c,d)+2r_{ab}x_{ab}\Sigma_{pq}(c,d)}\Biggr| < \tau \end{align} Steps and update the set of current leaves $L$ before repeating the reconstruction steps with a new undiscovered node.\\ \textbf{Algorithm Complexity:} Ignoring complexity of computing the second moments in Steps and (part of the data pre-processing), there are $N+K$ steps ($N$ load nodes and $K$ substation nodes) in the `while' loop, and at most $N$ comparisons in the `for' loop for each node. Therefore, the worst-case complexity of this algorithm is $O(N^2+ NK)$. Note that we use LC-PF model in Algorithm $1$. To design the DC-resistive version of Algorithm $1$, the condition in Step should be replaced with the result in Lemma . (Required modifications are straightforward and thus their description is omitted.) \section{Experiments} We perform a set of numerical experiments to test and demonstrate the performance of Algorithm $1$ in extracting the operational radial forest $\cal F$ from meshed ``as-designed'' distribution networks $\cal G$. We remind the reader that the observer in Algorithm $1$ has information of the full graph $\cal G$, the impedance (resistance and reactance) of all lines (operational or open) as well as the number of connected substation buses. Further, true second moments of active and reactive power injections at each non-substation node are assumed to be known. The set of measurements available as input with the observer comprises of deviations in voltage magnitudes ($\varepsilon$) at the grid nodes. Table summarizes the distribution grid test systems by the number of load busses, number of substation busses, and the number of tie switches. Additional information on these test systems can be found online at . In normal operation, each test grid consists of nodes in a forest- or tree-like configuration $\cal{F}$ with open tie-switches. We construct the complete meshed network $\cal{G}$ by closing all the tie-switches. To test the scalability of our algorithms, we increase the number of possible forest configuration by introducing several additional non-operational lines into each system as noted in Table . Although these contribute to $\cal{G}$, they are kept open and do not contribute to power flows in the operation forest $\cal{F}$. The impedances (reactances and resistances) of the additional lines are generated by assigning random values uniformly between the minimum and maximum impedances of the operational lines. Fig.~ displays the test networks $\cal{G}$ (solid and dashed lines) and the respective operational forests $\cal{F}$ (solid lines only). \begin{table}[ht] \caption{Summary of the tested distribution grids} \begin{center} \begin{tabular}{|p{1cm}|p{3cm}|p{2cm}|p{1cm}|} \hline Test Case & Number of buses / substations / tie-switches & Non-operational lines added & Source\\ \hline $bus\_13\_3$ & $13/3/3$ & $10$ & \\\hline $bus\_29\_1$ & $29/1/1$ & $20$ & \\\hline $bus\_83\_11$ & $83/11/13$ & $30$ & \\\hline \end{tabular} \end{center} \end{table} For each numerical experiment on a grid from Table , we pick an operational spanning forest $\cal{F}$ by opening tie switches. We also choose the statistics of the injections at each node using Gaussian distributions, unless otherwise specified. These distributions are used to generate multiple power injection samples, and from each vector-valued sample, we solve a PF to compute voltages and phases at every node in the network with the voltages at the substations fixed (i.e. they are slack busses). Averaging over all PF solutions, we compute empirical correlations of voltage magnitudes and phases. Using only these correlations, we run our algorithms and compare the resulting reconstruction with the actual operational configuration. In the reconstruction, we assume the observer has access to the resistance and reactance of all the lines in $\cal{G}$. All powers and voltages are presented in per unit (p.u.) values. Figs. - display the accuracy of Algorithm $1$ for the different test grids from Table . The relative error is defined to be the number of mislabeled lines (connected when actually open and vice versa) divided by the size of the operational edge set. The relative error is averaged by computing many reconstructions using the same nodal power injection distributions. Different curves (colors) in Fig.~ show the effect of changing the tolerance $\tau$ in Eq.~. The average fractional error in Figs. - decays exponentially with the number of samples used by the observer in the learning algorithm. For the tolerance values considered here, the majority of structural errors arise due to connected nodes not satisfying condition () and hence being labelled as open. The decay is then intuitive as an increase in the sample size makes empirical moments in voltage magnitudes approximate their true values better which in turn leads to an increase in the number of operational lines satisfying () and being correctly identified. On the other hand, if a sufficiently large value of $\tau$ is used, condition () will be relaxed and possibly be satisfied even by unconnected nodes. In such a case, a majority of errors will be recorded due to open lines being incorrectly labelled as operational. As errors of this type (open edges classified as operational) does not improve with the number of measurement samples, the average fractional errors will not decay with the sample size. This is elucidated in Fig. , where the $bus\_13\_3$ structure is learnt, in the presence of $50$ non-operational lines. Note that for larger values of $\tau$ in Fig. , the errors do not decay with the sample size whereas for smaller values, they do as justified in the preceding discussion. For larger sample sizes, a smaller value of $\tau$ is indeed preferable as observed in Fig. . \squeezeup \section{Conclusions \& Path Forward} Accurate structural estimation of distribution grids is important to many applications including failure identification, power flow optimization and estimation of state variables. In this manuscript, we have developed algorithms to learn the structure of a radial distribution grid using observed nodal voltage magnitude measurements. We have used a Linear Coupled (LC) approximation that relates complex nodal voltages to the complex power consumptions to prove that second moments of nodal voltage magnitudes in radial distribution grids follow certain statistical structure/ordering. Our algorithm relies on these results to reconstruct the operational tree in a bottom up fashion -- starting from the leaves and progressing to the root of the grid. The primary benefits of our approach are two-fold. First, our model is practical as voltage measurements are easily available at distribution grid nodes and individual devices. Second, the only assumption used regarding the statistics of the loads is the positivity of non-central correlation of nodal load profiles, which is natural for most distribution grids and more general than other assumptions discussed in the literature. We tested our algorithms on sample distribution grids and observed an exponential decay of average errors with increasing number of measurement samples. In the next part, we extend the work in this paper along the following directions: coupling structure learning with estimation of load statistics in the grid or estimation of line parameters in the grid, and learning grid structure even when measurement data is missing. Developing general algorithms for learning distribution grid operational forests that are not restricted to linearized power flow models remains a potential future direction of research. \appendix \section{Models of Power Flows} \subsection{Basic Power Flows} Note that the operational `base-constrained spanning forest' $\cal F$ can be thought of as a spanning tree over an extended graph, where an (artificial) super node is introduced and connected with (artificial) lines to all the sub-station nodes. This trick allows, without a loss of generality, to limit our discussion in the following to spanning trees, thus replacing ${\cal F}$ by ${\cal T}$. Let $z=(z_{ab}=r_{ab}+i x_{ab}|(ab)\in {\cal E})$ as the vector of complex line impedances in the grid. ($i^2=-1$.) Expressed in terms of the complex powers and potentials (voltages and phases) the Kirchoff laws over the operational (spanning tree) configuration ${\cal T}$ become \begin{align} \forall a\in{\cal V}: & P_a =p_a+i q_a=\underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum}\frac{v_a^2-v_a v_b\exp(i\theta_a-i\theta_b)}{z_{ab}^*} \end{align} \begin{comment} \begin{align} &= \underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum}\frac{r_{ab} +\hat{i}x_{ab}}{r_{ab}^2 +x_{ab}^2}(v_a^2 - v_av_b\cos(\theta_a-\theta_b)\nonumber\\ &\quad\quad\quad\quad\quad\quad~~~- \hat{i}v_av_b\sin(\theta_a-\theta_b)) \end{align} \end{comment} where the real valued scalars, $v_a$ and $\theta_a$, characterize voltage magnitude and phase respectively at node $a$. We assume that the power within the system considered is balanced through a slack bus, $a=0$. The set of Eqs.~(), expressing potentials via complex powers injected at the nodes of the power graph, are called \emph{Power Flow} (PF) equations. \subsection{Linear Coupled (LC) Approximation of Power Flows} We linearize the PF Eqs.~() in the first order jointly over phase difference and voltage deviations ($v_a -1=\varepsilon_a$) from nominal, i.e. the two corrections are considered on equal footing. We arrive at the following set of equations: \begin{align} p_a&=\underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum}\left(\beta_{ab}(\theta_a-\theta_b)+ g_{ab}(\varepsilon_a-\varepsilon_b)\right), \\ q_a&=\underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum} \left(-g_{ab}(\theta_a-\theta_b)+ \beta_{ab}(\varepsilon_a-\varepsilon_b)\right)\\ \text{where~} &\forall a\in{\cal V},\forall (ab)\in{\cal E}^{{\cal T}}:~|\varepsilon_a|\ll 1, |\theta_a-\theta_b| \ll 1,\\ &~~~~~~~~~~~~~~ g_{ab}\doteq\frac{r_{ab}}{x_{ab}^2+r_{ab}^2}, \beta_{ab}\doteq\frac{x_{ab}}{x_{ab}^2+r_{ab}^2} \end{align} Eqs.~(,) show coupling between phases and voltages, thus calling it the Linear-Coupled (LC) approximation, is proper. The linearity of LC-PF is used in the paper in deriving results to learn the grid structure. Two comments are in order. First of all, notice that the LC-PF approximation does not make any assumption about the relative strength of inductance and resistance, thus making it applicable to power distribution systems where the two line characteristics are typically of the same order. Second, expressions under the sum on the rhs of Eqs.~(,) represent active and reactive power flows which are antisymmetric ($p_{a\to b}=-p_{b\to a}$, $q_{a\to b}=-q_{b\to a}$). This emphasizes an important consequence of linear approximation in the LC-PF model -- both active and reactive losses in lines are ignored as such losses occur at second order. \subsection{DC-resistive approximation} In low-voltage distribution grids line inductances may be much smaller in magnitude than line resistances. Then the LC-PF model can be simplified even further. Indeed, taking this case to the extreme where inductance can be ignored in comparison with the resistance, $\forall \{a,b\}:\quad \beta_{ab}\ll g_{ab}$, we arrive at the following resistance-dominating version of Eqs.~(,) \begin{align} \forall a\in {\cal V}: p_a\approx\sum_{b:(ab)\in{\cal E}^{{\cal T}}} g_{ab}(\varepsilon_a-\varepsilon_b), q_a\approx\sum_{b:(ab)\in{\cal E}^{{\cal T}}}g_{ab}(\theta_b-\theta_a)\nonumber \end{align} We will coin this approximation DC-resistive PF. Its formulation is similar to the traditional DC flow model where active power flows are related to phase angles. The traditional DC model, used primarily for transmission networks, requires line inductances to dominate resistances, which is seldom observed in distribution grids. \subsection{From Power Flows to DistFlow and LinDistFlow} The DistFlow Eqs. introduced by Baran and Wu in are derived from the PF Eq.~() \begin{align} & p_{a\to b}-r_{ab}\frac{p_{a\to b}^2+q_{a\to b}^2}{v_a^2}=p_b+\sum_{(bc)\in{\cal E}^{{\cal T}};c\neq a} p_{b\to c}, \\ & q_{a\to b}-x_{ab}\frac{p_{a\to b}^2+q_{a\to b}^2}{v_a^2}=q_b+\sum_{(bc)\in{\cal E}^{{\cal T}};c\neq a} q_{b\to c}, \\ & v_b^2=v_a^2-2\left(r_{ab}p_{a\to b}+x_{ab} q_{a\to b}\right)+\left(r_{ab}^2+x_{ab}^2\right)\frac{p_{a\to b}^2+q_{a\to b}^2}{v_a^2} \end{align} where each of the three equations above are stated in terms of active, $p_{a\to b}$, and reactive, $q_{a\to b}$, powers over directed lines, $a\to b$, and voltages over nodes, $a\in{\cal V}$. If power losses at any line segment is negligible, Eqs.~(,,) reduce to . \begin{align} &p_{a\to b}\approx p_b+\sum_{\substack{(bc)\in{\cal E}^{{\cal T}}\\c \neq a}} p_{b\to c}, q_{a\to b}\approx q_b+\sum_{\substack{(bc)\in{\cal E}^{{\cal T}}\\c \neq a}} q_{b\to c}, \\ &\varphi_b\approx \varphi_a-2\left(r_{ab}p_{a\to b}+x_{ab} q_{a\to b}\right), \quad \varphi_a\equiv v_a^2 \end{align} On a general graph with loops, the number of these (directed) edge-related variables in Eq.~() is larger then the number of phases and voltages. However, the former becomes equal to the later if the grid is a tree, where the number of edges is equal to the number of nodes minus one. Therefore when the phase and voltage at one special node (each substation node per tree in our case) is fixed at $v_0$ (say), phases and voltages at all other nodes can be reconstructed from the directional line flows, and vice versa. Note that additional boundary conditions arise in a tree due to the requirement of active and reactive power flowing from/into any leaf being equal to its nodal injection. Thus, \begin{align} & \forall a\in{\cal V}_0:\quad v_a = v_0,\\ & \forall b\in V_l^{{\cal T}},\ \& \ (a\to b)\in{\cal E}^{{\cal T}}: p_{a\to b}=p_b,\quad q_{a\to b}=q_b, \end{align} where ${\cal V}_0\subset {\cal V}$ is the set of $K$ substations nodes (colored in red in Fig.~), and $V_l^{{\cal T}}$ is the set of leaf-nodes in tree ${\cal T}$. Based on the assumption that voltage drop across any line segment is sufficiently small, we linearize Eq. () by substituting $v_a=1+\varepsilon_a$, with $|\varepsilon_a|\ll 1$ and have \begin{equation} \forall (ab)\in{\cal E}^{{\cal T}}:\quad \varepsilon_a-\varepsilon_b\approx \left(r_{ab}p_{a\to b}+x_{ab} q_{a\to b}\right). \end{equation} Notice that the LinDistFlow Eqs.~(,), derived assuming that the grid graph is a tree, are exactly equivalent to Eqs.~(,) of the Linear Coupled (LC) Approximation. \section*{Acknowledgment} The work at LANL was funded by the Advanced Grid Modeling Program in the Office of Electricity in the US Department of Energy and was carried out under the auspices of the National Nuclear Security Administration of the U.S. Department of Energy at Los Alamos National Laboratory under Contract No. DE-AC52-06NA25396. \bibliographystyle{IEEETran} \bibliography{../../Bib/FIDVR,../../Bib/SmartGrid,../../Bib/voltage,../../Bib/trees} \begin{IEEEbiography}[{}]{Deepjyoti Deka} Deepjyoti Deka received his M.S. in Electrical Engineering from University of Texas, Austin in 2011, and his B.Tech in Electronics and Communication Engineering from IIT Guwahati, India, in 2009 for which he was awarded the Institute Silver Medal. He is currently a PhD candidate in Electrical Engineering at UT Austin. His research focusses on the design and analysis of power grid structure, operations and data security. He is also interested in modeling and optimization in social and physical networks. He has held internship positions at Los Alamos National Lab, Los Alamos NM, Electric Reliability Council of Texas, Taylor TX, and Qualcomm Inc, San Diego CA. \end{IEEEbiography} \vspace{-1.5cm} \begin{IEEEbiography}[{}]{ScottBackhaus} Scott Backhaus received the Ph.D. degree in physics from the University of California at Berkeley in 1997 in the area of experimental macroscopic quantum behavior of superfluid He-3 and He-4. In 1998, he came to Los Alamos, NM, was Director's Funded Postdoctoral Researcher from 1998 to 2000, a Reines Postdoctoral Fellow from 2001 to 2003, and a Technical Staff Member from 2003 to the present. While at Los Alamos, he has performed both experimental and theoretical research in the area of thermoacoustic energy conversion for which he received an R\&D 100 award in 1999 and Technology Review's Top 100 Innovators Under 35 [award in 2003]. Recently, his attention has shifted to other energy-related topics including the fundamental science of geologic carbon sequestration and grid-integration of renewable generation. \end{IEEEbiography} \vspace{-1.5cm} \begin{IEEEbiography}[{}]{Michael Chertkov} Dr. Chertkov's areas of interest include statistical and mathematical physics applied to energy and communication networks, machine learning, control theory, information theory, computer science, fluid mechanics and optics. Dr. Chertkov received his Ph.D. in physics from the Weizmann Institute of Science in 1996, and his M.Sc. in physics from Novosibirsk State University in 1990. After his Ph.D., Dr. Chertkov spent three years at Princeton University as a R.H. Dicke Fellow in the Department of Physics. He joined Los Alamos National Lab in 1999, initially as a J.R. Oppenheimer Fellow in the Theoretical Division. He is now a technical staff member in the same division. Dr. Chertkov has published more than 130 papers in these research areas. He is an editor of the Journal of Statistical Mechanics (JSTAT), associate editor of IEEE Transactions on Control of Network Systems, a fellow of the American Physical Society (APS), and a Founding Faculty Fellow of Skoltech (Moscow, Russia). \end{IEEEbiography} \title{Structure Learning and Statistical Estimation in Distribution Networks - Part I} \author{\authorblockN{Deepjyoti~Deka*, Scott~Backhaus\dag, and Michael~Chertkov\ddag\\} \authorblockA{*Corresponding Author. Electrical \& Computer Engineering, University of Texas at Austin\\ \dag MPA Division, Los Alamos National Lab\\ \ddag Theory Division and the Center for Nonlinear Systems, Los Alamos National Lab} \thanks{D. Deka is with the Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX 78712. Email: deepjyotideka@utexas.edu} \thanks{S. Backhaus is with the MPA Division of LANL, Los Alamos, NM 87544. Email: *backhaus@lanl.gov} \thanks{M. Chertkov is with the Theory Division and the Center for Nonlinear Systems of LANL, Los Alamos, NM 87544. Email: *chertkov@lanl.gov}} \maketitle \begin{abstract} Traditionally power distribution networks are either not observable or only partially observable. This complicates development and implementation of new smart grid technologies, such as those related to demand response, outage detection and management, and improved load-monitoring. In this two part paper, inspired by proliferation of metering technology, we discuss estimation problems in structurally loopy but operationally radial distribution grids from measurements, e.g. voltage data, which are either already available or can be made available with a relatively minor investment. In Part I, the objective is to learn the operational layout of the grid. Part II of this paper presents algorithms that estimate load statistics or line parameters in addition to learning the grid structure. Further, Part II discusses the problem of structure estimation for systems with incomplete measurement sets. Our newly suggested algorithms apply to a wide range of realistic scenarios. The algorithms are also computationally efficient -- polynomial in time -- which is proven theoretically and illustrated computationally on a number of test cases. The technique developed can be applied to detect line failures in real time as well as to understand the scope of possible adversarial attacks on the grid. \end{abstract} \begin{IEEEkeywords} Power Distribution Networks, Power Flows, Struture/graph Learning, Voltage measurements, Transmission Lines. \end{IEEEkeywords} \section{Introduction} The power grid is composed of a network of transmission and distribution lines that enable the transfer of electrical power from generators to loads. The design, operation and control of these networks is typically hierarchical with a major division occurring between the transmission network of high voltage lines connecting sub-stations and power plants, and the distribution network of medium and low voltage lines that connect the transmission sub-stations to the end-users. Here, we focus on distribution networks. The design of distribution networks may appear to be loopy or meshed, however for practical engineering concerns, the vast majority of distribution grids are operated as "radial" networks, i.e. as a set of non-overlapping trees. Switches in the network are used to achieve one radial configuration out of many possibilities. Each tree in the network has a substation at the root and customers positioned at the other nodes. Switching from one tree-like operational configuration to another is typically caused by system upsets, e.g. faults and outages, and may occur few times a day or even an hour. The radial configuration distinguishes distribution networks from transmission networks that generally have multiple loops energized all the time to guarantee continuous delivery of power to every node, even in case of occasional line faults and outages. Radial configurations and one-way flow of power have led to much less monitoring, observability, and state estimation in distribution as compared to meshed transmission networks . The recent proliferation of smart grid technology, including smart meters that measure electricity consumption at the node level, is creating a new opportunities to extract information important to grid operators and planners. Such efforts are also getting additional attention in view of mounting concerns over data security and protection of user privacy . In this paper (Part I), we seek to \textbf{\textit{develop low-complexity algorithms to learn the current operational structure in `radial' distribution networks using only nodal measurements}}. Nodal measurements may include voltage magnitudes, voltage phase (potentially), and power injections and are typically available at smart meters, pole-mount or pad-mount transformers, and distribution phasor measurement units (PMU). Accurate structural estimation impacts many important applications including failure identification , outage management, and recovery following major and minor disruptions (e.g. hurricanes to individual lightning strikes), grid reconfiguration for power flow optimization and generation scheduling, and quantifying the need for additional meter placement. From an adversarial viewpoint, our work can be viewed as low-intrusion learning by a rogue agent interested in estimating the grid structure for a data attack . In the subsequent part (Part II), we will look at developing algorithms that are able to estimate the statistics of power consumption at the grid nodes or estimate the parameters of operational lines in addition to determining the grid's radial structure. Further, we will analyze learning the operational grid structure with missing data, where observations from a subset of nodes are not available. \subsection{Related Work} Our work falls in the broad category of `graph learning' problems that have been approached from different directions. For general graphs and graphical models , maximum-likelihood structure estimation has been researched in several papers by utilizing prior information such as sparsity of the parameter space , size of the graph neighborhood , etc. Techniques employed include both traditional convex optimization as well as greedy learning . For power grids, structure estimation techniques discussed in the literature can be classified based on the type of measurements available as well as assumptions made regarding grid structure and user behavior. In , a maximum likelihood estimator (MLE) with regularizers for low-rank and sparsity is used to recover the grid structure using locational marginal prices (LMPs). In , a model using bus phase angles as a Markov random field for the DC power flow builds a dependency graph based approach to detect faults in grids. In work specific to radial distribution grids, provides a structure identification algorithm that uses signs within the inverse covariance matrix (or concentration matrix) of voltage measurements to generate a minimum spanning tree. In , topology identification with limited measurements in a distribution grid with Gaussian loads is used to design a machine learning (ML) estimate with approximate schemes. Our work uses ordering of second moments, not a ML approach, to reconstruct a radial grid sequentially from the leaves to the root, making it distinct from previous work. Our algorithm design is based on a linear coupled approximation for lossless AC power flow that is idealized but practical for analyzing distribution grids where the line and voltage characteristics limit the accuracy of traditional approximations. Unlike related work, our topology learning algorithm is agnostic to the load profile distributions or variability in line impedances and requires only a less restrictive assumption on the correlation of load profiles. The rest of the paper is organized as follows. We start the next section with a description of the distribution grid and summary of the learning problems discussed in the manuscript. Section and Appendix describe the linear coupled (LC) power flow model and its special case, the DC-resistive power flow model. Statistical trends in observed nodal measurements are discussed in Section . Next, we use the derived results to design algorithms for learning the distribution grid structure using the power flow models in Section . Simulation results elucidating the performance of our algorithms on test distribution grids are presented in Section . Finally, Section concludes and suggests future directions that will be explored in a subsequent work. \section{Technical Preliminaries} The structure of a radial distribution network has important features that motivates our algorithm development. We discuss the radial structure in detail here and introduce the notation used in this paper. We then formulate the learning problem tackled in this paper in terms of its input data and deliverables and discuss the underlying motivation. \subsection{Structure of Radial Distribution Network} We consider a meshed distribution network which is operated as a union of non-intersecting `radial' trees, i.e. a spanning forest, by configuring switches as shown in Fig.~. There are exponentially many (in the number of switches) possible configurations of spanning forests. The grid-graph with all the switches closed is denoted ${\cal G}=({\cal V},{\cal E})$, where ${\cal V}$ is the set of nodes of the graph and ${\cal E}$ is the set of undirected edges of the graph. We denote nodes with single Roman letter subscripts $a$ and undirected edges with pairs of Roman letter subscripts $(ab)$. The operational grid is a forest denoted by ${\cal F}$ which spans all the nodes in ${\cal V}$. Specifically, ${\cal F}$ is a special subgraph of ${\cal G}$ (${\cal F}\subset {\cal G}$) such that \begin{itemize} \item ${\cal F}$ is a union of $K$ non-overlapping trees covering all the nodes of the graph \item Each tree contains exactly one of the $K$ bases (substations), ${\cal F}=\cup_{k=1,\cdots,K}{\cal T}_k$. \end{itemize} The distribution system ${\cal F}$ is a \textit{\textbf{`base-constrained spanning forest'}} with operational edges ${\cal E}^{\cal F}$ where ${\cal E}^{\cal F}\subset {\cal E}$. Table~ provides other relevant notations (nomenclature) used through out this manuscript to denote various nodal and edge features of the grid $\cal G$ and the operational forest $\cal F$. \begin{table}[ht] \caption{Notation Table} \begin{center} \begin{tabular}{|l|l|} \hline ${\cal F}\in{\cal G}$ & \begin{tabular}{|l} a particular forest configuration\\ of the physical distribution network ${\cal G}$ \end{tabular}\\ $\cal V$ & vertex set of ${\cal G}$\\ $N$ & $\#$ of nodes other than sub-stations in $\cal V$ \\ $K$ & number of sub-stations in the network \\ $\cal E$ & edge set of ${\cal G}$\\ ${\cal E}^{\cal F}$ & set of edges operational within $\cal F$ \\ ${\cal T}_k\in{\cal F}$ & \begin{tabular}{|l} tree within ${\cal F}$ containing\\ the $k^{th}$ sub-station\end{tabular} \\ $M_k$ & reduced incidence matrix of the tree ${\cal T}_k$ \\ ${\cal E}^{{\cal T}_k}$ & set of edges in ${\cal T}_k$ \\ ${\cal V}^{{\cal T}_k}$ & set of nodes in ${\cal T}_k$\\ path between $a, b\in {\cal V}^{{\cal T}_k}$ & \begin{tabular}{|l} subset of edges from ${\cal E}^{{\cal T}_k}$ s.t. each\\ node with edge in the subset, except\\ $a$ and $b$, contributes exactly two edges\end{tabular} \\ ${\cal E}_a^{{\cal T}_k}$ & path from $a$ to slack bus in ${{\cal T}_k}$ \\ $b$ is a descendant of $a$ & $a$ contributes ${\cal E}_b^{{\cal T}_k}$ \\ $b$ is the parent of $a$ & $(ab)\in {\cal E}_a^{{\cal T}_k}$ and $b$ contributes ${\cal E}_a^{{\cal T}_k}$ \\ $\theta_a, v_a$ & voltage phase and magnitude resp. at bus $a$ \\ $\theta, v$ & \begin{tabular}{|l} Vector of non-substation voltage\\ phases and magnitudes resp. \end{tabular}\\ $\varepsilon_a$ & $=1-v_a$, voltage deviation at node $a$\\ $\varepsilon$ & $=1-v$, voltage deviation \\ $p_a, q_a$ & \begin{tabular}{|l} resp. active and reactive power\\ injection/consumption ($+/-$) at bus $a$\end{tabular} \\ $p, q$ & \begin{tabular}{|l} Vector of non-substation \\ power injections/consumptions \end{tabular}\\ $\beta_{ab},g_{ab},r_{ab},x_{ab}$ & \begin{tabular}{|l} susceptance, conductance, resistance,\\ reactance resp. of edge $(ab)$ \end{tabular}\\ $\beta,g,r,x$ & \begin{tabular}{|l} diagonal matrix of line susceptances,\\ conductances, resistances, reactances resp.\end{tabular} \\ $\Sigma_y$ & matrix of second moments for variable $y$\\ $\Omega_y$ & matrix of covariances for variable $y$\\ $\mu_y$ & vector of means for variable $y$ \\ $H_y$ & \begin{tabular}{|l} reduced weighted Laplacian matrix\\ with edge weights in $y$ \end{tabular}\\ $D^{{\cal T}_k}_a$ & Descendants (including itself) of $a$ in ${\cal T}_k$\\ $\cal M$ & set of unobserved nodes\\ \hline \end{tabular} \end{center} \end{table} \subsection{Problem Formulation and Contribution} We consider large distribution grids where the utility (observer) is unsure of the grid configuration because of insufficient or inaccurate switching data, perhaps caused by a recent system upset. Alternatively, we could take the point of view of a third party observer, who may be an aggregator or adversary, trying to extract the current forest configuration from available nodal measurements. We assume the current spanning forest configuration is kept intact sufficiently long for load profiles at grid nodes to attain a steady distribution (longer than the fluctuations but shorter than changes in the mean load). We assume that the observer has access to nodal measurements, but not edge measurements -- an assumption consistent with the recent expansion of smart grid monitoring devices. Smart meters generally provide nodal voltages and power injections at fine spatial resolution, i.e. at the individual customer level, but they do not provide any edge flow data. Additional instrumentation is emerging for pole-mount or pad-mount transformers , however, these new devices still only provide nodal voltages and aggregated customer power injections. Some edge flow data is available to utilities, however, this is generally at a few select locations in the distribution grid, e.g. at the substation/root node, voltage regulators, reclosers, or other major utility equipment. These select locations may also have nodal and edge data from another emerging technology, i.e. distribution grid PMUs . However, we continue to restrict our input data to nodal values, which is consistent with the new, ubiquitous sensing provided by smart meters. The nodal devices provide the observer with temporal samples of the nodal voltage magnitudes. The observer seeks to use these samples to learn the current configuration of switches that determine the `base-constrained spanning forest'. To supplement the voltage magnitude samples, the observer has historical information about statistics of the nodal consumption. \section{Power Flow Models and Statistical Correlations} Our approach to the structure learning problem relies on linearized PF models on radial spanning forests that enable efficient reconstruction of the grid structure via a second-moment analysis. The most general of the two, termed the Linear Coupling (LC) model, ignores losses of active and reactive powers and consistently assumes small voltage magnitude and phase drops between connected nodes. For tree-like distribution grids, the LC-PF model becomes equivalent to the LinDistFlow PF model in . The second model considered in the paper, coined the DC-resistive model, corresponds to the special resistance dominating case of the LC-PF model. These PF models are described in more detail in Appendix . \subsection{Linear Coupled Power Flow (LC-PF) model} As noted in Eqs.~(,) in Appendix , the LC-PF model is derived from the general AC power flow model by assuming small voltage magnitude deviations and phase differences between neighboring buses in the grid. It is convenient to restate the linear equations in LC-PF model in matrix form as: \begin{eqnarray} && p= H_g\varepsilon+H_{\beta}\theta, ~~ q= H_{\beta}\varepsilon-H_g\theta \end{eqnarray} where $p,q,\varepsilon$ and $\theta$ are defined in Table . $H_g$ and $H_{\beta}$ are the weighted graph Laplacian matrices associated with forest ${\cal F}$ such that \begin{eqnarray} {\huge H}_g(a,b)&=\begin{cases}\sum_{c:(a,c) \in {\cal E}^{\cal F}}g_{ac} & \quad\text{if~} b = a\\ -g_{ab} & \quad\text{if~} (ab) \in {\cal E}^{\cal F}\\ 0 & \quad\text{otherwise}\end{cases} \end{eqnarray} $H_{\beta}$ has a similar structure with $g$-weights replaced by $\beta$-weights. The weighted graph Laplacians can be stated in terms of the directed incidence matrix $M$ as \begin{eqnarray} H_g = M^Tg^{\cal F}M, \quad H_{\beta} = M^T{\beta}^{\cal F}M \end{eqnarray} Here, $g^{\cal F}$ and ${\beta}^{\cal F}$ are diagonal matrices representing, respectively, line conductances and susceptances for edges within ${\cal F}$. $M$ is the edge to node directed incidence matrix of $\cal F$. See Fig. for an example. Every row $m_{ab}$ in $M$ is equal to $\pm(e_a^T -e_b^T)$ and represents the directed edge $(ab)$, where the direction of an edge is chosen arbitrarily. $e_a \in \mathbb{R}^{N+k}$ is the standard basis vector associated with the vertex $a$, with $1$ at the $a^{th}$ position and zero everywhere else. We can combine Eqs.~() and express the complex power flows as: \begin{eqnarray} p+\hat{i}q = M^T(g^{\cal F}+ \hat{i}{\beta}^{\cal F})M(\varepsilon - \hat{i}\theta) \end{eqnarray} Both $H_\beta$ and $H_g$ are weighted graph Laplacians and are degenerate --- showing $K$ zero-eigenvalues associated with the freedom in fixing phase and voltage deviation (from nominal) at any node within each tree of the forest. It is natural to fix phases and voltages at the sub-stations making these `slack buses' $a_k$ for trees ${\cal T}_k$ of the (operational) forest, ${\cal F}$ such that $\theta_{a_k}=\varepsilon_{a_k}=0, $ for any $k$, $1 \leq k \leq K$. Formally, elimination of the set of $K$ sub-stations corresponds to elimination of $K$ components from all the vectors contributing Eqs.~(), and reduction of $K$ rows and $K$ columns from the weighted Laplacian matrices. All the eigenvalues of the resulting reduced graph Laplacian matrices are thus strictly positive. Without loss of generality, we will use the same notation for the original and reduced dimension variables $\theta, \varepsilon, p$ and $q$ and also refer to Eqs.~(,) as applied to the reduced vectors of dimension $N\times 1$. We will also keep notations, $H_{\beta}$ and $H_g$ for the reduced graph Laplacian matrices, and $M$ for the reduced incidence matrix respectively. The reduced $M$ has a block diagonal structure: $M=\mbox{diag}(M_1,M_2,\cdots, M_K)$, where, $M_k$ is the invertible reduced incidence matrix of tree ${\cal T}_k$ in $\cal F$. Thus, $M$ and correspondingly $H_\beta$ and $H_g$ are full rank, invertible and block-diagonal matrices. Inverting the linear non-degenerate Eqs.~() we arrive at \begin{eqnarray} \theta=& M^{-1}x^{\cal F}{M^{-1}}^Tp -M^{-1}r^{\cal F}{M^{-1}}^Tq = H^{-1}_{1/x}p - H^{-1}_{1/r}q\\ \varepsilon=& M^{-1}r^{\cal F}{M^{-1}}^Tp +M^{-1}x^{\cal F}{M^{-1}}^Tq = H^{-1}_{1/r}p + H^{-1}_{1/x}q \end{eqnarray} where $H_{1/r} \doteq M^T{r^{\cal F}}^{-1}M$ and $H_{1/x} \doteq M^T{x^{\cal F}}^{-1}M$. $r^{\cal F}$ and $x^{\cal F}$ are diagonal matrices representing, respectively, line resistances and reactances within the forest ${\cal F}$. Their relation to $g^{\cal F}$ and ${\beta}^{\cal F}$ are expressed in Eqs.~(). \subsection{Relations between second moments} The real and reactive nodal power injections $p$ and $q$ in Eqs.~(,) fluctuate because of exogenous processes, and their second moments are related by: \begin{align} \mathbb{E}[\theta\theta^T] =& H^{-1}_{1/x}\mathbb{E}[pp^T]H^{-1}_{1/x} + H^{-1}_{1/r}\mathbb{E}[qq^T]H^{-1}_{1/r}\nonumber\\ &~- H^{-1}_{1/x}\mathbb{E}[pq^T]H^{-1}_{1/r}- H^{-1}_{1/r}\mathbb{E}[qp^T]H^{-1}_{1/x}\nonumber\\ \Rightarrow~ \Sigma_{\theta} =& H^{-1}_{1/x}\Sigma_pH^{-1}_{1/x} + H^{-1}_{1/r}\Sigma_qH^{-1}_{1/r} \nonumber\\ &~- H^{-1}_{1/x}\Sigma_{pq}H^{-1}_{1/r}- \left[H^{-1}_{1/x}\Sigma_{pq}H^{-1}_{1/r}\right]^T \\ \text{Similarly,}~~\Sigma_{\varepsilon} =& H^{-1}_{1/r}\Sigma_pH^{-1}_{1/r} + H^{-1}_{1/x}\Sigma_qH^{-1}_{1/x}\nonumber\\ &~+ H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}+\left[H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}\right]^T\\ \Sigma_{\theta\varepsilon} =& H^{-1}_{1/x}\Sigma_{p}H^{-1}_{1/r} - H^{-1}_{1/r}\Sigma_qH^{-1}_{1/x}\nonumber\\ &~+ H^{-1}_{1/x}\Sigma_{pq}H^{-1}_{1/x} - H^{-1}_{1/r}\Sigma_{qp}H^{-1}_{1/r} \end{align} These formulas are the basis for reconstruction/learning analysis in the rest of the paper. \subsection{DC-resistive model} The DC-resistive PF model (see Appendix ) is an extremal case of the LC-PF model realized when line reactance can be ignored in comparison with resistance ($x/r\to 0$). The relation between the statistics of active powers and voltage second-order moments deviations reduces to \begin{eqnarray} \mathbb{E}[pp^T] = H_g\mathbb{E}[\varepsilon\varepsilon^T]H_g \Rightarrow~ \Sigma_{\varepsilon} =H^{-1}_g\Sigma_pH^{-1}_g \end{eqnarray} \section{Trends in Second Moments over Tree Networks} We now derive key results related to the second moments in voltage magnitudes that arise from the properties of the forest $\cal F$. We denote the unique path from node $a$ to the slack bus in tree ${\cal T}_k$ by ${\cal E}_a^{{\cal T}_k}$. From , the inverse of the reduced incidence matrix of a tree has the following special structure: \begin{align} \squeezeup {\huge M}_k^{-1}(a,r)=\begin{cases}1 & \text{if edge $r\in {\cal E}_a^{{\cal T}_k}$ is directed}\\ &\text{along path from $a$ to slack bus},\\ -1 & \text{if edge $r\in {\cal E}_a^{{\cal T}_k}$ is directed}\\ &\text{against path from $a$ to slack bus}, \\ 0 & \text{if edge~} r \not\in {\cal E}_a^{{\cal T}_k} \end{cases} \squeezeup \end{align} Here, the direction of edge $r = (cd)$ is specified by its representative row $m_{cd}$ in the directed incidence matrix. For example, if $m_{cd} = e_c^T - e_d^T$, the edge is directed from node $c$ to node $d$, whereas for $m_{cd} = e_d^T - e_c^T$, the direction is from node $d$ to node $c$. An immediate corollary of () is that $M^{-1}(a,r) = 0$ if edge $r$ and node $a$ lie on separate trees within the forest $\cal F$, a fact consistent with the block diagonal structure of $M$. Using () in $H_g^{-1} = M^{-1}{g^{\cal F}}^{-1}{M^{-1}}^T$, we derive for forest $\cal F$ \begin{align} H_g^{-1}(a,b)&= 0 \text{~if $a,b$ are on different trees ${\cal T}_{k}$ and}\\ H_g^{-1}(a,b)&= \sum_{r \in {\cal E}^{{\cal T}_k}} M^{-1}(a,r){g^{{\cal F}}}^{-1}(r,r)M^{-1}(b,r) \text{~if~} a,b \in {\cal T}_k\nonumber\\ &= \sum_{(cd) \in {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_b^{{\cal T}_k}} \frac{1}{g_{cd}} \text{~if~} a,b \in {\cal T}_k \end{align} Thus, $H_g^{-1}(a,b)$ is equal to \textit{the sum of the inverse conductances of lines that are common to the paths from both nodes to the slack bus.} If no such line exists, the corresponding entry in $H_g^{-1}$ is $0$. See Fig.~ for illustration. Similar results hold for other measurement matrices like \begin{eqnarray} H^{-1}_{1/r}(a,b) &&=\begin{cases} \sum_{(cd) \in {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_b^{{\cal T}_k}} r_{cd} \text{~~if nodes~} a,b \in {\cal T}_k\\ 0 ~~ \text{otherwise,} \end{cases} \end{eqnarray} Let $D^{{\cal T}_k}_a$ denote the set of descendants of node $a$ within the tree ${\cal T}_k$. We call $b$ a descendent of $a$, if $a$ lies on the (unique) path from $b$ to the slack bus of ${\cal T}_k$, also including $a$ itself in the set of its descendants. We call $b$ the parent of $a$ within ${\cal T}_k$ (there can only be one) if $(ab)\in{\cal T}_k$ and $a$ is an immediate descendant of $b$ as illustrated in Fig. . The following statement holds. \begin{lemma} For two nodes, $a$ and its parent $b$, in tree ${\cal T}_k$ \begin{align} {\huge H}_g^{-1}(a,c)-{\huge H}_g^{-1}(b,c)&&=\begin{cases}\frac{1}{g_{ab}} & \quad\text{if node $c \in D^{{\cal T}_k}_a$}\\ 0 & \quad\text{otherwise,} \end{cases} \end{align} \end{lemma} \begin{proof} For any node $c$ which belongs to a tree not containing nodes $a$ and $b$, $H_g^{-1}(a,c)-H_g^{-1}(b,c) = 0$ according to (). Now, focus on nodes contained, together with $a$ and $b$, within ${\cal T}_k$. Since $b$ is $a$'s parent, ${\cal E}_a^{{\cal T}_k} = {\cal E}_b^{{\cal T}_k}\bigcup \{(ab)\}$ and we derive (also validating on the illustrative example in Fig.~) that for any node $c$ in tree ${\cal T}_k$, \begin{align*} {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k} &= {\cal E}_b^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k} \quad &&\text{if node $c \not\in D^{{\cal T}_k}_a$}\\ {\cal E}_a^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k} &= [{\cal E}_b^{{\cal T}_k}\bigcap {\cal E}_c^{{\cal T}_k}] \bigcup \{(ab)\} \quad &&\text{if node $c \in D^{{\cal T}_k}_a$} \end{align*} resulting in Eq.~(). \end{proof} We now prove our main results regarding trends in second moments of deviations in voltage magnitudes ($\varepsilon$) along any tree in the network. The results are conditioned on the following assumption regarding correlations in power injections at the non-substation buses. \textbf{Assumption $1$}: For any two buses $a$ and $b$ drawing power from the same distribution sub-station, $\Sigma_p(a,b)>0, \Sigma_q(a,b)>0, \Sigma_{pq}(a,b)>0$. Note that this assumption holds, in particular, if the overall node balance is such that each non-substation node $a$ always consumes strictly more than it produces in active and reactive powers, i.e. $p_a <0, q_a <0$. This assumption is certainly true in any distribution grid with small and/or moderate penetration of renewables . However, the assumption is also reasonable for a system with significant penetration of generation which is still dominated in average by the consumption. \begin{theorem} If node $a \neq b$ is a descendant of node $b$ within tree ${\cal T}_k$, then for the DC-resistive model, $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(b,b)$. \end{theorem} \begin{proof} We first show that for any node $a$ and its parent $b$, $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(b,b)$. Consider $\Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b)$. From Eq.~(), we derive \begin{align} \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) =& \sum_{c,d}H_g^{-1}(a,c)\Sigma_p(c,d)\nonumber\\ &~~\left(H_g^{-1}(a,d)-H_g^{-1}(b,d)\right) \\ (\text{using Lemma })=& \sum_{c,d}H_g^{-1}(a,c)\Sigma_p(c,d)\frac{1}{g_{ab}}\textbf{1}(d \in D^{{\cal T}_k}_a) \nonumber\\ \Rightarrow \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) >& 0 ~(\text{using Assumption $1$})\\ \text{Also,}\Sigma_{\varepsilon}(a,b) - \Sigma_{\varepsilon}(b,b) &= \smashoperator[lr]{\sum_{c \in D^{{\cal T}_k}_a,d}}H_g^{-1}(b,d)\Sigma_p(c,d)\frac{1}{g_{ab}} > 0 \end{align} Combining Eqs.~() and () we derive $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(a,b) > \Sigma_{\varepsilon}(b,b)$. Since node $a$ is a descendant of node $b$, there is a path $a, c_1,...c_r,b$, such that each node in the path is a parent of its predecessor. Then, we derive $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(c_1,c_1) >...> \Sigma_{\varepsilon}(c_r,c_r) > \Sigma_{\varepsilon}(b,b)$. \end{proof} The following theorem is the LC-PF version of Theorem . \begin{theorem} If node $a \neq b$ is a descendant of node $b$ on tree ${\cal T}_k$ in forest $\cal F$, then $\Sigma_{\varepsilon}(a,a) > \Sigma_{\varepsilon}(b,b)$ in the LC-PF model. \end{theorem} \begin{proof} Consider Eq.~(). Notice that the right hand side has four constituent terms ($H^{-1}_{1/r}\Sigma_pH^{-1}_{1/r}$, $H^{-1}_{1/x}\Sigma_qH^{-1}_{1/x}$, $H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}$ and $H^{-1}_{1/x}\Sigma_{qp}H^{-1}_{1/r}$). We denote each of these terms by $\Sigma^j_{\varepsilon}$ where $j \in \{1,2,3,4\}$. For each individual term, applying Assumption $1$ and the analysis in Theorem , we find that $\Sigma^j_{\varepsilon}(a,a) > \Sigma^j_{\varepsilon}(b,b)$ if node $a$ is a descendant of node $b$, other than itself. Thus, the statement also holds for the sum. \end{proof} We now focus on evaluating the term $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2]$, which is the expected value of the squared difference between two node voltage deviations ($\varepsilon$). For any two nodes $a$ and $b$, the DC-resistive model yields: \begin{align} \mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] &= \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) - \Sigma_{\varepsilon}(b,a) + \Sigma_{\varepsilon}(b,b)\nonumber\\ (\text{using ()}) &= \sum_{c,d}H_g^{-1}(a,c)\Sigma_p(c,d)\left(H_g^{-1}(a,d)- H_g^{-1}(b,d)\right)\nonumber\\ &~- \sum_{c,d}H_g^{-1}(b,c)\Sigma_p(c,d)\left(H_g^{-1}(a,d)- H_g^{-1}(b,d)\right)\nonumber\\ &= \sum_{c,d}\left(H_g^{-1}(a,c)- H_g^{-1}(b,c)\right)\Sigma_p(c,d)\nonumber\\ &\quad\quad~\left(H_g^{-1}(a,d)- H_g^{-1}(b,d)\right) \end{align} Our next Lemma follows directly by applying Lemma to Eq.~(). \begin{lemma} For two nodes, $a$ and its parent $b$ belonging to tree ${\cal T}_k$, in the DC-resistive model, we derive $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] = \sum_{c,d \in D^{{\cal T}_k}_a} \frac{1}{g_{ab}^2}\Sigma_p(c,d)$ \end{lemma} For the LC-PF model, we evaluate the expression $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2]$ as well. In this case, for two nodes, $a$ and its parent $b$, that lie on tree ${\cal T}_k$, we arrive at \squeezeup \begin{eqnarray} \mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] &&= \Sigma_{\varepsilon}(a,a) - \Sigma_{\varepsilon}(a,b) - \Sigma_{\varepsilon}(b,a) + \Sigma_{\varepsilon}(b,b)\nonumber \end{eqnarray} where $\Sigma_{\varepsilon}$ is given by Eq.~(). Let $\Sigma^1_{\varepsilon}$ and $\Sigma^2_{\varepsilon}$ represent the symmetric terms $H^{-1}_{1/r}\Sigma_pH^{-1}_{1/r}$ and $H^{-1}_{1/x}\Sigma_qH^{-1}_{1/x}$ respectively in $\Sigma_{\varepsilon}$. Extending the result of Lemma , we derive \begin{align} \Sigma^1_{\varepsilon}(a,a) -\Sigma^1_{\varepsilon}(a,b) - \Sigma^1_{\varepsilon}(b,a) + \Sigma^1_{\varepsilon}(b,b) = \smashoperator[lr]{\sum_{c,d \in D^{{\cal T}_k}_a}} r_{ab}^2\Sigma_p(c,d)\\ \Sigma^2_{\varepsilon}(a,a) -\Sigma^2_{\varepsilon}(a,b) - \Sigma^2_{\varepsilon}(b,a) + \Sigma^2_{\varepsilon}(b,b) = \smashoperator[r]{\sum_{c,d \in D^{{\cal T}_k}_a}} x_{ab}^2\Sigma_q(c,d) \end{align} Similarly, let $\Sigma^3_{\varepsilon}= H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}$ and $\Sigma^4_{\varepsilon}= {\Sigma^3}^T_{\varepsilon}= \left[H^{-1}_{1/r}\Sigma_{pq}H^{-1}_{1/x}\right]^T$, the non-symmetric terms in $\Sigma_{\varepsilon}$. Using Lemma with Eq.~(), we get \begin{align} &\Sigma^3_{\varepsilon}(a,a) -\Sigma^3_{\varepsilon}(a,b) - \Sigma^3_{\varepsilon}(b,a) + \Sigma^3_{\varepsilon}(b,b) =\smashoperator[r]{\sum_{c,d\in D^{{\cal T}_k}_a}}r_{ab}\Sigma_{pq}(c,d)x_{ab}\nonumber\\ &=\Sigma^4_{\varepsilon}(a,a) -\Sigma^4_{\varepsilon}(a,b) - \Sigma^4_{\varepsilon}(b,a) + \Sigma^4_{\varepsilon}(b,b) \end{align} Combining Eqs.~(,,) we arrive at the following Lemma. \begin{lemma} In the LC-PF model, $\mathbb{E}[(\varepsilon_a-\varepsilon_b)^2] = \sum_{c,d \in D^{{\cal T}_k}_a} r_{ab}^2\Sigma_p(c,d)+x_{ab}^2\Sigma_q(c,d)+2r_{ab}x_{ab}\Sigma_{pq}(c,d)$, holds for a node $a$ and its parent $b$ belonging to the (operational) tree ${\cal T}_k$. \end{lemma} \section{Learning Structure of Base-Constrained Spanning Forest} Here, we propose Algorithm $1$ to learn the structure of the distribution network using properties of voltage deviations for the LC-PF model. The polynomial time algorithm, based on the Theorems proved in the previous section, requires positivity of the correlation between nodal power injections (Assumption $1$). The Algorithm is also agnostic to the probability distribution of active and reactive power injections. To reconstruct the Base-Constrained Spanning Forest (${\cal F}=\cup_{k=1,\cdots,K}{\cal T}_k$), Algorithm $1$ takes as input `$m$' measurements of nodal voltage deviations. These measurements are used to create the empirical voltage deviation second moment matrix $\Sigma_{\varepsilon}$. The voltage deviations $\varepsilon^j_a$ at the substation nodes are assumed to be zero which implies the elements of the row $\Sigma_{\varepsilon}(a,:)$=0 for each of the $K$ substation nodes. The observer has prior information (or estimates using power injection measurements) for the true second moment matrix of power injections $\Sigma_p$ for non-substation nodes. \begin{algorithm} \caption{Base Constrained Spanning Forest Learning: LC-PF Model} \textbf{Input:} True $\Sigma_p, \Sigma_q$ and $\Sigma_{pq}$, $m$ voltage deviation observations $\varepsilon^j, 1\leq j \leq m$, all line resistances $r$ and reactances $x$\\ \begin{algorithmic}[1] \State Compute $\Sigma_{\varepsilon}(a,a) = \sum_{j = 1}^m\varepsilon^j_a\varepsilon^j_a/m$ for all nodes $a$. \State Undiscovered Set $U \gets \{1,2,...,N+K\}$, Leaf Set $L \gets \phi$, Descendant Sets $D_a \gets \{a\} \forall$ nodes $a$. \While {($U \neq \phi)$} \State $b^* \gets \max_{b \in U} \Sigma_{\varepsilon}(b,b)$ \ForAll{$a \in L$} \If {$\sum_{j=1}^m(\varepsilon^j_a-\varepsilon^j_{b^*})^2/m = \sum_{c,d \in D_a} r_{ab}^2\Sigma_p(c,d)+x_{ab}^2\Sigma_q(c,d)+2r_{ab}x_{ab}\Sigma_{pq}(c,d)$} \State Draw edge between nodes $a$ and $b^*$ \State $D_{b^*} \gets D_{b^*} \bigcup D_a$ \State $L \gets L - \{a\}$ \EndIf \EndFor \State $L \gets L \bigcup \{b^*\}$ \EndWhile \end{algorithmic} \end{algorithm} \textbf{Algorithm 1 Overview:} We reconstruct each tree within the distribution grid forest sequentially moving from the leaves to the root nodes. At every stage, $U$ represents the set of undiscovered nodes that are not part of the current reconstructed tree while $L$ represents the set of `current leaves' (nodes that are in the current reconstructed tree but with undiscovered parents). At each iteration, Step selects the node $b^*$ from set $U$ with the largest second moment of voltage deviation. Next Step adds edges between node $b^*$ and nodes in set $L$ of the growing tree using Lemma . In the ideal case when infinitely many voltage magnitude samples are collected, second-order moments of the power injections satisfy the relation in Lemma . However, we have a finite number of samples. Thus the presence of an edge is determined in Algorithm $1$ by checking if the relative difference between the reals on the left and right sides of the condition in Step is less than a predefined tolerance, $\tau$: \begin{align} 1 - \Biggl|\frac{\sum_{j=1}^m(\varepsilon^j_a-\varepsilon^j_{b^*})^2/m}{\sum_{c,d \in D_a} r_{ab}^2\Sigma_p(c,d)+x_{ab}^2\Sigma_q(c,d)+2r_{ab}x_{ab}\Sigma_{pq}(c,d)}\Biggr| < \tau \end{align} Steps and update the set of current leaves $L$ before repeating the reconstruction steps with a new undiscovered node.\\ \textbf{Algorithm Complexity:} Ignoring complexity of computing the second moments in Steps and (part of the data pre-processing), there are $N+K$ steps ($N$ load nodes and $K$ substation nodes) in the `while' loop, and at most $N$ comparisons in the `for' loop for each node. Therefore, the worst-case complexity of this algorithm is $O(N^2+ NK)$. Note that we use LC-PF model in Algorithm $1$. To design the DC-resistive version of Algorithm $1$, the condition in Step should be replaced with the result in Lemma . (Required modifications are straightforward and thus their description is omitted.) \section{Experiments} We perform a set of numerical experiments to test and demonstrate the performance of Algorithm $1$ in extracting the operational radial forest $\cal F$ from meshed ``as-designed'' distribution networks $\cal G$. We remind the reader that the observer in Algorithm $1$ has information of the full graph $\cal G$, the impedance (resistance and reactance) of all lines (operational or open) as well as the number of connected substation buses. Further, true second moments of active and reactive power injections at each non-substation node are assumed to be known. The set of measurements available as input with the observer comprises of deviations in voltage magnitudes ($\varepsilon$) at the grid nodes. Table summarizes the distribution grid test systems by the number of load busses, number of substation busses, and the number of tie switches. Additional information on these test systems can be found online at . In normal operation, each test grid consists of nodes in a forest- or tree-like configuration $\cal{F}$ with open tie-switches. We construct the complete meshed network $\cal{G}$ by closing all the tie-switches. To test the scalability of our algorithms, we increase the number of possible forest configuration by introducing several additional non-operational lines into each system as noted in Table . Although these contribute to $\cal{G}$, they are kept open and do not contribute to power flows in the operation forest $\cal{F}$. The impedances (reactances and resistances) of the additional lines are generated by assigning random values uniformly between the minimum and maximum impedances of the operational lines. Fig.~ displays the test networks $\cal{G}$ (solid and dashed lines) and the respective operational forests $\cal{F}$ (solid lines only). \begin{table}[ht] \caption{Summary of the tested distribution grids} \begin{center} \begin{tabular}{|p{1cm}|p{3cm}|p{2cm}|p{1cm}|} \hline Test Case & Number of buses / substations / tie-switches & Non-operational lines added & Source\\ \hline $bus\_13\_3$ & $13/3/3$ & $10$ & \\\hline $bus\_29\_1$ & $29/1/1$ & $20$ & \\\hline $bus\_83\_11$ & $83/11/13$ & $30$ & \\\hline \end{tabular} \end{center} \end{table} For each numerical experiment on a grid from Table , we pick an operational spanning forest $\cal{F}$ by opening tie switches. We also choose the statistics of the injections at each node using Gaussian distributions, unless otherwise specified. These distributions are used to generate multiple power injection samples, and from each vector-valued sample, we solve a PF to compute voltages and phases at every node in the network with the voltages at the substations fixed (i.e. they are slack busses). Averaging over all PF solutions, we compute empirical correlations of voltage magnitudes and phases. Using only these correlations, we run our algorithms and compare the resulting reconstruction with the actual operational configuration. In the reconstruction, we assume the observer has access to the resistance and reactance of all the lines in $\cal{G}$. All powers and voltages are presented in per unit (p.u.) values. Figs. - display the accuracy of Algorithm $1$ for the different test grids from Table . The relative error is defined to be the number of mislabeled lines (connected when actually open and vice versa) divided by the size of the operational edge set. The relative error is averaged by computing many reconstructions using the same nodal power injection distributions. Different curves (colors) in Fig.~ show the effect of changing the tolerance $\tau$ in Eq.~. The average fractional error in Figs. - decays exponentially with the number of samples used by the observer in the learning algorithm. For the tolerance values considered here, the majority of structural errors arise due to connected nodes not satisfying condition () and hence being labelled as open. The decay is then intuitive as an increase in the sample size makes empirical moments in voltage magnitudes approximate their true values better which in turn leads to an increase in the number of operational lines satisfying () and being correctly identified. On the other hand, if a sufficiently large value of $\tau$ is used, condition () will be relaxed and possibly be satisfied even by unconnected nodes. In such a case, a majority of errors will be recorded due to open lines being incorrectly labelled as operational. As errors of this type (open edges classified as operational) does not improve with the number of measurement samples, the average fractional errors will not decay with the sample size. This is elucidated in Fig. , where the $bus\_13\_3$ structure is learnt, in the presence of $50$ non-operational lines. Note that for larger values of $\tau$ in Fig. , the errors do not decay with the sample size whereas for smaller values, they do as justified in the preceding discussion. For larger sample sizes, a smaller value of $\tau$ is indeed preferable as observed in Fig. . \squeezeup \section{Conclusions \& Path Forward} Accurate structural estimation of distribution grids is important to many applications including failure identification, power flow optimization and estimation of state variables. In this manuscript, we have developed algorithms to learn the structure of a radial distribution grid using observed nodal voltage magnitude measurements. We have used a Linear Coupled (LC) approximation that relates complex nodal voltages to the complex power consumptions to prove that second moments of nodal voltage magnitudes in radial distribution grids follow certain statistical structure/ordering. Our algorithm relies on these results to reconstruct the operational tree in a bottom up fashion -- starting from the leaves and progressing to the root of the grid. The primary benefits of our approach are two-fold. First, our model is practical as voltage measurements are easily available at distribution grid nodes and individual devices. Second, the only assumption used regarding the statistics of the loads is the positivity of non-central correlation of nodal load profiles, which is natural for most distribution grids and more general than other assumptions discussed in the literature. We tested our algorithms on sample distribution grids and observed an exponential decay of average errors with increasing number of measurement samples. In the next part, we extend the work in this paper along the following directions: coupling structure learning with estimation of load statistics in the grid or estimation of line parameters in the grid, and learning grid structure even when measurement data is missing. Developing general algorithms for learning distribution grid operational forests that are not restricted to linearized power flow models remains a potential future direction of research. \appendix \section{Models of Power Flows} \subsection{Basic Power Flows} Note that the operational `base-constrained spanning forest' $\cal F$ can be thought of as a spanning tree over an extended graph, where an (artificial) super node is introduced and connected with (artificial) lines to all the sub-station nodes. This trick allows, without a loss of generality, to limit our discussion in the following to spanning trees, thus replacing ${\cal F}$ by ${\cal T}$. Let $z=(z_{ab}=r_{ab}+i x_{ab}|(ab)\in {\cal E})$ as the vector of complex line impedances in the grid. ($i^2=-1$.) Expressed in terms of the complex powers and potentials (voltages and phases) the Kirchoff laws over the operational (spanning tree) configuration ${\cal T}$ become \begin{align} \forall a\in{\cal V}: & P_a =p_a+i q_a=\underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum}\frac{v_a^2-v_a v_b\exp(i\theta_a-i\theta_b)}{z_{ab}^*} \end{align} \begin{comment} \begin{align} &= \underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum}\frac{r_{ab} +\hat{i}x_{ab}}{r_{ab}^2 +x_{ab}^2}(v_a^2 - v_av_b\cos(\theta_a-\theta_b)\nonumber\\ &\quad\quad\quad\quad\quad\quad~~~- \hat{i}v_av_b\sin(\theta_a-\theta_b)) \end{align} \end{comment} where the real valued scalars, $v_a$ and $\theta_a$, characterize voltage magnitude and phase respectively at node $a$. We assume that the power within the system considered is balanced through a slack bus, $a=0$. The set of Eqs.~(), expressing potentials via complex powers injected at the nodes of the power graph, are called \emph{Power Flow} (PF) equations. \subsection{Linear Coupled (LC) Approximation of Power Flows} We linearize the PF Eqs.~() in the first order jointly over phase difference and voltage deviations ($v_a -1=\varepsilon_a$) from nominal, i.e. the two corrections are considered on equal footing. We arrive at the following set of equations: \begin{align} p_a&=\underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum}\left(\beta_{ab}(\theta_a-\theta_b)+ g_{ab}(\varepsilon_a-\varepsilon_b)\right), \\ q_a&=\underset{b:(ab)\in{\cal E}^{{\cal T}}}{\sum} \left(-g_{ab}(\theta_a-\theta_b)+ \beta_{ab}(\varepsilon_a-\varepsilon_b)\right)\\ \text{where~} &\forall a\in{\cal V},\forall (ab)\in{\cal E}^{{\cal T}}:~|\varepsilon_a|\ll 1, |\theta_a-\theta_b| \ll 1,\\ &~~~~~~~~~~~~~~ g_{ab}\doteq\frac{r_{ab}}{x_{ab}^2+r_{ab}^2}, \beta_{ab}\doteq\frac{x_{ab}}{x_{ab}^2+r_{ab}^2} \end{align} Eqs.~(,) show coupling between phases and voltages, thus calling it the Linear-Coupled (LC) approximation, is proper. The linearity of LC-PF is used in the paper in deriving results to learn the grid structure. Two comments are in order. First of all, notice that the LC-PF approximation does not make any assumption about the relative strength of inductance and resistance, thus making it applicable to power distribution systems where the two line characteristics are typically of the same order. Second, expressions under the sum on the rhs of Eqs.~(,) represent active and reactive power flows which are antisymmetric ($p_{a\to b}=-p_{b\to a}$, $q_{a\to b}=-q_{b\to a}$). This emphasizes an important consequence of linear approximation in the LC-PF model -- both active and reactive losses in lines are ignored as such losses occur at second order. \subsection{DC-resistive approximation} In low-voltage distribution grids line inductances may be much smaller in magnitude than line resistances. Then the LC-PF model can be simplified even further. Indeed, taking this case to the extreme where inductance can be ignored in comparison with the resistance, $\forall \{a,b\}:\quad \beta_{ab}\ll g_{ab}$, we arrive at the following resistance-dominating version of Eqs.~(,) \begin{align} \forall a\in {\cal V}: p_a\approx\sum_{b:(ab)\in{\cal E}^{{\cal T}}} g_{ab}(\varepsilon_a-\varepsilon_b), q_a\approx\sum_{b:(ab)\in{\cal E}^{{\cal T}}}g_{ab}(\theta_b-\theta_a)\nonumber \end{align} We will coin this approximation DC-resistive PF. Its formulation is similar to the traditional DC flow model where active power flows are related to phase angles. The traditional DC model, used primarily for transmission networks, requires line inductances to dominate resistances, which is seldom observed in distribution grids. \subsection{From Power Flows to DistFlow and LinDistFlow} The DistFlow Eqs. introduced by Baran and Wu in are derived from the PF Eq.~() \begin{align} & p_{a\to b}-r_{ab}\frac{p_{a\to b}^2+q_{a\to b}^2}{v_a^2}=p_b+\sum_{(bc)\in{\cal E}^{{\cal T}};c\neq a} p_{b\to c}, \\ & q_{a\to b}-x_{ab}\frac{p_{a\to b}^2+q_{a\to b}^2}{v_a^2}=q_b+\sum_{(bc)\in{\cal E}^{{\cal T}};c\neq a} q_{b\to c}, \\ & v_b^2=v_a^2-2\left(r_{ab}p_{a\to b}+x_{ab} q_{a\to b}\right)+\left(r_{ab}^2+x_{ab}^2\right)\frac{p_{a\to b}^2+q_{a\to b}^2}{v_a^2} \end{align} where each of the three equations above are stated in terms of active, $p_{a\to b}$, and reactive, $q_{a\to b}$, powers over directed lines, $a\to b$, and voltages over nodes, $a\in{\cal V}$. If power losses at any line segment is negligible, Eqs.~(,,) reduce to . \begin{align} &p_{a\to b}\approx p_b+\sum_{\substack{(bc)\in{\cal E}^{{\cal T}}\\c \neq a}} p_{b\to c}, q_{a\to b}\approx q_b+\sum_{\substack{(bc)\in{\cal E}^{{\cal T}}\\c \neq a}} q_{b\to c}, \\ &\varphi_b\approx \varphi_a-2\left(r_{ab}p_{a\to b}+x_{ab} q_{a\to b}\right), \quad \varphi_a\equiv v_a^2 \end{align} On a general graph with loops, the number of these (directed) edge-related variables in Eq.~() is larger then the number of phases and voltages. However, the former becomes equal to the later if the grid is a tree, where the number of edges is equal to the number of nodes minus one. Therefore when the phase and voltage at one special node (each substation node per tree in our case) is fixed at $v_0$ (say), phases and voltages at all other nodes can be reconstructed from the directional line flows, and vice versa. Note that additional boundary conditions arise in a tree due to the requirement of active and reactive power flowing from/into any leaf being equal to its nodal injection. Thus, \begin{align} & \forall a\in{\cal V}_0:\quad v_a = v_0,\\ & \forall b\in V_l^{{\cal T}},\ \& \ (a\to b)\in{\cal E}^{{\cal T}}: p_{a\to b}=p_b,\quad q_{a\to b}=q_b, \end{align} where ${\cal V}_0\subset {\cal V}$ is the set of $K$ substations nodes (colored in red in Fig.~), and $V_l^{{\cal T}}$ is the set of leaf-nodes in tree ${\cal T}$. Based on the assumption that voltage drop across any line segment is sufficiently small, we linearize Eq. () by substituting $v_a=1+\varepsilon_a$, with $|\varepsilon_a|\ll 1$ and have \begin{equation} \forall (ab)\in{\cal E}^{{\cal T}}:\quad \varepsilon_a-\varepsilon_b\approx \left(r_{ab}p_{a\to b}+x_{ab} q_{a\to b}\right). \end{equation} Notice that the LinDistFlow Eqs.~(,), derived assuming that the grid graph is a tree, are exactly equivalent to Eqs.~(,) of the Linear Coupled (LC) Approximation. \section*{Acknowledgment} The work at LANL was funded by the Advanced Grid Modeling Program in the Office of Electricity in the US Department of Energy and was carried out under the auspices of the National Nuclear Security Administration of the U.S. Department of Energy at Los Alamos National Laboratory under Contract No. DE-AC52-06NA25396. \bibliographystyle{IEEETran} \bibliography{../../Bib/FIDVR,../../Bib/SmartGrid,../../Bib/voltage,../../Bib/trees} \begin{IEEEbiography}[{}]{Deepjyoti Deka} Deepjyoti Deka received his M.S. in Electrical Engineering from University of Texas, Austin in 2011, and his B.Tech in Electronics and Communication Engineering from IIT Guwahati, India, in 2009 for which he was awarded the Institute Silver Medal. He is currently a PhD candidate in Electrical Engineering at UT Austin. His research focusses on the design and analysis of power grid structure, operations and data security. He is also interested in modeling and optimization in social and physical networks. He has held internship positions at Los Alamos National Lab, Los Alamos NM, Electric Reliability Council of Texas, Taylor TX, and Qualcomm Inc, San Diego CA. \end{IEEEbiography} \vspace{-1.5cm} \begin{IEEEbiography}[{}]{ScottBackhaus} Scott Backhaus received the Ph.D. degree in physics from the University of California at Berkeley in 1997 in the area of experimental macroscopic quantum behavior of superfluid He-3 and He-4. In 1998, he came to Los Alamos, NM, was Director's Funded Postdoctoral Researcher from 1998 to 2000, a Reines Postdoctoral Fellow from 2001 to 2003, and a Technical Staff Member from 2003 to the present. While at Los Alamos, he has performed both experimental and theoretical research in the area of thermoacoustic energy conversion for which he received an R\&D 100 award in 1999 and Technology Review's Top 100 Innovators Under 35 [award in 2003]. Recently, his attention has shifted to other energy-related topics including the fundamental science of geologic carbon sequestration and grid-integration of renewable generation. \end{IEEEbiography} \vspace{-1.5cm} \begin{IEEEbiography}[{}]{Michael Chertkov} Dr. Chertkov's areas of interest include statistical and mathematical physics applied to energy and communication networks, machine learning, control theory, information theory, computer science, fluid mechanics and optics. Dr. Chertkov received his Ph.D. in physics from the Weizmann Institute of Science in 1996, and his M.Sc. in physics from Novosibirsk State University in 1990. After his Ph.D., Dr. Chertkov spent three years at Princeton University as a R.H. Dicke Fellow in the Department of Physics. He joined Los Alamos National Lab in 1999, initially as a J.R. Oppenheimer Fellow in the Theoretical Division. He is now a technical staff member in the same division. Dr. Chertkov has published more than 130 papers in these research areas. He is an editor of the Journal of Statistical Mechanics (JSTAT), associate editor of IEEE Transactions on Control of Network Systems, a fellow of the American Physical Society (APS), and a Founding Faculty Fellow of Skoltech (Moscow, Russia). \end{IEEEbiography} |
1501.04132 | Title: IFC Inside: Retrofitting Languages with Dynamic Information Flow Control
(Extended Version)
Abstract: Many important security problems in JavaScript, such as browser extension
security, untrusted JavaScript libraries and safe integration of mutually
distrustful websites (mash-ups), may be effectively addressed using an
efficient implementation of information flow control (IFC). Unfortunately
existing fine-grained approaches to JavaScript IFC require modifications to the
language semantics and its engine, a non-goal for browser applications. In this
work, we take the ideas of coarse-grained dynamic IFC and provide the
theoretical foundation for a language-based approach that can be applied to any
programming language for which external effects can be controlled. We then
apply this formalism to server- and client-side JavaScript, show how it
generalizes to the C programming language, and connect it to the Haskell LIO
system. Our methodology offers design principles for the construction of
information flow control systems when isolation can easily be achieved, as well
as compositional proofs for optimized concrete implementations of these
systems, by relating them to their isolated variants.
Body: \newcommand{\flows}{\sqsubseteq} \newcommand{\lub}{\sqcup} \newcommand{\glb}{\sqcap} \newcommand{\ifc}[1]{\ensuremath{{\color{blue} #1}}} \newcommand{\lcurr}{\ensuremath{\ifc{l_{\textrm{cur}}}}} \newcommand{\tar}[1]{\ensuremath{\mathbf{\color{red} #1}}} \newcommand{\Coloneqq}{::=} \newcommand{\dom}[1]{\ensuremath{{\textrm{dom}} #1}} \newcommand{\fresh}[1]{\ensuremath{\textrm{fresh}(#1)}} \title{ IFC Inside: Retrofitting Languages with Dynamic Information Flow Control } \ifextended \subtitle{Extended Version} \fi \author{ Stefan Heule\inst{1} \and Deian Stefan\inst{1} \and Edward Z. Yang\inst{1} \and John C. Mitchell\inst{1} \and Alejandro Russo\inst{2}\protect } \institute{Stanford University \and Chalmers University } \maketitle \begin{abstract} Many important security problems in JavaScript, such as browser extension security, untrusted JavaScript libraries and safe integration of mutually distrustful websites (mash-ups), may be effectively addressed using an efficient implementation of information flow control (IFC). Unfortunately existing fine-grained approaches to JavaScript IFC require modifications to the language semantics and its engine, a non-goal for browser applications. In this work, we take the ideas of coarse-grained dynamic IFC and provide the theoretical foundation for a language-based approach that can be applied to any programming language for which external effects can be controlled. We then apply this formalism to server- and client-side JavaScript, show how it generalizes to the C programming language, and connect it to the Haskell LIO system. Our methodology offers design principles for the construction of information flow control systems when isolation can easily be achieved, as well as compositional proofs for optimized concrete implementations of these systems, by relating them to their isolated variants. \end{abstract} \section{Introduction} Modern web content is rendered using a potentially large number of different components with differing provenance. Disparate and untrusting components may arise from browser extensions (whose JavaScript code runs alongside website code), web applications (with possibly untrusted third-party libraries), and mashups (which combine code and data from websites that may not even be aware of each other's existence.) While just-in-time combination of untrusting components offers great flexibility, it also poses complex security challenges. In particular, maintaining data privacy in the face of malicious extensions, libraries, and mashup components has been difficult. Information flow control (IFC) is a promising technique that provides security by tracking the flow of sensitive data through a system. Untrusted code is confined so that it cannot exfiltrate data, except as per an information flow policy. Significant research has been devoted to adding various forms of IFC to different kinds of programming languages and systems. In the context of the web, however, there is a strong motivation to preserve JavaScript's semantics and avoid JavaScript-engine modifications, while retrofitting it with dynamic information flow control. The Operating Systems community has tackled this challenge (e.g., in~) by taking a \textit{coarse-grained} approach to IFC: dividing an application into coarse computational units, each with a single label dictating its security policy, and only monitoring communication between them. This coarse-grained approach provides a number of advantages when compared to the fine-grained approaches typically employed by language-based systems. First, adding IFC does not require intrusive changes to an existing programming language, thereby also allowing the reuse of existing programs. Second, it has a small runtime overhead because checks need only be performed at isolation boundaries instead of (almost) every program instruction~(e.g.,~). Finally, associating a single security label with the entire computational unit simplifies understanding and reasoning about the security guarantees of the system, without reasoning about most of the technical details of the semantics of the underlying programming language. In this paper, we present a framework which brings coarse-grained IFC ideas into a language-based setting: an information flow control system should be thought of as multiple instances of completely isolated language runtimes or \emph{tasks}, with information flow control applied to inter-task communication. We describe a formal system in which an IFC system can be designed once and then applied to any programming language which has control over external effects (e.g., JavaScript or C with access to hardware privilege separation). We formalize this system using an approach by Matthews and Findler~ for combining operational semantics and prove non-interference guarantees that are independent of the choice of a specific target language. There are a number of points that distinguish this setting from previous coarse-grained IFC systems. First, even though the underlying semantic model involves communicating tasks, these tasks can be coordinated together in ways that simulate features of traditional languages. In fact, simulating features in this way is a useful \emph{design tool} for discovering what variants of the features are permissible and which are not. Second, although completely separate tasks are semantically easy to reason about, real-world implementations often blur the lines between tasks in the name of efficiency. Characterizing what optimizations are permissible is subtle, since removing transitions from the operational semantics of a language can break non-interference. We partially address this issue by characterizing isomorphisms between the operational semantics of our abstract language and a concrete implementation, showing that if this relationship holds, then non-interference in the abstract specification carries over to the concrete implementation. Our contributions can be summarized as follows: \vspace*{-0.3em} \begin{itemize} \item We give formal semantics for a core coarse-grained dynamic information flow control language free of non-IFC constructs. We then show how a large class of target languages can be combined with this IFC language and prove that the result provides non-interference. (Sections~ and ) \item We provide a proof technique to show the non-interference of a concrete semantics for a potentially optimized IFC language by means of an isomorphism and show a class of restrictions on the IFC language that preserves non-interference. (Section~) \item We have implemented an IFC system based on these semantics for Node.js, and we connect our formalism to another implementation based on this work for client-side JavaScript~. Furthermore, we outline an implementation for the C programming language and describe improvements to the Haskell LIO system that resulted from this framework. (Section~) \end{itemize} \ifextended \else In the extended version of this paper we give all the relevant proofs and extend our IFC language with additional features~. \fi \cut{ (Something about the importance of IFC). One barrier to the adoption of information flow control has been the fact that it often incurs a large performance cost. This usually stems from the fact that most existing programming languages do not have facilities for enforcing isolation. Thus information flow control checks must be applied at a very fine-grained level, e.g.\ all values in the system must be labeled, resulting in large overhead for ordinary operations. Motivated by these problems, there has been increasing interest in coarse-grained IFC systems, which trade-off precision for reduced overhead. These systems are characterized by floating label associated with a thread of execution, so that access to labeled data taints the entire thread, solving the problem of flow-sensitivity. Typically, these systems require strong isolation between threads, which previously has been enforced by the type system. (LIO) This has made this method difficult to apply to languages which are unable to statically enforce such strong isolation. In this paper, we describe a simple but general methodology for using isolation of \emph{execution contexts}, e.g.\ an instance of the JavaScript engine, in order to add information flow control to an existing language. This is a very practical approach, as there are many languages which have built-in capabilities for strong isolation by forking an execution context (i.e. Web Workers). To validate this methodology, we describe its application to Haskell (LIO), to JavaScript (Browbound), to C (HipStar) and to Java (Aeolis). The essential idea is to combine the formal models of the source language and a minimal IFC language, using the embedding technique described in Matthews and Findler '07. Our contributions are as follows: \begin{itemize} \item We define a minimal IFC language which describes the essence of coarse-grained information flow control. \item We describe how to combine this IFC language with an existing source language in the style of Matthews-Findler, with a twist: there are arbitrarily many copies of the source language, which do not share execution contexts. Mediating between these languages requires serialization of some sort (we make this notion precise in our paper), making our combined system a distributed one. \item We carry out this methodology on four existing languages, and show its adequacy with respect to systems that were specialized for these languages. In particular, our definitions are general enough to enable relaxed isolation when the source language is able to give stronger static guarantees. \item We show how to extend the Matthews-Findler method from just languages that are not simply expressions evaluating to values, but may be collections of threads executing nondeterministically. \end{itemize} The organization of the paper is as follows: first, we describe how to add information flow control to JavaScript, showing the general outline of our procedure. Next, we describe the procedure in generality. Finally, we apply the procedure to a number of systems. } \section{Retrofitting Languages with IFC} Before moving on to the formal treatment of our system, we give a brief primer of information flow control and describe some example programs in our system, emphasizing the parallel between their implementation in a multi-task setting, and the traditional, ``monolithic'' programming language feature they simulate. Information flow control systems operate by associating data with \emph{labels}, and specifying whether or not data tagged with one label \ensuremath{l_{1}} can flow to another label \ensuremath{l_{2}} (written as \ensuremath{l_{1}\flows{}l_{2}}). These labels encode the desired security policy (for example, confidential information should not flow to a public channel), while the work of specifying the semantics of an information flow language involves demonstrating that impermissible flows cannot happen, a property called \emph{non-interference}~. In our coarse-grained floating-label approach, labels are associated with tasks. The task label---we refer to the label of the currently executing task as the \emph{current label}---serves to protect everything in the task's scope; all data in a task shares this common label. As an example, here is a program which spawns a new isolated task, and then sends it a mutable reference: \begin{align*} & \ensuremath{\mathbf{let}\;\Varid{i}\mathrel{=}\tar{_{\textrm{TI}}\lfloor}\mathbf{sandbox}\;(\mathbf{blockingRecv}\ \Varid{x},\anonymous \;\mathbf{in}\;\ifc{^{\textrm{IT}}\lceil}\mathbin{!}\tar{_{\textrm{TI}}\lfloor}\Varid{x}\tar{\rfloor}\ifc{\rceil})\tar{\rfloor}}\\ & \ensuremath{\mathbf{in}\;\tar{_{\textrm{TI}}\lfloor}\mathbf{send}\;\ifc{^{\textrm{IT}}\lceil}\Varid{i}\ifc{\rceil}\;l\;\ifc{^{\textrm{IT}}\lceil}\mathbf{ref}\;\mathbf{true}\ifc{\rceil}\tar{\rfloor}} \end{align*} For now, ignore the tags \ensuremath{\tar{_{\textrm{TI}}\lfloor}\cdot \tar{\rfloor}} and \ensuremath{\ifc{^{\textrm{IT}}\lceil}\cdot \ifc{\rceil}}: roughly, this code creates a new \ensuremath{\mathbf{sandbox}}ed task with identifier $i$ which waits (\textbf{blockingRecv}, binding $x$ with the received message) for a message, and then \textbf{send}s the task a mutable reference \ensuremath{(\mathbf{ref}\;\mathbf{true})} which it labels $l$. If this operation actually shared the mutable cell between the two tasks, it could be used to violate information flow control if the tasks had differing labels. At this point, the designer of an IFC system might add label checks to mutable references, to check the labels of the reader and writer. While this solves the leak, for languages like JavaScript, where references are prevalently used, this also dooms the performance of the system. Our design principles suggest a different resolution: when these constructs are treated as isolated tasks, each of which have their own heaps, it is obviously the case that there is no sharing; in fact, the sandboxed task receives a dangling pointer. Even if there is only one heap, if we enforce that references not be shared, the two systems are morally equivalent. (We elaborate on this formally in Section~.) Finally, this semantics strongly suggests that one should restrict the types of data which may be passed between tasks (for example, in JavaScript, one might only allow JSON objects to be passed between tasks, rather than general object structures). Existing language-based, coarse-grained IFC systems~ allow a sub-computation to temporarily raise the floating-label; after the sub-computation is done, the floating-label is restored to its original label. When this occurs, the enforcement mechanism must ensure that information does not leak to the (less confidential) program continuation. The presence of exceptions adds yet more intricacies. For instance, exceptions should not automatically propagate from a sub-computation directly into the program continuation, and, if such exceptions are allowed to be inspected, the floating-label at the point of the exception-raise must be tracked alongside the exception value~. In contrast, our system provides the same flexibility and guarantees with no extra checks: tasks are used to execute sub-computations, but the mere definition of isolated tasks guarantees that (a) tasks only transfer data to the program continuation by using inter-task communication means, and (b) exceptions do cross tasks boundaries automatically. \subsection{Preliminaries} Our goal now is to describe how to take a \textbf{{\color{red} target language}} with a formal operational semantics and combine it with an \textit{{\color{blue} information flow control language}}. For example, taking ECMAScript as the target language and combining it with our IFC language should produce the formal semantics for the core part of COWL~. In this presentation, we use a simple, untyped lambda calculus with mutable references and fixpoint in place of ECMAScript to demonstrate some the key properties of the system (and, because the embedding does not care about the target language features); we discuss the proper embedding in more detail in Section~. \vspace{1pt} \noindent \textit{Notation} We have typeset nonterminals of the target language using \textbf{{\color{red} bold font}} while the nonterminals of the IFC language have been typeset with \textit{{\color{blue} italic font}}. Readers are encouraged to view a color copy of this paper, where target language nonterminals are colored \textbf{{\color{red} red}} and IFC language nonterminals are colored \textit{{\color{blue} blue}}. \subsection{Target Language: Mini-ES} In Fig.~, we give a simple, untyped lambda calculus with mutable references and fixpoint, prepared for combination with an information flow control language. The presentation is mostly standard, and utilizes Felleisen-Hieb reduction semantics~ to define the operational semantics of the system. One peculiarity is that our language defines an evaluation context \ensuremath{\tar{E}}, but, the evaluation rules have been expressed in terms of a different evaluation context \ensuremath{\mathcal{E}_{\tar{\Sigma}}}; Here, we follow the approach of Matthews and Findler~ in order to simplify combining semantics of multiple languages. To derive the usual operational semantics for this language, the evaluation context merely needs to be defined as \ensuremath{\mathcal{E}_{\tar{\Sigma}}\left[\tar{e}\right]\triangleq{}\tar{\Sigma},\tar{E}\left[\tar{e}\right]}. However, when we combine this language with an IFC language, we reinterpret the meaning of this evaluation context. In general, we require that a target language be expressed in terms of some global machine state \ensuremath{\tar{\Sigma}}, some evaluation context \ensuremath{\tar{E}}, some expressions \ensuremath{\tar{e}}, some set of values \ensuremath{\tar{v}} and a \emph{deterministic} reduction relation on full configurations $\ensuremath{\tar{\Sigma}} \times \ensuremath{\tar{E}} \times \ensuremath{\tar{e}}$. \subsection{IFC Language} As mentioned previously, most modern, dynamic information flow control languages encode policy by associating a label with data. Our embedding is agnostic to the choice of labeling scheme; we only require the labels to form a lattice~ with the partial order $\sqsubseteq$, join \ensuremath{\lub}, and meet \ensuremath{\glb}. In this paper, we simply represent labels with the metavariable $l$, but do not discuss them in more detail. To enforce labels, the IFC monitor inspects the current label before performing a read or a write to decide whether the operation is permitted. A task can only write to entities that are at least as sensitive. Similarly, it can only read from entities that are less sensitive. However, as in other floating-label systems, this current label can be raised to allow the task to read from more sensitive entities at the cost of giving up the ability to write to others. In Fig.~, we give the syntax and \emph{single-task} evaluation rules for a minimal information flow control language. Ordinarily, information flow control languages are defined by directly stating a base language plus information flow control operators. In contrast, our language is purposely minimal: it does not have sequencing operations, control flow, or other constructs. However, it contains support for the following core information flow control features: \begin{itemize} \item First-class labels, with label values $l$ as well as operations for computing on labels (\ensuremath{\flows{}}, \ensuremath{\lub} and \ensuremath{\glb}). \item Operations for inspecting (\textbf{getLabel}) and modifying (\textbf{setLabel}) the current label of the task (a task can only increase its label). \item Operations for non-blocking inter-task communication (\textbf{send} and \textbf{recv}), which interact with the global store of per-task message queues \ensuremath{\ifc{\Sigma}}. \item A sandboxing operation used to spawn new isolated tasks. In concurrent settings \ensuremath{\mathbf{sandbox}} corresponds to a fork-like primitive, whereas in a sequential setting, it more closely resembles computations which might temporarely raise the current floating-label~. \end{itemize} These operations are all defined with respect to an evaluation context \ensuremath{\mathcal{E}_{\ifc{\Sigma}}^{\ifc{\Varid{i}},\ifc{l}}} that represents the context of the current task. The evaluation context has three important pieces of state: the global message queues \ensuremath{\ifc{\Sigma}}, the current label \ensuremath{\ifc{l}} and the task ID \ensuremath{\ifc{\Varid{i}}}. We note that first-class labels, tasks (albeit named differently), and operations for inspecting the current label are essentially universal to all floating-label systems. However, our choice of communication primitives is motivated by those present in browsers, namely \texttt{postMessage}~. Of course, other choices, such as blocking communication or labeled channels, are possible. These asynchronous communication primitives are worth further discussion. When a task is sending a message using \ensuremath{\mathbf{send}}, it also labels that message with a label \ensuremath{\ifc{l}'} (which must be at or above the task's current label \ensuremath{\ifc{l}}). Messages can only be received by a task if its current label is at least as high as the label of the message. Specifically, receiving a message using $\ensuremath{\mathbf{recv}\ \ifc{\Varid{x}}_{1},\ifc{\Varid{x}}_{2}\;\mathbf{in}\;\ifc{\Varid{e}}_{1}\ \mathbf{else}\ \ifc{\Varid{e}}_{2}}$ binds the message and the sender's task identifier to local variables \ensuremath{\ifc{\Varid{x}}_{1}} and \ensuremath{\ifc{\Varid{x}}_{2}}, respectively, and then executes \ensuremath{\ifc{\Varid{e}}_{1}}. Otherwise, if there are no messages, that task continues its execution with \ensuremath{\ifc{\Varid{e}}_{2}}. We denote the filtering of the message queue by \ensuremath{\ifc{\Theta}\preceq\ifc{l}}, which is defined as follows. If \ensuremath{\ifc{\Theta}} is the empty list \ensuremath{\mathbf{nil}}, the function is simply the identity function, i.e., \ensuremath{\mathbf{nil}\preceq\ifc{l}\mathrel{=}\mathbf{nil}}, and otherwise: \[ \ensuremath{((\ifc{l}',\ifc{\Varid{i}},\ifc{\Varid{e}}),\ifc{\Theta})\preceq\ifc{l}} = \left\{ \begin{array}{l l} \ensuremath{(\ifc{l}',\ifc{\Varid{i}},\ifc{\Varid{e}}),(\ifc{\Theta}\preceq\ifc{l})} & \quad \text{if \ensuremath{\ifc{l}'\;\flows{}\;\ifc{l}}}\\ \ensuremath{\ifc{\Theta}\preceq\ifc{l}} & \quad \text{otherwise} \end{array} \right. \] This ensures that tasks cannot receive messages that are more sensitive than their current label would allow. \subsection{The Embedding} Fig.~ provides all of the rules responsible for actually carrying out the embedding of the IFC language within the target language. The most important feature of this embedding is that every task maintains its own copy of the target language global state and evaluation context, thus enforcing isolation between various tasks. In more detail: \begin{itemize} \item We extend the values, expressions and evaluation contexts of both languages to allow for terms in one language to be embedded in the other, as in~. In the target language, an IFC expression appears as \ensuremath{\tar{_{\textrm{TI}}\lfloor}\ifc{\Varid{e}}\tar{\rfloor}} (``\Red{T}arget-outside, IFC-inside''); in the IFC language, a target language expression appears as \ensuremath{\ifc{^{\textrm{IT}}\lceil}\tar{e}\ifc{\rceil}} ( ``{\color{blue}{I}}FC-outside, target-inside''). \item We reinterpret \ensuremath{\mathcal{E}} to be evaluation contexts on task lists, providing definitions for \ensuremath{\mathcal{E}_{\tar{\Sigma}}} and \ensuremath{\mathcal{E}_{\ifc{\Sigma}}^{\ifc{\Varid{i}},\ifc{l}}}. These rules only operate on the first task in the task list, which by convention is the only task executing. \item We reinterpret \ensuremath{\rightarrow}, an operation on a single task, in terms of \ensuremath{\hookrightarrow}, operation on task lists. The correspondence is simple: a task executes a step and then is rescheduled in the task list according to schedule policy \ensuremath{\alpha}. Fig.~ defines two concrete schedulers. \item Finally, we define some rules for scheduling, handling sandboxing tasks (which interact with the state of the target language), and intermediating between the borders of the two languages. \end{itemize} The \textsc{I-sandbox} rule is used to create a new isolated task that executes separately from the existing tasks (and can be communicated with via \ensuremath{\mathbf{send}} and \ensuremath{\mathbf{recv}}). When the new task is created, there is the question of what the target language state of the new task should be. Our rule is stated generically in terms of a function \ensuremath{\kappa}. Conservatively, \ensuremath{\kappa} may be simply thought of as the identity function, in which case the semantics of \ensuremath{\mathbf{sandbox}} are such that the state of the target language is \emph{cloned} when sandboxing occurs. However, this is not necessary: it is also valid for \ensuremath{\kappa} to remove entries from the state. In Section~, we give a more detailed discussion of the implications of the choice of \ensuremath{\kappa}, but all our security claims will hold regardless of the choice of \ensuremath{\kappa}. The rule \textsc{I-noStep} says something about configurations for which it is not possible to take a transition. The notation $\ensuremath{\ifc{c}}\not\overset{\alpha}{\hookrightarrow}$ in the premise is meant to be understood as follows: If the configuration \ensuremath{\ifc{c}} cannot take a step by any rule other than \textsc{I-noStep}, then \textsc{I-noStep} applies and the stuck task gets removed. Rules \textsc{I-done} and \textsc{I-noStep} define the behavior of the system when the current thread has reduced to a value, or gotten stuck, respectively. While these definitions simply rely on the underlying scheduling policy \ensuremath{\alpha} to modify the task list, as we describe in Sections~ and~, these rules (notably, \textsc{I-noStep}) are crucial to proving our security guarantees. For instance, it is unsafe for the whole system to get stuck if a particular task gets stuck, since a sensitive thread may then leverage this to leak information through the termination channel. Instead, as our example round-robin (\ensuremath{\textsc{RR}}) scheduler shows, such tasks should simply be removed from the task list. Many language runtime or Operating System schedulers implement such schedulers. Moreover, techniques such as instruction-based scheduling~ can be further applied close the gap between specified semantics and implementation. As in~, rules \textsc{T-border} and \textsc{I-border} define the syntactic boundaries between the IFC and target languages. Intuitively, the boundaries respectively correspond to an upcall into and downcall from the IFC runtime. As an example, taking \ensuremath{\Red{\lambda_{\text{ES}}}} as the target language, we can now define a blocking receive (inefficiently) in terms of the asynchronous \ensuremath{\mathbf{recv}} as series of cross-language calls: \begin{hscode}\SaveRestoreHook \column{B}{@{}>{\hspre}l<{\hspost}@{}} \column{E}{@{}>{\hspre}l<{\hspost}@{}} \>[B]{}\mathbf{blockingRecv}\ \ifc{\Varid{x}}_{1},\ifc{\Varid{x}}_{2}\;\mathbf{in}\;\ifc{\Varid{e}}\triangleq{}\ifc{^{\textrm{IT}}\lceil}\mathbf{fix}\;(\lambda \Varid{k}.\tar{_{\textrm{TI}}\lfloor}\mathbf{recv}\ \ifc{\Varid{x}}_{1},\ifc{\Varid{x}}_{2}\;\mathbf{in}\;\ifc{\Varid{e}}\ \mathbf{else}\ \ifc{^{\textrm{IT}}\lceil}\Varid{k}\ifc{\rceil}\tar{\rfloor})\ifc{\rceil}{}\<[E] \ColumnHook \end{hscode}\resethooks For any target language \ensuremath{\Red{\lambda}} and scheduling policy \ensuremath{\alpha}, this embedding defines an IFC language, which we will refer to as \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda})}. \section{Security Guarantees} We are interested in proving non-interference about many programming languages. This requires an appropriate definition of this notion that is language agnostic, so in this section, we present a few general definitions for what an information flow control language is and what non-interference properties it may have. In particular, we show that \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda})}, with an appropriate scheduler \ensuremath{\alpha}, satisfies non-interference~, without making any reference to properties of \ensuremath{\Red{\lambda}}. We state the appropriate theorems here, and provide the formal proofs in \appref{sec:app:proof}. \cut{ The precise scheduling policy dictates what guarantee we can achieve for programs with diverging tasks. For a sequential scheduler \ensuremath{\textsc{Seq}}, we will only be able to show \emph{termination-insensitive non-interference}, where a program may diverge based on secret data; this allows attackers to observe secrets by observing the termination of tasks. For the concurrent round-robin schedule \ensuremath{\textsc{RR}}, we can show a stronger result known as \emph{termination-sensitive non-interference}, where termination attacks cannot leak information. } \subsection{Erasure Function} When defining the security guarantees of an information flow control, we must characterize what the \emph{secret inputs} of a program are. Like other work~, we specify and prove non-interference using \emph{term erasure}. Intuitively, term erasure allows us to show that an attacker does not learn any sensitive information from a program if the program behaves identically (from the attackers point of view) to a program with all sensitive data ``erased''. To interpret a language under information flow control, we define a function \ensuremath{\varepsilon_{l}} that performs erasures by mapping configurations to erased configurations, usually by rewriting (parts of) configurations that are more sensitive than \ensuremath{l} to a new syntactic construct \ensuremath{\bullet}. We define an information flow control language as follows: \begin{definition}[Information flow control language] An information flow control language \ensuremath{\textbf{L}} is a tuple \ensuremath{(\Delta,\hookrightarrow,\varepsilon_{l})}, where $\ensuremath{\Delta}$ is the type of machine configurations (members of which are usually denoted by the metavariable \ensuremath{c}), \ensuremath{\hookrightarrow} is a reduction relation between machine configurations and \ensuremath{\varepsilon_{l}\mathbin{:}\Delta\rightarrow\varepsilon(\Delta)} is an erasure function parametrized on labels from machine configurations to \emph{erased} machine configurations \ensuremath{\varepsilon(\Delta)}. Sometimes, we use \ensuremath{V} to refer to set of terminal configurations in \ensuremath{\Delta}, i.e., configurations where no further transitions are possible. \end{definition} Our language \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda}\;\!\!)} fulfills this definition as \ensuremath{(\ifc{\Delta},\overset{\alpha}{\hookrightarrow},\varepsilon_{\ifc{l}})}, where $\ensuremath{\ifc{\Delta}} = \ensuremath{\ifc{\Sigma}} \times \operatorname{List}(\ensuremath{\ifc{\Varid{t}}})$. The set of terminal conditions $\ensuremath{\ifc{V}}$ is $\ensuremath{\ifc{\Sigma}} \times \ensuremath{\ifc{\Varid{t}}}_V$, where $\ensuremath{\ifc{\Varid{t}}}_V \subset \ensuremath{\ifc{\Varid{t}}}$ is the type for tasks whose expressions have been reduced to values.\footnote{ Here, we abuse notation by describing types for configuration parts using the same metavariables as the ``instance'' of the type, e.g., \ensuremath{\ifc{\Varid{t}}} for the type of task. } The erased configuration \ensuremath{\varepsilon(\ifc{\Delta})} extends \ensuremath{\ifc{\Delta}} with configurations containing \ensuremath{\bullet}, and Fig.~ gives the precise definition for our erasure function \ensuremath{\varepsilon_{\ifc{l}}}. Essentially, a task and its corresponding message queue is completely erased from the task list if its label does not flow to the attacker observation level \ensuremath{\ifc{l}}. Otherwise, we apply the erasure function homomorphically and remove any messages from the task's message queue that are more sensitive than \ensuremath{\ifc{l}}. The definition of an erasure function is quite important: it captures the attacker model, stating what can and cannot be observed by the attacker. In our case, we assume that the attacker cannot observe sensitive tasks or messages, or even the number of such entities. While such assumptions are standard~\cite{Castellani:Boudol:ICALP01, stefan:addressing-covert}, our definitions allow for stronger attackers that may be able to inspect resource usage.\footnote{ We believe that we can extend \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda}\;\!\!)} to such models using the resource limits techniques of~. We leave this extension to future work. } \subsection{Non-Interference} Given an information flow control language, we can now define non-interference. Intuitively, we want to make statements about the attacker's observational power at some security level \ensuremath{l}. This is done by defining an equivalence relation called \ensuremath{l}-equivalence on configurations: an attacker should not be able to distinguish two configurations that are \ensuremath{l}-equivalent. Since our erasure function captures what an attacker can or cannot observe, we simply define this equivalence as the syntactic-equivalence of erased configurations~. \begin{definition}[\ensuremath{l}-equivalence] In a language \ensuremath{(\Delta,\hookrightarrow,\varepsilon_{l})}, two machine configurations \ensuremath{c,c'\in\Delta} are considered $l$-equivalent, written as \ensuremath{c\approx_lc'}, if \ensuremath{\varepsilon_{l}(c)\mathrel{=}\varepsilon_{l}(c')}. \end{definition} We can now state that a language satisfies non-interference if an attacker at level \ensuremath{l} cannot distinguish the runs of any two \ensuremath{l}-equivalent configurations. This particular property is called termination sensitive non-interference (TSNI). Besides the obvious requirement to not leak secret information to public channels, this definition also requires the termination of public tasks to be independent of secret tasks. Formally, we define TSNI as follows: \begin{definition}[Termination Sensitive Non-Interference (TSNI)] A language \ensuremath{(\Delta,\hookrightarrow,\varepsilon_{l})} satisfies termination sensitive non-interference if for any label \ensuremath{l}, and configurations $\ensuremath{c_{1},c_{1}',c_{2}}\in\ensuremath{\Delta}$, if \begin{equation} \ensuremath{c_{1}} \approx_{\ensuremath{l}} \ensuremath{c_{2}} \qquad \text{and} \qquad \ensuremath{c_{1}} \ensuremath{\hookrightarrow}^* \ensuremath{c_{1}'} \end{equation} then there exists a configuration $\ensuremath{c_{2}'}\in\ensuremath{\Delta}$ such that \begin{equation} \ensuremath{c_{1}'} \approx_{\ensuremath{l}} \ensuremath{c_{2}'} \qquad \text{and} \qquad \ensuremath{c_{2}} \ensuremath{\hookrightarrow}^* \ensuremath{c_{2}'} \ \text{.} \end{equation} \end{definition} In other words, if we take two \ensuremath{l}-equivalent configurations, then for every intermediate step taken by the first configuration, there is a corresponding number of steps that the second configuration can take to result in a configuration that is \ensuremath{l}-equivalent to the first resultant configuration. By symmetry, this applies to all intermediate steps from the second configuration as well. \ifextended We remark that this notion of non-interfernce is similar to \emph{progress sensitive non-interference (PSNI)}, which accounts for leakage via progress (or termination) channels, as used for static systems~. \fi Our language satisfies TSNI \ifextended (and thus PSNI) \fi under the round-robin scheduler \ensuremath{\textsc{RR}} of Fig.~. \begin{theorem}[Concurrent IFC language is TSNI] For any target language \ensuremath{\Red{\lambda}}, \ensuremath{L_\text{IFC}(\textsc{RR},\Red{\lambda})} satisfies TSNI. \end{theorem} In general, however, non-interference will not hold for an arbitrary scheduler \ensuremath{\alpha}. For example, \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda})} with a scheduler that inspects a sensitive task's current state when deciding which task to schedule next will in general break non-interference~. However, even non-adversarial schedulers are not always safe. Consider, for example, the sequential scheduling policy \ensuremath{\textsc{Seq}} given in Fig.~. It is easy to show that \ensuremath{L_\text{IFC}(\textsc{Seq},\Red{\lambda})} does not satisfy TSNI: consider a target language similar to \ensuremath{\Red{\lambda_{\text{ES}}}} with an additional expression terminal \ensuremath{\tar{\Uparrow}} that denotes a divergent computation, i.e., \ensuremath{\tar{\Uparrow}} always reduces to \ensuremath{\tar{\Uparrow}} and a simple label lattice \ensuremath{\{\mskip1.5mu \mathsf{pub},\mathsf{sec}\mskip1.5mu\}} such that \ensuremath{\mathsf{pub}\flows{}\mathsf{sec}}, but \ensuremath{\mathsf{sec}\not\flows{}\mathsf{pub}}. Consider the following two configurations in this language: \begin{hscode}\SaveRestoreHook \column{B}{@{}>{\hspre}l<{\hspost}@{}} \column{71}{@{}>{\hspre}l<{\hspost}@{}} \column{E}{@{}>{\hspre}l<{\hspost}@{}} \>[B]{}\ifc{c}_{1}\mathrel{=}\ifc{\Sigma};\langle \tar{\Sigma}_{1}, \ifc{^{\textrm{IT}}\lceil}\;\mathbf{if}\;\mathbf{false}\;\mathbf{then}\;\tar{\Uparrow}\;\mathbf{else}\;\mathbf{true}\ifc{\rceil}\rangle^{\mathrm{1}}_{\mathsf{sec}},{}\<[71] \>[71]{}\langle \tar{\Sigma}_{2}, \ifc{\Varid{e}}\rangle^{\mathrm{2}}_{\mathsf{pub}}{}\<[E] \\ \>[B]{}\ifc{c}_{2}\mathrel{=}\ifc{\Sigma};\langle \tar{\Sigma}_{1}, \ifc{^{\textrm{IT}}\lceil}\;\mathbf{if}\;\mathbf{true}\;\mathbf{then}\;\tar{\Uparrow}\;\mathbf{else}\;\mathbf{true}\ifc{\rceil}\rangle^{\mathrm{1}}_{\mathsf{sec}},{}\<[71] \>[71]{}\langle \tar{\Sigma}_{2}, \ifc{\Varid{e}}\rangle^{\mathrm{2}}_{\mathsf{pub}}{}\<[E] \ColumnHook \end{hscode}\resethooks These two configurations are \ensuremath{\mathsf{pub}}-equivalent, but \ensuremath{\ifc{c}_{1}} will reduce (in two steps) to \ensuremath{\ifc{c}_{1}'\mathrel{=}\ifc{\Sigma};\langle \tar{\Sigma}_{1}, \ifc{^{\textrm{IT}}\lceil}\mathbf{true}\ifc{\rceil}\rangle^{\mathrm{2}}_{\mathsf{pub}}}, whereas \ensuremath{\ifc{c}_{2}} will not make any progress. Suppose that \ensuremath{\ifc{\Varid{e}}} is a computation that writes to a \ensuremath{\mathsf{pub}} channel,\footnote{ Though we do not model labeled channels, extending the calculus with such a feature is straightforward, see Section~.} then the \ensuremath{\mathsf{sec}} task's decision to diverge or not is directly leaked to a public entity. To accommodate for sequential languages, or cases where a weaker guarantee is sufficient, we consider an alternative non-interference property called termination insensitive non-interference (TINI). This property can also be upheld by sequential languages at the cost of leaking through (non)-termination~. \begin{definition}[Termination insensitive non-interference (TINI)] A language \ensuremath{(\Delta,V,\hookrightarrow,\varepsilon_{l})} is termination insensitive non-interfering if for any label \ensuremath{l}, and configurations $\ensuremath{c_{1},c_{2}}\in\ensuremath{\Delta}$ and $\ensuremath{c_{1}',c_{2}'}\in\ensuremath{V}$, it holds that \[ (\ensuremath{c_{1}} \approx_{\ensuremath{l}} \ensuremath{c_{2}} \;\;\land\;\; \ensuremath{c_{1}} \ensuremath{\hookrightarrow}^* \ensuremath{c_{1}'} \;\;\land\;\; \ensuremath{c_{2}} \ensuremath{\hookrightarrow}^* \ensuremath{c_{2}'}) \implies \ensuremath{c_{1}'} \approx_{\ensuremath{l}} \ensuremath{c_{2}'} \] \end{definition} TINI states that if we take two \ensuremath{l}-equivalent configurations, and both configurations reduce to final configurations (i.e., configurations for which there are no possible further transitions), then the end configurations are also \ensuremath{l}-equivalent. We highlight that this statement is much weaker than TSNI: it only states that terminating programs do not leak sensitive data, but makes no statement about non-terminating programs. As shown by compilers~, interpreters~, and libraries~, TINI is useful for sequential settings. In our case, we show that our IFC language with the sequential scheduling policy \ensuremath{\textsc{Seq}} satisfies TINI. \begin{theorem}[Sequential IFC language is TINI] For any target language \ensuremath{\Red{\lambda}}, \ensuremath{L_\text{IFC}(\textsc{Seq},\Red{\lambda})} satisfies TINI. \end{theorem} \section{Isomorphisms and Restrictions} \newcommand{\con}[1]{\ensuremath{{\color{red} #1}}} \newcommand{\abs}[1]{\ensuremath{{\color{blue} #1}}} The operational semantics we have defined in the previous section satisfy non-interference by design. We achieve this general statement that works for a large class of languages by having different tasks executing completely isolated from each other, such that every task has its own state. In some cases, this strong separation is desirable, or even necessary. Languages like C provide direct access to memory locations without mechanisms in the language to achieve a separation of the heap. On the other hand, for other languages, this strong isolation of tasks can be undesirable, e.g., for performance reasons. For instance, for the language \ensuremath{\Red{\lambda_{\text{ES}}}}, our presentation so far requires a separate heap per task, which is not very practical. Instead, we would like to more tightly couple the integration of the target and IFC languages by reusing existing infrastructure. In the running example, a concrete implementation might use a single global heap. More precisely, instead of using a configuration of the form $\ensuremath{\ifc{\Sigma};\langle \tar{\Sigma}_{1}, \ifc{\Varid{e}}_{1}\rangle^{\ifc{\Varid{i}}_{1}}_{\ifc{l}_{1}},\langle \tar{\Sigma}_{2}, \ifc{\Varid{e}}_{2}\rangle^{\ifc{\Varid{i}}_{2}}_{\ifc{l}_{2}}\;\ldots}$ we would like a single global heap as in $\ensuremath{\ifc{\Sigma};\tar{\Sigma};\langle \ifc{\Varid{e}}_{1}\rangle^{\ifc{\Varid{i}}_{1}}_{\ifc{l}_{1}},\langle \ifc{\Varid{e}}_{2}\rangle^{\ifc{\Varid{i}}_{2}}_{\ifc{l}_{2}},\ldots}$ If the operational rules are adapted na\"ively to this new setting, then non-interference can be violated: as we mentioned earlier, shared mutable cells could be used to leak sensitive information. What we would like is a way of characterizing safe modifications to the semantics which preserve non-interference. The intention of our single heap implementation is to permit efficient execution while \emph{conceptually maintaining isolation between tasks} (by not allowing sharing of references between them). This intuition of having a different (potentially more efficient) concrete semantics that behaves like the abstract semantics can be formalized by the following definition: \begin{definition}[Isomorphism of information flow control languages] A language \ensuremath{(\Delta,\hookrightarrow,\varepsilon_{l})} is \textit{isomorphic} to a language \ensuremath{(\Delta',\hookrightarrow',\varepsilon'_{l})} if there exist total functions \ensuremath{\Varid{f}\mathbin{:}\Delta\rightarrow\Delta'} and \ensuremath{\Varid{f}^{-1}\mathbin{:}\Delta'\rightarrow\Delta} such that \ensuremath{\Varid{f}\circ\Varid{f}^{-1}\mathrel{=}id_{\Delta}} and \ensuremath{\Varid{f}^{-1}\circ\Varid{f}\mathrel{=}id_{\Delta'}}. Furthermore, \ensuremath{\Varid{f}} and \ensuremath{\Varid{f}^{-1}} are functorial (e.g., if $x'\ R'\ y'$ then $f(x')\ R\ f(y')$) over both $l$-equivalences and \ensuremath{\hookrightarrow}. If we weaken this restriction such that \ensuremath{\Varid{f}^{-1}} does not have to be functorial over \ensuremath{\hookrightarrow}, we call the language \ensuremath{(\Delta,\hookrightarrow,\varepsilon_{l})} \textit{weakly isomorphic} to \ensuremath{(\Delta',\hookrightarrow',\varepsilon'_{l})}. \end{definition} Providing an isomorphism between the two languages allows us to preserve (termination sensitive or insensitive) non-interference as the following two theorems state. \begin{theorem}[Isomorphism preserves TSNI] If $L$ is isomorphic to $L'$ and $L'$ satisfies TSNI, then $L$ satisfies TSNI. \end{theorem} \begin{proof} Shown by transporting configurations and reduction derivations from $L$ to $L'$, applying TSNI, and then transporting the resulting configuration, $l$-equivalence and multi-step derivation back. \qed \end{proof} Only weak isomorphism is necessary for TINI. Intuitively, this is because it is not necessary to back-translate reduction sequences in $L'$ to $L$; by the definition of TINI, we have both reduction sequences in $L$ by assumption. \begin{theorem}[Weak isomorphism preserves TINI] If a language $L$ is weakly isomorphic to a language $L'$, and $L'$ satisfies TINI, then $L$ satisfies TINI. \end{theorem} \begin{proof} Shown by transporting configurations and reduction derivations from $L$ to $L'$, applying TINI and transporting the resulting equivalence back using functoriality of \ensuremath{\Varid{f}^{-1}} over $l$-equivalences. \qed \end{proof} Unfortunately, an isomorphism is often too strong of a requirement. To obtain an isomorphism with our single heap semantics, we need to mimic the behavior of several heaps with a single actual heap. The interesting cases are when we sandbox an expression and when messages are sent and received. The rule for sandboxing is parametrized by the strategy \ensuremath{\kappa} (see Section~), which defines what heap the new task should execute with. We have considered two choices: \begin{itemize} \item When we sandbox into an empty heap, existing addresses in the sandboxed expression are no longer valid and the task will get stuck (and then removed by \textsc{I-noStep}). Thus, we must rewrite the sandboxed expression so that all addresses point to fresh addresses guaranteed to not occur in the heap. Similarly, sending a memory address should be rewritten. \item When we clone the heap, we have to copy everything reachable from the sandboxed expression and replace all addresses correspondingly. Even worse, the behavior of sending a memory address now depends on whether that address existed at the time the receiving task was sandboxed; if it did, then the address should be rewritten to the existing one. \end{itemize} Isomorphism demands we implement this convoluted behavior, despite our initial motivation of a more efficient implementation. \subsection{Restricting the IFC Language} A better solution is to forbid sandboxed expressions as well as messages sent to other tasks to contain memory addresses in the first place. In a statically typed language, the type system could prevent this from happening. In dynamically typed languages such as \ensuremath{\Red{\lambda_{\text{ES}}}}, we might restrict the transition for \ensuremath{\mathbf{sandbox}} and \ensuremath{\mathbf{send}} to only allow expressions without memory addresses. While this sounds plausible, it is worth noting that we are modifying the IFC language semantics, which raises the question of whether non-interference is preserved. This question can be subtle: it is easy to remove a transition from a language and invalidate TSNI. Intuitively if the restriction depends on secret data, then a public thread can observe if some other task terminates or not, and from that obtain information about the secret data that was used to restrict the transition. With this in mind, we require semantic rules to get restricted only based on information observable by the task triggering them. This ensures that non-interference is preserved, as the restriction does not depend on confidential information. Below, we give the formal definition of this condition for the abstract IFC language \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda})}. \begin{definition}[Restricted IFC language] For a family of predicates $\mathcal P$ (one for every reduction rule), we call \ensuremath{L_\text{IFC}^{\mathcal{P}}(\alpha,\Red{\lambda})} a restricted IFC language if its definition is equivalent to the abstract language \ensuremath{L_\text{IFC}(\alpha,\Red{\lambda})}, with the following exception: the reduction rules are restricted by adding a predicate $P \in \mathcal P$ to the premise of all rules other than \textsc{I-noStep}. Furthermore, the predicate $P$ can depend only on the \textit{erased} configuration \ensuremath{\varepsilon_{\ifc{l}}(\ifc{c})}, where \ensuremath{\ifc{l}} is the label of the first task in the task list and \ensuremath{\ifc{c}} is the full configuration. \end{definition} By the following theorem, the restricted IFC language with an appropriate scheduling policy is non-interfering. \begin{theorem} For any target language \ensuremath{\Red{\lambda}} and family of predicates $\mathcal{P}$, the restricted IFC language \ensuremath{L_\text{IFC}^{\mathcal{P}}(\textsc{RR},\Red{\lambda})} is TSNI. Furthermore, the IFC language \ensuremath{L_\text{IFC}^{\mathcal{P}}(\textsc{Seq},\Red{\lambda})} is TINI. \end{theorem} In~\appref{sec:single-heap} we give an example how this formalism can be used to show non-intereference of an implementation of IFC with a single heap. \section{Real World Languages} Our approach can be used to retrofit any language for which we can achieve isolation with information flow control. Unfortunately, controlling the external effects of a real-world language, as to achieve isolation, is language-specific and varies from one language to another.\footnote{ Though we apply our framework to several real-world languages, it is conceivable that there are languages for which isolation cannot be easily achieved. } Indeed, even for a single language (e.g., JavaScript), how one achieves isolation may vary according to the language runtime or embedding (e.g., server and browser). In this section, we describe several implementations and their approaches to isolation. In particular, we describe two JavaScript IFC implementations building on the theoretical foundations of this work. Then, we consider how our formalism could be applied to the C programming language and connect it to a previous IFC system for Haskell. \subsection{JavaScript} JavaScript, as specified by ECMAScript~, does not have any built-in functionality for I/O. For this language, which we denote by \ensuremath{\Red{\lambda_{\text{JS}}}}, the IFC system \ensuremath{L_\text{IFC}(\textsc{RR},\Red{\lambda_{\text{JS}}})} can be implemented by exposing IFC primitives to JavaScript as part of the runtime, and running multiple instances of the JavaScript virtual machine in separate OS-level threads. Unfortunately, this becomes very costly when a system, such as a server-side web application, relies on many tasks. Luckily, this issue is not unique to our work---browser layout engines also rely on isolating code executing in separate iframes (e.g., according to the same-origin policy). Since creating an OS thread for each iframe is expensive, both the V8 and SpiderMonkey JavaScript engines provide means for running JavaScript code in isolation within a single OS thread, on disjoint sub-heaps. In V8, this unit of isolation is called a \emph{context}; in SpiderMonkey, it is called a \emph{compartment}. (We will use these terms interchangeably.) Each context is associated with a global object, which, by default, implements the JavaScript standard library (e.g., \text{\tt Object}, \text{\tt Array}, etc.). Naturally, we adopt contexts to implement our notion of tasks. When JavaScript is embedded in browser layout engines, or in server-side platforms such as Node.js, additional APIs such as the Document Object Model (DOM) or the file system get exposed as part of the runtime system. These features are exposed by extending the global object, just like the standard library. For this reason, it is easy to modify these systems to forbid external effects when implementing an IFC system, ensuring that important effects can be reintroduced in a safe manner. \paragraph{Server-side IFC for Node.js:} We have implemented \ensuremath{L_\text{IFC}(\textsc{Seq},\Red{\lambda_{\text{JS}}})} for Node.js in the form of a library, without modifying Node.js or the V8 JavaScript engine. Our implementation.} provides a library for creating new tasks, i.e., contexts whose global object only contains the standard JavaScript library and our IFC primitives (e.g., \ensuremath{\mathbf{send}} and \ensuremath{\mathbf{sandbox}}). When mapped to our formal treatment, \ensuremath{\mathbf{sandbox}} is defined with \ensuremath{\kappa(\tar{\Sigma})\mathrel{=}\tar{\Sigma}_{0}}, where \ensuremath{\tar{\Sigma}_{0}} is the global object corresponding to the standard JavaScript library and our IFC primitives. These IFC operations are mediated by the trusted library code (executing as the main Node.js context), which tracks the state (current label, messages, etc.) of each task. An example for \ensuremath{\mathbf{send}}/\ensuremath{\mathbf{recv}} is shown in Fig.~. Our system conservatively restricts the kinds of messages that can be exchanged, via \ensuremath{\mathbf{send}} (and \ensuremath{\mathbf{sandbox}}), to string values. In our formalization, this amounts to restricting the IFC language rule for \ensuremath{\mathbf{send}} in the following way: \newcommand{\str}{"string"} \begin{mathpar} \inferrule[JS-send] { \ensuremath{\ifc{l}\;\flows{}\;\ifc{l}'}\\ \ensuremath{\ifc{\Sigma}\;(\ifc{\Varid{i}}')\mathrel{=}\ifc{\Theta}}\\ \ensuremath{\ifc{\Sigma}'\mathrel{=}\ifc{\Sigma}\;[\mskip1.5mu \ifc{\Varid{i}}'\mapsto{}(\ifc{l}',\ifc{\Varid{i}},\ifc{\Varid{v}}),\ifc{\Theta}\mskip1.5mu]}\\ \ensuremath{\ifc{\Varid{e}}\mathrel{=}\ifc{^{\textrm{IT}}\lceil}\tar{e}\ifc{\rceil}}\\ \ensuremath{\mathcal{E}_{\tar{\Sigma}}\left[\texttt{typeOf}(\tar{e})\texttt{ === \str}\right]\rightarrow\mathcal{E}_{\tar{\Sigma}}\left[\texttt{true}\right]} } {\ensuremath{\ifc{\Sigma};\langle \tar{\Sigma}, \ifc{E}\ifc{[}\mathbf{send}\;\ifc{\Varid{i}}'\;\ifc{l}'\;\ifc{\Varid{v}}\ifc{]}_\ifc{I}\rangle^{\ifc{\Varid{i}}}_{\ifc{l}},\ldots\hookrightarrow\ifc{\Sigma}';\alpha_{{\tiny\mathrm{\text{step}}}}(\langle \tar{\Sigma}, \ifc{E}\ifc{[}\langle\rangle\ifc{]}_\ifc{I}\rangle^{\ifc{\Varid{i}}}_{\ifc{l}},\ldots)}} \end{mathpar} Of course, we provide a convenience library which marshals JSON objects to/from strings. We remark that this is not unlike existing message-passing JavaScript APIs, e.g., \texttt{postMessage}, which impose similar restrictions as to avoid sharing references between concurrent code. While the described system implements \ensuremath{L_\text{IFC}(\textsc{Seq},\Red{\lambda_{\text{JS}}})}, applications typically require access to libraries (e.g., the file system library \textsf{fs}) that have external effects. Exposing the Node.js APIs directly to sandboxed tasks is unsafe. Instead, we implement libraries (like a labeled version of \textsf{fs}) as message exchanges between the sandboxed tasks (e.g., \textsf{task-1} in Fig.~) and the main Node.js task that implements the IFC monitor. While this is safer than simply wrapping unsafe objects, which can potentially be exploited to access objects outside the context (e.g., as seen with \ifextended ADSafe, FBJS, and Caja~), \else ADSafe~), \fi adding features such as the \textsf{fs} requires the code in the main task to ensures that labels are properly propagated and enforced. Unfortunately, while imposing such a proof burden is undesirable, this also has to be expected: different language environments expose different libraries for handling external I/O, and the correct treatment of external effects is application specific. We do not extend our formalism to account for the particular interface to the file system, HTTP client, etc., as this is specific to the Node.js implementation and does not generalize to other systems. \paragraph{Client-side IFC:} This work provides the formal basis for the core part of the COWL client-side JavaScript IFC system~. Like our Node.js implementation, COWL takes a coarse-grained approach to providing IFC for JavaScript programs. However, COWL's IFC monitor is implemented in the browser layout engine instead (though still leaving the JavaScript engine unmodified). Furthermore, COWL repurposes existing contexts (e.g., iframes and pages) as IFC tasks, only imposing additional constraints on how they communicate. As with Node.js, at its core, the global object of a COWL task should only contain the standard JavaScript libraries and \texttt{postMessage}, whose semantics are modeled by our \textsc{JS-send} rule. However, existing contexts have objects such as the DOM, which require COWL to restrict a task's external effects. To this end, COWL mediates any communication (even via the DOM) at the context boundary. Simply disallowing all the external effects is overly-restricting for real-world applications (e.g., pages typically load images, perform network requests, etc.). In this light, COWL allows safe network communication by associating an implicit label with remote hosts (a host's label corresponds to its origin). In turn, when a task performs a request, COWL's IFC monitor ensures that the task label can flow to the remote origin label. While the external effects of COWL can be formally modeled, we do not model them in our formalism, since, like for the Node.js case, they are specific to this system. \subsection{Haskell} Our work borrows ideas from the LIO Haskell coarse-grained IFC system~. LIO relies on Haskell's type system and monadic encoding of effects to achieve isolation and define the IFC sub-language. Specifically, LIO provides the \text{\tt LIO} monad as a way of restricting (almost all) side-effects. In the context of our framework, LIO can be understood as follows: the \emph{pure subset} of Haskell is the target language, while the monadic subset of Haskell, operating in the \text{\tt LIO} monad, is the IFC language. Unlike our proposal, LIO originally associated labels with exceptions, in a similar style to fine-grained systems~. In addition to being overly complex, the interaction of exceptions with clearance (which sets an upper bound on the floating label, see \appref{sec:clearance}) was incorrect: the clearance was restored to the clearance at point of the catch. Furthermore, pure exceptions (e.g., divide by zero) always percolated to trusted code, effectively allowing for denial of service attacks. The insights gained when viewing coarse-grained IFC as presented in this paper led to a much cleaner, simpler treatment of exceptions, which has now been adopted by LIO. \subsection{C} C programs are able to execute arbitrary (machine) code, access arbitrary memory, and perform arbitrary system calls. Thus, the confinement of C programs must be imposed by the underlying OS and hardware. For instance, our notion of isolation can be achieved using Dune's hardware protection mechanisms~, similar to Wedge~\cite{Belay:2012:DSU:2387880.2387913, Bittau:2008:WSA:1387589.1387611}, but using an information flow control policy. Using page tables, a (trusted) IFC runtime could ensure that each task, implemented as a lightweight process, can only access the memory it allocates---tasks do not have access to any shared memory. In addition, ring protection could be used to intercept system calls performed by a task and only permit those corresponding to our IFC language (such as \ensuremath{\mathbf{getLabel}} or \ensuremath{\mathbf{send}}). Dune's hardware protection mechanism would allow us to provide a concrete implementation that is efficient and relatively simple to reason about, but other sandboxing mechanisms could be used in place of Dune. In this setting, the combined language of Section~ can be interpreted in the following way: calling from the target language to the IFC language corresponds to invoking a system call. Creating a new task with the \ensuremath{\mathbf{sandbox}} system call corresponds to \emph{forking} a process. Using page tables, we can ensure that there will be no shared memory (effectively defining \ensuremath{\kappa(\tar{\Sigma})\mathrel{=}\tar{\Sigma}_{0}}, where \ensuremath{\tar{\Sigma}_{0}} is the set of pages necessary to bootstrap a lightweight process). Similarly, control over page tables and protection bits allows us to define a \ensuremath{\mathbf{send}} system call that copies pages to our (trusted) runtime queue; and, correspondingly, a \ensuremath{\mathbf{recv}} that copies the pages from the runtime queue to the (untrusted) receiver. Since C is not memory safe, conditions on these system calls are meaningless. We leave the implementation of this IFC system for C as future work. \section{Extensions and Limitations} While the IFC language presented thus far provides the basic information flow primitives, actual IFC implementations may wish to extend the minimal system with more specialized constructs. For example, COWL provides a labeled version of the XMLHttpRequest (XHR) object, which is used to make network requests. Our system can be extended with constructs such as labeled values, labeled mutable references, clearance, and privileges. For space reasons, we provide details of this, including the soundness proof with the extensions, in \appendixextfirst{}. Here, we instead discuss a limitation of our formalism: the lack of external effects. Specifically, our embedding assumes that the target language does not have any primitives that can induce external effects. As discussed in Section~, imposing this restriction can be challenging. Yet, external effects are crucial when implementing more complex real-world applications. For example, code in an IFC browser must load resources or perform XHR to be useful. Like labeled references, features with external effects must be modeled in the IFC language; we must reason about the precise security implications of features that otherwise inherently leak data. Previous approaches have modeled external effects by internalizing the effects as operations on labeled channels/references~. Alternatively, it is possible to model such effects as messages to/from certain labeled tasks, an approach taken by our Node.js implementation. These ``special'' tasks are trusted with access to the unlabeled primitives that can be used to perform the external effects; since the interface to these tasks is already part of the IFC language, the proof only requires showing that this task does not leak information. Instead of restricting or wrapping unsafe primitives, COWL allow for controlled network communication at the context boundary. (By restricting the default XHR object, for example, COWL allows code to communicate with hosts according to the task's current label.) \section{Related Work} Our information flow control system is closely related to the coarse-grained information systems used in operating systems such as Asbestos~, HiStar~, and Flume~, as well as language-based \emph{floating-label IFC systems} such as LIO~, and Breeze~, where there is a monotonically increased label associated with threads of execution. Our treatment of termination-sensitive and termination-insensitive interference originates from Smith and Volpano~. One information flow control technique designed to handle legacy code is secure multi-execution (SME)~. SME runs multiple copies of the program, one per security level, where the semantics of I/O interactions is altered. Bielova et al.~ use a transition system to describe SME, where the details of the underlying language are hidden. Zanarini et al.~ propose a novel semantics for programs based on interaction trees~, which treats programs as black-boxes about which nothing is known, except what can be inferred from their interaction with the environment. Similar to SME, our approach mediates I/O operations; however, our approach only runs the program once. One of the primary motivations behind this paper is the application of information flow control to JavaScript. Previous systems retrofitted JavaScript with fine-grained \ifextended IFC~. \else IFC~. \fi While fine-grained IFC can result in fewer false alarms and target legacy code, it comes at the cost of complexity: the system must accommodate the entirety of JavaScript's semantics~. By contrast, coarse-grained approaches to security tend to have simpler implications~. The constructs in our IFC language, as well as the behavior of inter-task communication, are reminiscent of distributed systems like Erlang~. In distributed systems, isolation is required due to physical constraints; in information flow control, isolation is required to enforce non-interference. Papagiannis et al.~ built an information flow control system on top of Erlang that shares some similarities to ours. However, they do not take a floating-label approach (processes can find out when sending a message failed due to a forbidden information flow), nor do they provide security proofs. There is limited work on general techniques for retrofitting arbitrary languages with information flow control. However, one time-honored technique is to define a fundamental calculus for which other languages can be desugared into. Abadi et al.~ motivate their core calculus of dependency by showing how various previous systems can be encoded in it. Tse and Zdancewic~, in turn, show how this calculus can be encoded in System F via parametricity. Broberg and Sands~ encode several IFC systems into Paralocks. However, this line of work is primarily focused on static enforcements. \cut{ \begin{itemize} \item Monads and IFC (Abadi~, Tse~, Harrison-Hook~, Crary~, Devriese-Piessens~) \item \Red{Operating systems approaches to IFC, which use coarse grained (Asbestos~, HiStar~)} \item \Red{Floating-label IFC systems, look at LIO paper (LIO~, Breeze~)} \item JavaScript IFC related work, look at BrowBound. (ADSafe, xBook \Red{needs lookup}) Classify approaches into fine-grained (Hedin-Sabelfield~+JSFlow~, FlowFox~, ConDOM~, Austin-Flanagan (sparse information labeling)~) or coarse-grained (BFlow~, see Deian) \item Secure multi-execution (FlowFox~, Russo~) \item Distributed programming languages (Cloud Haskell) plus IFC (Erlang (they check IFC when sending messages, no security proofs, assumes true isolation, not a floating labels)~) \end{itemize} \Red{ Monads and IFC: Basically, monads and IFC go together like toads and holes, or pigs and blankets, etc. But the role the monad plays varies widely. First, you have static systems (Abadi, Tse-Zdancewic), which utilize a lattice of monads: if a value is in a monad with a label that's too high, you're not allowed to unpeel it and look at the value inside. Next, you have dynamic systems. It seems pretty clear that making the IFC TCB be in a monad is a good idea, and the cluster of papers around LIO are all about that. Some of these papers talk about monad transformers: Devriese/Piessens focuses on how you might implement something like LIO by applying a monad transformer to some base monad to run the appropriate restrictions; Harrison/Hook do some more close to what we're doing, where they actually try to say something like "transformers do not interfere." But I can't tell if they're actually doing what we're doing, and their work has a number of unrelated technical restrictions. In more detail: Information Flow Enforcement in Monadic Libraries (Devriese/Piessens): Talks about how to use a monad transformer to take a non-IFC base monad and turn it into a IFC monad, namely the lifting operation should enforce extra constraints. We take closely related ideas, and apply it to settings where there are not any monads. Devriese/Piessens doesn't note the idea that untrusted monad transformers can be applied to add extra effects. Achieving information flow security through monadic control of effects (Harrison/Hook): separation kernel is the basic idea behind our constructed (everything is partioned). Notion of atomic noninterference between layers of effects, ``sufficient condition for atomic noninterference to be inherited through monad transformer application''. But atomic noninterference is not proper noninterference as we've defined here! No proof of separation of kernels, instead prove weaker property no write down. A Core Calculus of Dependency (Abadi et al): define a calculus, which calculi you want to prove IFC secure can be translated into. It's a static system. Why do they use monads? Something about that } IFC for any language: Gotta talk about secure multi-execution } \section{Conclusion} In this paper, we argued that when designing a coarse-grained IFC system, it is better to start with a fully isolated, multi-task system and work one's way back to the model of a single language equipped with IFC. We showed how systems designed this way can be proved non-interferent without needing to rely on details of the target language, and we provided conditions on how to securely refine our formal semantics to consider optimizations required in practice. We connected our semantics to two IFC implementations for JavaScript based on this formalism, explained how our methodology improved an exiting IFC system for Haskell, and proposed an IFC system for C using hardware isolation. By systematically applying ideas from IFC in operating systems to programming languages for which isolation can be achieved, we hope to have elucidated some of the core design principles of coarse-grained, dynamic IFC systems. { \small \paragraph{Acknowledgements} We thank the POST 2015 anonymous reviewers, Adriaan Larmuseau, Sergio Maffeis, and David Mazi\`eres for useful comments and suggestions. This work was funded by DARPA CRASH under contract \#N66001-10-2-4088, by the NSF, by the AFOSR, by multiple gifts from Google, by a gift from Mozilla, and by the Swedish research agencies VR and the Barbro Oshers Pro Suecia Foundation. Deian Stefan and Edward Z. Yang were supported by the DoD through the NDSEG. } { \ifextended \else \frenchspacing\scriptsize \setlength{\bibsep}{2pt} \fi \bibliographystyle{abbrvnat} \bibliography{local} } \ifextended \clearpage \balance \appendix \section{Full Semantics for \ensuremath{\Red{\lambda_{\text{ES}}}}} In Fig.~ we give the full semantics for \ensuremath{\Red{\lambda_{\text{ES}}}}. A subset of them has been given in Fig.~ earlier in the paper. \section{Example IFC Language with a Single Heap} As a concrete instantiation of this proof technique, we show how to make implement our IFC language using a single heap and ensure its non-interference using the techniques presented. First, we can construct the restricted language \ensuremath{L_\text{IFC}^{\mathcal{P}_\text{norefs}}(\alpha,\Red{\lambda_{\text{ES}}})}, where \ensuremath{\mathcal{P}_\text{norefs}} is the family of always valid predicates, except for the ones for \textsc{I-sandbox} and \textsc{I-send}, which we define as $ P(\ensuremath{\ifc{\Varid{e}}}) = (\mathcal{AV}(\ensuremath{\ifc{\Varid{e}}}) = \emptyset{}) $ where $\mathcal{AV}(\ensuremath{\ifc{\Varid{e}}})$ denotes the set of address variables in \ensuremath{\ifc{\Varid{e}}}. That is, we do not restrict any rules except for \textsc{I-sandbox} and \textsc{I-send}. Since $P$ only depends on \ensuremath{\ifc{\Varid{e}}}, which is part of the current task and thus never erased w.r.t.\ the label of the first task, this language satisfies non-interference by Theorem~. The essential parts of the semantics for the concrete language with a single heap, which we call \ensuremath{L_\text{IFC}^{\Red{\text{Heap}}}(\alpha)}, are given in Fig.~. Most rules are straight-forward translations of the rules in Figs.~ and~ but for a single heap. For conciseness, we only show the interesting ones. Now, we can show an isomorphism between this language and \ensuremath{L_\text{IFC}^{\mathcal{P}_\text{norefs}}(\alpha,\Red{\lambda_{\text{ES}}})}, which (by Theorem~ and~) guarantees non-interference for an appropriate scheduling policy \ensuremath{\alpha}. To this end, we represent addresses in the concrete language as pairs $(\ensuremath{\ifc{\Varid{i}}},\ensuremath{\tar{a}})$ where \ensuremath{\ifc{\Varid{i}}} is a task identifier, and \ensuremath{\tar{a}} an address in the abstract system\footnote{Note that this does not make the isomorphism trivial, as in the single heap, there is nothing preventing task 1 to access an address (2,\ensuremath{\tar{a}}). Furthermore, it is common to represent addresses in this way for efficient garbage collection of dead tasks.}. We also formulate the following well-formedness condition for configurations: \[ \ensuremath{\operatorname{wf}(c)} = \forall \ensuremath{\langle \ifc{\Varid{e}}\rangle^{\ifc{\Varid{i}}}_{\ifc{l}}} \in \ensuremath{\ifc{c}}.\ \{ (\ensuremath{\ifc{\Varid{i}}'},\ensuremath{\ifc{\Varid{e}}'}) \in \mathcal{AV}(\ensuremath{\ifc{\Varid{e}}}) \ \vert\ \ensuremath{\ifc{\Varid{i}}} \neq \ensuremath{\ifc{\Varid{i}}'} \} = \emptyset \] Essentially, every address in a given task must have the correct identifier as the first part of the address. It is easy to see that the initial configuration satisfies this condition, and any step in the concrete semantics preserves the condition. Therefore, we only need to consider well-formed configurations, which allows us to give the two required functions \ensuremath{\Varid{f}} and \ensuremath{\Varid{f}^{-1}} for the isomorphism. For conciseness, we only give the interesting parts of their definition, and leave out the straight-forward proof that they actually provide an isomorphism. \begin{itemize} \item Addresses can be directly translated with $\ensuremath{\Varid{f}}((\ensuremath{\ifc{\Varid{i}}},\ensuremath{\tar{a}}))=\ensuremath{\tar{a}}$, and $\ensuremath{\Varid{f}^{-1}}(\ensuremath{\tar{a}})=\ensuremath{(\ifc{\Varid{i}},\tar{a})}$ for an address \ensuremath{\tar{a}} that occurs in task \ensuremath{\ifc{\Varid{i}}}. \item \ensuremath{\Varid{f}} splits the single heap into multiple heaps based on the \ensuremath{\ifc{\Varid{i}}} of the addresses. \ensuremath{\Varid{f}^{-1}} produces a single heap by translating the addresses and collapsing everything to a single store. \end{itemize} \section{Extending the Core Calculus} As mentioned in the main body of this paper, actual IFC implementations may wish to extend the minimal system with more specialized constructs. In this section we show how to extend the language with several such constructs. \subsection{Labeled values} In traditional language-based dynamic IFC systems, a label is associated with values. Hence, a program that, for example, simply writes labeled messages to a labeled log can operate on both public and sensitive values. Similarly, a task that receives a sensitive value and forwards it to another task does not have be be at a sensitive level, if the value is not inspected. In its simplest form, our coarse grained system requires that the current label of a task be at least at the level of the sensitive data to reflect the fact that such data is in scope. If such fine-grained labeling of values is required, our base IFC system can be extended with explicitly labeled values, much like those of LIO and Breeze~: \ensuremath{\ifc{\Varid{v}}\Coloneqq\cdots\;|\;\mathbf{Labeled}\;\ifc{l}\;\ifc{\Varid{e}}}. Following LIO, we say that the expression \ensuremath{\ifc{\Varid{e}}} is protected by label \ensuremath{\ifc{l}}, while the label \ensuremath{\ifc{l}} itself is protected by the task's current label. The label of such values can be inspected the task without requiring the current label to be raised. However, when a task wishes to inspect the protected value \ensuremath{\ifc{\Varid{e}}}, it must first raise its label to at least \ensuremath{\ifc{l}} to reflect that it is incorporating data at such sensitivity level in its scope. When creating labeled values the label \ensuremath{\ifc{l}} must be above the current label; otherwise it cannot be said that protection has been transferred from the current label to \ensuremath{\ifc{l}}. In Fig.~, we formally show how to add this extension to the language. We assume that the constructor \ensuremath{\mathbf{Labeled}} is not part of the surface syntax, but rather an internal construct. \subsection{Labeled mutable references/variables/channels} Extending the calculus with other labeled features, such as references, mutable variables (MVars)~, or channels, can be done in a similar manner: these references are implemented in the IFC language, separately from any preexisting notions of mutable references in the target language. There is some minor additional state to track: specifically, by amending \ensuremath{\ifc{\Sigma}}, as in~\cite{lio, stefan:addressing-covert}, we can allow threads to use these constructs to synchronize, or communicate with constructs other than \ensuremath{\mathbf{send}}/\ensuremath{\mathbf{recv}} in a safe manner. For example, when extending the calculus with labeled references, \ensuremath{\ifc{\Sigma}} additionally contains a store that maps addresses to a value and a label which can be read and written to by different tasks through a labeled reference implementations. In Fig.~ details labeled references formally. The construct \ensuremath{\ifc{\Varid{a}}_{\ifc{l}}} is internal in the labeled reference implementation, and not part of the surface syntax. The changes to the language for labeled values and references require us to update the erasure function \ensuremath{\varepsilon_{\ifc{l}}}, whose full definition is shown in Fig.~. \subsection{Clearance} Systems like LIO, COWL, and Breeze additionally provide a discretionary access control (DAC) mechanism---called \emph{clearance}---at the language level~. This mechanisms is used to restrict a computation from allocating and accessing data (or communicating with entities) above a specified label, the clearance. Amending our IFC language with clearance is straight forward, and, can be done using our notation of a restricted language. To this end, we first extend tasks to track a clearance label alongside the current label, and amend the core IFC language with two new terminals for retrieving and setting this value. Since this extension only adds a per-task mutable variable whose value has no influence on the system, all security guarantees still hold, by essentially the same proofs. However, this does not implement any DAC mechanism yet. To do so, we can restrict the language with a family of predicates $\mathcal{P}_\text{clearance}$: All rules that raise the current label (e.g., \textsc{I-setLabel}), perform allocation (e.g., \textsc{I-sandbox} and \textsf{I-send}), or set the clearance (clearance should not be arbitrarily raised), a predicate that uses the clearance to impose DAC is used. For instance, the predicate for \textsc{I-setLabel} prevents the current label from being raised above the clearance (and thus permit reads above the clearance). The predicate $P := \ensuremath{\ifc{l}\;\flows{}\;\ifc{l}'}$ achieves this restriction, where \ensuremath{\ifc{l}'} is the clearance and \ensuremath{\ifc{l}} is the current label. The other predicates are defined in a similar way and omitted for brevity. \subsection{Privileges} Decentralized IFC extends IFC with the decentralized label model of Myers and Liskov~ to allow for more general applications, including systems consisting of mutually distrustful parties. In a decentralized system, a computation is executed with a set of \emph{privileges}, which, when exercised, allow the computation to declassify data (e.g., by lowering the current label). Practical IFC systems (e.g.,~\cite{Zeldovich:2006, lio, Hritcu:2013:YIB:2497621.2498098, myers:jif}) rely on privileges to implement many applications. The challenge with such an extension lies in the precise security guarantees that must be proved, which to the best of our knowledge is an open research problem. Our implementation for Node.js and COWL both provide privileges, but we have not formalized this part any further. \section{Non-Interference Proof} In this section we prove the theorems we have stated in the paper. Note that we prove soundness of the system including the formally defined extensions from Appendix~. We first observe that the non-interference claims for the languages \ensuremath{L_\text{IFC}(\textsc{Seq},\Red{\lambda})} and \ensuremath{L_\text{IFC}(\textsc{RR},\Red{\lambda})} in Theorems~ and~ follow directly from Theorem~, where the set of predicates is the set of always valid predicates (i.e., no restriction). Before we proceed with the proof of Theorem~, we state and proof two lemmas we will use. \begin{lemma} For any task \ensuremath{\ifc{\Varid{t}}}, task lists \ensuremath{\ifc{\Varid{ts}}}, store \ensuremath{\ifc{\Sigma}}, and label \ensuremath{\ifc{l}}, if $\ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}})}=\ensuremath{\bullet}$, then there exists a task list \ensuremath{\ifc{\Varid{ts}}'} and a store \ensuremath{\ifc{\Sigma}'} such that \begin{align} \ensuremath{\ifc{\Sigma};\ifc{\Varid{t}},\ifc{\Varid{ts}}} \ensuremath{\hookrightarrow} \ensuremath{\ifc{\Sigma}';\ifc{\Varid{ts}},\ifc{\Varid{ts}}'} \\ \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{ts}}')}=\ensuremath{\mathbf{nil}} \\ \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Sigma}')}=\ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Sigma})} \end{align} \end{lemma} \begin{proof} From $\ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}})}=\ensuremath{\bullet}$ we know that the current label \ensuremath{\lcurr} of \ensuremath{\ifc{\Varid{t}}} must be above \ensuremath{\ifc{l}}. Furthermore, tasks can always take a step (if no regular rule applies, then \textsc{I-noStep} can be used), and thus we consider all rules that could be applied to execute \ensuremath{\ifc{\Varid{t}}}. \begin{description} \item[Case \textsc{I-noStep} and \textsc{I-done}] In this case, the task \ensuremath{\ifc{\Varid{t}}} is dropped, and thus \ensuremath{\ifc{\Varid{ts}}'\mathrel{=}\mathbf{nil}} and \ensuremath{\ifc{\Sigma}'\mathrel{=}\ifc{\Sigma}} satisfy conditions~\eqref{eq:hnb-2} and~\eqref{eq:hnb-3}. \item[Case \textsc{I-sandbox}] The newly created task has a label of at least \ensuremath{\lcurr}, and will thus be erased, as required by condition~\eqref{eq:hnb-2}. Furthermore, the state only changes for the newly created thread, and thus the state change is erased, showing~\eqref{eq:hnb-3}. \end{description} In all other rules, no new tasks are created, and thus \ensuremath{\ifc{\Varid{ts}}'} consists of just the one task \ensuremath{\ifc{\Varid{t}}'}, to which \ensuremath{\ifc{\Varid{t}}} executed. Since the tasks label can only increase, \ensuremath{\ifc{\Varid{t}}'} is still erased, showing condition~\eqref{eq:hnb-2}. We are left to show condition~\eqref{eq:hnb-3} for the remaining rules. \begin{description} \item[Case \textsc{I-send}] A new message triple with label \ensuremath{\ifc{l}'} gets added to the message queue of the receiving thread. However, since \ensuremath{\lcurr\;\flows{}\;\ifc{l}'}, the triple will get erased. \item[Case \textsc{I-recv} and \textsc{I-noRecv}] In this case, only the queue of task \ensuremath{\ifc{\Varid{t}}} can change, which gets erased. \item[Case \textsc{I-new}] The newly allocated address has to be at a label at least as high as \ensuremath{\lcurr}, and will thus be erased. \item[Case \textsc{I-write}] Only addresses with a label \ensuremath{\ifc{l}'} above \ensuremath{\lcurr} can be written, thus the change in \ensuremath{\ifc{\Sigma}_{1}} will get erased. \item[Otherwise.] None of the other rules modify the state \ensuremath{\ifc{\Sigma}}, and thus \ensuremath{\ifc{\Sigma}'\mathrel{=}\ifc{\Sigma}} will trivially satisfy condition~\eqref{eq:hnb-3}. \end{description} \qed \end{proof} \begin{lemma} We consider, for any target language \ensuremath{\Red{\lambda}}, the restricted IFC language \ensuremath{L_\text{IFC}^{\mathcal{P}}(\alpha,\Red{\lambda})} (according to Definition~). Then, for any configurations \ensuremath{\ifc{c}_{1}}, \ensuremath{\ifc{c}_{1}'}, \ensuremath{\ifc{c}_{2}}, and label \ensuremath{\ifc{l}} where \begin{equation} \ensuremath{\ifc{c}_{1}} \approx_{\ensuremath{\ifc{l}}} \ensuremath{\ifc{c}_{2}} \qquad \text{and} \qquad \ensuremath{\ifc{c}_{1}} \ensuremath{\hookrightarrow} \ensuremath{\ifc{c}_{1}'} \end{equation} there exists a configuration \ensuremath{\ifc{c}_{2}'} such that \begin{equation} \ensuremath{\ifc{c}_{1}'} \approx_{\ensuremath{\ifc{l}}} \ensuremath{\ifc{c}_{2}'} \qquad \text{and} \qquad \ensuremath{\ifc{c}_{2}} \ensuremath{\hookrightarrow}^* \ensuremath{\ifc{c}_{2}'} \ \text{.} \end{equation} \end{lemma} \begin{proof} First, we observe there must be at least one task in \ensuremath{\ifc{c}_{1}}, otherwise it could not take a step. Thus, \ensuremath{\ifc{c}_{1}} is of the form \ensuremath{\ifc{\Sigma}_{1};\ifc{\Varid{t}}_{1},\ifc{\Varid{ts}}_{1}}. Furthermore, let \ensuremath{\ifc{c}_{2}} be \ensuremath{\ifc{\Sigma}_{2};\ifc{\Varid{ts}}_{2}}. Consider two cases: \begin{itemize} \item $\ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}}_{1})}=\ensuremath{\bullet}$. By the definition of \ensuremath{\varepsilon_{\ifc{l}}}, we know that \ensuremath{\ifc{l}\;\flows{}\;\lcurr} where \ensuremath{\lcurr} is the label of \ensuremath{\ifc{\Varid{t}}_{1}}. In this case, we do not need to take a step for \ensuremath{\ifc{c}_{2}}, because \ensuremath{\ifc{c}_{2}'\mathrel{=}\ifc{c}_{2}} will already be \ensuremath{\ifc{l}}-equivalent to \ensuremath{\ifc{c}_{1}'}. To show this, note that the tasks \ensuremath{\ifc{\Varid{ts}}_{1}} in \ensuremath{\ifc{c}_{1}} are left in the same order and unmodified (the scheduling policy only modifies the first task). The task \ensuremath{\ifc{\Varid{t}}_{1}} either gets dropped (by \textsc{I-noStep}), or transforms into a task \ensuremath{\ifc{\Varid{t}}_{1}'} as well as potentially spawning a new task \ensuremath{\ifc{\Varid{t}}_{1}''}. Since both \ensuremath{\ifc{\Varid{t}}_{1}'} and \ensuremath{\ifc{\Varid{t}}_{1}''} have a label that is at least as high as the label of \ensuremath{\ifc{\Varid{t}}_{1}} (can be seen by inspecting all reduction rules), they will get filtered by \ensuremath{\varepsilon_{\ifc{l}}} in \ensuremath{\ifc{c}_{1}'}. Therefore, the \ensuremath{\ifc{l}}-equivalence of the task list is guaranteed. Lets consider the possible changes to \ensuremath{\ifc{\Sigma}_{1}}: Only five reduction interact with \ensuremath{\ifc{\Sigma}_{1}}, thus it suffices to consider these cases: \begin{description} \item[Case \textsc{I-send}] A new message triple with label \ensuremath{\ifc{l}'} gets added to the message queue of the receiving thread. However, since \ensuremath{\lcurr\;\flows{}\;\ifc{l}'}, the triple will get erased. \item[Case \textsc{I-recv} and \textsc{I-noRecv}] In this case, only the queue of task \ensuremath{\ifc{\Varid{t}}_{1}} can change, which gets erased. \item[Case \textsc{I-new}] The newly allocated address has to be at a label at least as high as \ensuremath{\lcurr}, and will thus be erased. \item[Case \textsc{I-write}] Only addresses with a label \ensuremath{\ifc{l}'} above \ensuremath{\lcurr} can be written, thus the change in \ensuremath{\ifc{\Sigma}_{1}} will get erased. \end{description} This ensures that $\ensuremath{\ifc{c}_{1}'}\approx_{\ensuremath{\ifc{l}}}\ensuremath{\ifc{c}_{2}'}$, as well as $\ensuremath{\ifc{c}_{2}} \ensuremath{\hookrightarrow}^* \ensuremath{\ifc{c}_{2}'}$ (in zero steps), as claimed. \item $\ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}}_{1})}\neq\ensuremath{\bullet}$. By the definition of \ensuremath{\varepsilon_{\ifc{l}}}, the task list \ensuremath{\ifc{\Varid{ts}}_{2}} in \ensuremath{\ifc{c}_{2}} must be of the form \ensuremath{\ifc{\Varid{ts}}_{2}',\ifc{\Varid{t}}_{2},\ifc{\Varid{ts}}_{2}''} (for some task lists \ensuremath{\ifc{\Varid{ts}}_{2}'}, \ensuremath{\ifc{\Varid{ts}}_{2}''} and some task \ensuremath{\ifc{\Varid{t}}_{2}}) where \begin{align} \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{ts}}_{2}')} = \ensuremath{\mathbf{nil}} \\ \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}}_{2})} = \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}}_{1})} \\ \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{ts}}_{2}'')} = \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{ts}}_{1})} \end{align} (where \ensuremath{\mathbf{nil}} is the empty list of tasks). Now, intuitively we will first execute a number of steps to process the tasks in \ensuremath{\ifc{\Varid{ts}}_{2}'} (execute them one step and move them to the back of the task list, or drop them if they are done or stuck). Then, the task \ensuremath{\ifc{\Varid{t}}_{2}} can take the same step as \ensuremath{\ifc{\Varid{t}}_{1}}, which will result in a configuration \ensuremath{\ifc{c}_{2}'} with the desired properties. More formally, we can proceed as follows: First, we can apply Lemma~ continuously for all the task in \ensuremath{\ifc{\Varid{ts}}_{2}'}, until we reach a configuration \ensuremath{\ifc{c}_{2}''\mathrel{=}\ifc{\Sigma}_{2}';\ifc{\Varid{t}}_{2},\ifc{\Varid{ts}}_{2}'',\ifc{\Varid{ts}}_{2}'''} for some \ensuremath{\ifc{\Varid{ts}}_{2}'''} such that \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{ts}}_{2}''')\mathrel{=}\mathbf{nil}} and \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Sigma}_{2})} = \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Sigma}_{2}')}. We note that \ensuremath{\varepsilon_{\ifc{l}}(\ifc{c}_{1})\mathrel{=}\varepsilon_{\ifc{l}}(\ifc{c}_{2}'')} (by the definition of \ensuremath{\varepsilon_{\ifc{l}}}). Now, the first task \ensuremath{\ifc{\Varid{t}}_{2}} in \ensuremath{\ifc{c}_{2}''} is \ensuremath{\ifc{l}}-equivalent to the task \ensuremath{\ifc{\Varid{t}}_{1}}. This implies that the two tasks must have the same id, label and can only differ in the expression or store if some subexpression is of the form \ensuremath{\mathbf{Labeled}\;\ifc{l}'\;\ifc{\Varid{e}}}. In this case, the expression \ensuremath{\ifc{\Varid{e}}} could be different in the two threads if \ensuremath{\lcurr\;\flows{}\;\ifc{l}'}. However, none of the reduction rules depend on an expression in that position, and there is never a hole in that position where evaluation could take place. Thus, the same rules will syntactically match for both task, and we are left to argue that all premises evaluate to the same values for \ensuremath{\ifc{\Varid{t}}_{1}} and \ensuremath{\ifc{\Varid{t}}_{2}}, as well as that the resulting states \ensuremath{\ifc{\Sigma}_{1}'} and \ensuremath{\ifc{\Sigma}_{2}''} are \ensuremath{\ifc{l}}-equivalent. The additional premises $P$ that follow the condition in Definition~ are not a problem, since those predicates only depend on \ensuremath{\varepsilon_{\ifc{l}}(\ifc{c}_{1})}, which is equivalent to \ensuremath{\varepsilon_{\ifc{l}}(\ifc{c}_{2}'')}, and thus those predicates evaluate in the same way. All other premises are either on the threads labels (which are the same), or on the state \ensuremath{\ifc{\Sigma}_{1}}, or \ensuremath{\ifc{\Sigma}_{2}'}, respectively. Because \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Sigma}_{1})\mathrel{=}\varepsilon_{\ifc{l}}(\ifc{\Sigma}_{2}')}, all of these also evaluate in the same way, as can be seen by simply considering all rules that involve or change the state: \begin{description} \item[Case \textsc{I-send}] Here, the task \ensuremath{\ifc{\Varid{t}}_{2}} will send the same message to the same receiver queue. This queue is either completely erased, or it is \ensuremath{\ifc{l}}-equivalent. In both cases, \ensuremath{\ifc{l}}-equivalence of \ensuremath{\ifc{\Sigma}_{1}'} and \ensuremath{\ifc{\Sigma}_{2}'} is preserved. \item[Case \textsc{I-recv} and \textsc{I-noRecv}] When the tasks are receiving a message, then by the reduction rules we know that they first filter the queue by the label \ensuremath{\lcurr} of \ensuremath{\ifc{\Varid{t}}_{1}}. We also know that the queues are equivalent when filtered by the less restrictive label \ensuremath{\ifc{l}}, thus the messages received (or dropped) from the queue are equivalent. \item[Case \textsc{I-new}] The newly allocated address can be the same for both \ensuremath{\ifc{\Varid{t}}_{1}} and \ensuremath{\ifc{\Varid{t}}_{2}}, thus resulting in \ensuremath{\ifc{l}}-equivalent states. \item[Case \textsc{I-write}] By \ensuremath{\varepsilon_{\ifc{l}}(\ifc{\Varid{t}}_{1})\mathrel{=}\Varid{earse}\;\ifc{l}\;\ifc{\Varid{t}}_{2}} both tasks write the same value, and therefore the resulting states will still be \ensuremath{\ifc{l}}-equivalent. \end{description} After \ensuremath{\ifc{\Varid{t}}_{2}} has taken a step, we finally arrive in the desired configuration \ensuremath{\ifc{c}_{2}'\mathrel{=}\ifc{\Sigma}_{2}'';\ifc{\Varid{ts}}_{2}'',\ifc{\Varid{ts}}_{2}''',\ifc{\Varid{ts}}_{2}''''}, where \ensuremath{\ifc{\Varid{ts}}_{2}''''} contains the task resulting from executing \ensuremath{\ifc{\Varid{t}}_{2}} (and might contain, zero (if the task was done or stuck), one (for most steps) or two tasks if a new task was launched). As required, we have \[ \ensuremath{\ifc{c}_{2}} \ensuremath{\hookrightarrow}^* \ensuremath{\ifc{c}_{2}''} \ensuremath{\hookrightarrow} \ensuremath{\ifc{c}_{2}'} \quad \land \quad \ensuremath{\ifc{c}_{1}'} \approx_{\ensuremath{\ifc{l}}} \ensuremath{\ifc{c}_{2}'} \text{.}\] \end{itemize} \qed \end{proof} With this, it is easy to proof Theorem~ as follows. \begin{proof}[Proof of Theorem~, TSNI] We proof the theorem by induction on the length of the derivation sequence in~\eqref{eq:tsni-lhs}. The base case for derivations of length 0 is trivial, allowing us to simple chose $\ensuremath{\ifc{c}_{2}'\mathrel{=}\ifc{c}_{2}}$. In the step case, we assume the theorem holds for derivation sequences of length up to $n$, and show that it also holds for those of length $n+1$. We split the derivation sequence from~\eqref{eq:tsni-lhs} as follows: \[ \ensuremath{\ifc{c}_{1}} \ensuremath{\hookrightarrow} \ensuremath{\ifc{c}_{1}''} \ensuremath{\hookrightarrow}^n \ensuremath{\ifc{c}_{1}'} \] for some configuration \ensuremath{\ifc{c}_{1}''}. By Lemma~, we get \ensuremath{\ifc{c}''} with \begin{equation} \ensuremath{\ifc{c}_{1}''} \approx_{\ensuremath{\ifc{l}}} \ensuremath{\ifc{c}_{2}''} \qquad \text{and} \qquad \ensuremath{\ifc{c}_{2}} \ensuremath{\hookrightarrow}^* \ensuremath{\ifc{c}_{2}''} \end{equation} Applying the induction hypothesis to $\ensuremath{\ifc{c}_{1}''} \ensuremath{\hookrightarrow}^n \ensuremath{\ifc{c}_{1}'}$, we get \ensuremath{\ifc{c}_{2}'} with \begin{equation} \ensuremath{\ifc{c}_{1}'} \approx_{\ensuremath{\ifc{l}}} \ensuremath{\ifc{c}_{2}'} \qquad \text{and} \qquad \ensuremath{\ifc{c}_{2}''} \ensuremath{\hookrightarrow}^* \ensuremath{\ifc{c}_{2}'} \end{equation} Stitching together the derivation sequences from~\eqref{eq:tsni-proof-1} and~\eqref{eq:tsni-proof-2} directly gives us the right-hand side of the implication in the TSNI definition~\eqref{eq:tsni-rhs}, which concludes the proof. \qed \end{proof} \fi |
1501.04134 | Title: Gravitational collapse and formation of universal horizons
Abstract: In this paper, we first generalize the definition of stationary universal
horizons to dynamical ones, and then show that (dynamical) universal horizons
can be formed from realistic gravitational collapse. This is done by
constructing analytical models of a collapsing spherically symmetric star with
finite thickness in Einstein-aether theory.
Body: \newcommand{\bq}{\begin{equation}} \newcommand{\eq}{\end{equation}} \newcommand{\bqn}{\begin{eqnarray}} \newcommand{\eqn}{\end{eqnarray}} \newcommand{\nb}{\nonumber} \newcommand{\lb}{\label} \newcommand{\PRL}{Phys. Rev. Lett.} \newcommand{\PL}{Phys. Lett.} \newcommand{\PR}{Phys. Rev.} \newcommand{\CQG}{Class. Quantum Grav.} \title{Gravitational collapse and formation of universal horizons} \author{Miao Tian $^{a,b}$} \author{Xinwen Wang $^{b}$} \author{ M.F. da Silva $^{c}$} \author{Anzhong Wang $^{b, c, d}$ } \email{Anzhong_Wang@baylor.edu} \affiliation{$^{a}$ School of Mathematics and Physics, Lanzhou Jiaotong University, Lanzhou 730070, China\\ $^{b}$ GCAP-CASPER, Physics Department, Baylor University, Waco, TX 76798-7316, USA\\ $^{c}$ Departamento de F\'isica Te\'orica, Universidade do Estado do Rio de Janeiro, Rua S\~ao Francisco Xavier 524, Maracan\~a, CEP 20550�013, Rio de Janeiro, RJ, Brazil\\ $^{d}$ Institute for Advanced Physics $\&$ Mathematics, Zhejiang University of Technology, Hangzhou 310032, China } \date{\today} \begin{abstract} In this paper, we first generalize the definition of stationary universal horizons to dynamical ones, and then show that (dynamical) universal horizons can be formed from realistic gravitational collapse. This is done by constructing analytical models of a collapsing spherically symmetric star with finite thickness in Einstein-aether theory. \end{abstract} \pacs{04.50.Kd, 04.70.Bw, 04.20.Jb, 97.60.-s} \maketitle \section{ Introduction } \renewcommand{\theequation}{1.\arabic{equation}} \setcounter{equation}{0} The invariance under the Lorentz symmetry group is a cornerstone of modern physics, and is strongly supported by observations. In fact, all the experiments carried out so far are consistent with it , and no evidence to show that such a symmetry must be broken at certain energy scales, although it is arguable that such constraints in the gravitational sector are much weaker than those in the matter sector . Nevertheless, there are various reasons to construct gravitational theories with broken Lorentz invariance (LI) . In particular, our understanding of space-times at Plank scale is still highly limited, and the renomalizability and unitarity of gravity often lead to the violation of LI . One concrete example is the Ho\v{r}ava theory of quantum gravity , in which the LI is broken via the anisotropic scaling between time and space in the ultraviolet (UV), \bq \lb{0.1} t \rightarrow b^{-z} t,\; \;\; x^{i} \rightarrow b^{-1} x^{i}, \; (i = 1, 2, 3), \eq where $z$ denotes the dynamical critical exponent. This is a reminiscent of Lifshitz scalars in condensed matter physics , hence the theory is often referred to as the Ho\v{r}ava-Lifshitz (HL) quantum gravity at a Lifshitz fixed point. The anisotropic scaling () provides a crucial mechanism: The gravitational action can be constructed in such a way that only higher-dimensional spatial (but not time) derivative operators are included, so that the UV behavior of the theory is dramatically improved. In particular, for $z \ge 3$ it becomes power-counting renormalizable . The exclusion of high-dimensional time derivative operators, on the other hand, prevents the ghost instability, whereby the unitarity of the theory is assured . In the infrared (IR) the lower dimensional operators take over, and a healthy low-energy limit is presumably resulted \footnote{It should be emphasized that, the breaking of LI can have significant effects on the low-energy physics through the interactions between gravity and matter, no matter how high the scale of symmetry breaking is . Recently, Pospelov and Tamarit proposed a mechanism of SUSY breaking by coupling a Lorentz-invariant supersymmetric matter sector to non-supersymmetric gravitational interactions with Lifshitz scaling, and showed that it can lead to a consistent HL gravity .}. It is remarkable to note that, despite of the stringent observational constraints of the violation of the LI , the nonrelativistic general covariant HL gravity constructed in is consistent with all the solar system tests and cosmology . In addition, it has been recently embedded in string theory via the nonrelativistic AdS/CFT correspondence . Another version of the HL gravity, the health extension , is also self-consistent and passes all the solar system, astrophysical and cosmological tests in a particular gauge , whereby the consistence of the theory with observations can be deduced.}. Another example that violates LI is the Einstein-aether theory, in which the breaking is realized by a timelike vector field, while the gravitational action is still generally covariant . This theory is consistent with all the solar system tests and binary pulsar observations . However, once the LI is broken, speeds of particles can be greater than that of light. In particular, the dispersion relation generically becomes nonlinear , \bq \lb{0.2} E^2 = c_{p}^2 p^2\left(1 + \alpha_1 \left(\frac{p}{M_{*}}\right)^2 + \alpha_2 \left(\frac{p}{M_{*}}\right)^4\right), \eq where $E$ and $p$ are the energy and momentum of the particle considered, and $c_p, \; \alpha_i$ are coefficients, depending on the species of the particle, while $M_{*}$ denotes the suppression energy scale of the higher-dimensional operators. Then, one can see that both phase and group velocities of the particles are unbounded with the increase of energy. This suggests that black holes may not exist at all in theories with broken LI, and makes such theories questionable, as observations strongly indicate that black holes exist in our universe . Lately, a potential breakthrough was the discovery that there still exist absolute causal boundaries, the so-called {\em universal horizons}, in the theories with broken LI . Particles even with infinitely large velocities would just move around on these boundaries and cannot escape to infinity . The universal horizon radiates like a blackbody at a fixed temperature, and obeys the first law of black hole mechanics . Recently, we studied the existence of universal horizons in the three well-known black hole solutions, the Schwarzschild, Schwarzschild anti-de Sitter, and Reissner-Nordstr\"om, and found that in all of them universal horizons always exist inside their Killing horizons . In particular, the peeling-off behavior of the globally timelike khronon field $u_{\mu}$ was found only at the universal horizons, whereby the surface gravity $\kappa_{\text{peeling}}$ is calculated and found equal to , \bq \lb{0.3} \kappa_{\text{UH}} \equiv \frac{1}{2} u^{\alpha} D_{\alpha} \left(u_{\lambda} \zeta^{\lambda}\right), \eq where $\zeta^{\mu}$ and $D_{\mu}$ denote, respectively, the time translation Killing field and covariant derivative with respect to the given space-time metric $g_{\mu\nu}\; (\mu, \nu = 0, 1, 2, 3)$. For the Schwarzschild solution, the universal horizon and surface gravity are given, respectively, by , \bq \lb{0.4} R_{UH}^{\text{Sch.}} = \frac{3r_s}{4},\;\;\; \kappa_{\text{peeling}}^{\text{Sch.}} = \left(\frac{2}{3}\right)^{3/2}\frac{1}{r_s}, \eq where $r_s$ denotes the Schwarzschild radius. In this paper, we shall study the formation of the universal horizons from realistic gravitational collapse of a spherically symmetric star \footnote{Here ``realistic" means that the collapsing object satisfies at least the weak energy condition .}. To be more concrete, we shall consider such a collapsing object in the Einstein-aether theory . To make the problem tractable, we further assume that the effects of the aether are negligible, so the space-time outside of the star is still described by the Schwarzschild solution, while inside the star we assume that the distribution of the matter is homogeneous and isotropic, so the internal space-time is that of the Friedmann-Robertson-Walker (FRW). Although the model is very ideal, it is sufficient to serve our current purposes, that is, to show explicitly that universal horizons can be formed from realistic gravitational collapse. Specifically, the paper is organized as follows: In Sec. II, we shall present a brief review on the definition of the stationary universal horizons, and then generalize it to dynamical spacetimes. This is realized by replacing Killing horizons by apparent horizons , and in the stationary limit, the latter reduces to the former. In Sec. III, we study a collapsing spherically symmetric star with a finite thickness in the framework of the Einstein-aether theory. When the effects of the aether are negligible, the vacuum space-time outside the star is uniquely described by the Schwarzschild solution and the junction condition across the surface of the star reduces to those of Israel . Once this is done, we find that the khronon equation can be solved analytically when the speed of the khronon is infinitely large, for which the sound horizon of the khronon coincides with the universal horizon. It is remarkable that this is also the case for the Schwarzschild solution , for which the universal horizon and surface gravity are given by Eq.(). The paper is ended in Sec. IV, in which our main conclusions are presented. \section{Dynamical universal horizons and black holes} \renewcommand{\theequation}{2.\arabic{equation}} \setcounter{equation}{0} A necessary condition for the existence of a universal horizon of a given space-time is the existence of a globally time-like foliation . This foliation is usually characterized by a scalar field $\phi$, dubbed khronon , and the normal vector $u_{\mu}$ of the foliation is always time-like, \bq \lb{1.2} u^{\lambda}u_{\lambda} = -1, \eq where \bq \lb{1.1} u_{\mu} = \frac{\phi_{,\mu}}{\sqrt{X}}, \;\;\; \phi_{,\mu} \equiv \frac{\partial \phi}{\partial x^{\mu}}, \;\;\; X \equiv -g^{\alpha\beta}\partial_{\alpha} \phi \partial_{\beta} \phi. \eq In this paper, we choose the signature of the metric as ($-1, 1, 1, 1$). It is important to note that such a defined khronon field is unique only up to the following gauge transformation, \bq \lb{1.3} \tilde{\phi} = {\cal{F}}(\phi), \eq where ${\cal{F}}(\phi)$ is a monotonically increasing (or decreasing) and otherwise arbitrary function of $\phi$. Clearly, under the above gauge transformations, we have $\tilde{u}_{\mu} = u_{\mu}$. The khronon field $\phi$ is described by the action , \bqn \lb{1.4} S_{\phi} &=& \int d^{4}x \sqrt{|g|}\Big[V_0\left(D_{\mu}u_{\nu}\right)^2 + V_0 \left(D_{\mu}u^{\mu}\right)^2\nb\\ && ~~ + c_3 \left(D^{\mu}u^{\nu}\right)\left( D_{\nu}u_{\mu}\right) - c_4 a^{\mu}a_{\mu} \Big], \eqn where $c_i$'s are arbitrary constants, and $a_{\mu} \equiv u^{\alpha}D_{\alpha}u_{\mu}$. The operator $D_{\mu}$ denotes the covariant derivative with respect to the background metric $g_{\mu\nu}$, as mentioned above. Note that the above action is the most general one in the sense that the resulting differential equations in terms of $u_{\mu}$ are second-order . However, when $u_{\mu}$ is written in the form of Eq.(), the relation \bq \lb{1.6} u_{[\nu}D_{\alpha}u_{\beta]} = 0, \eq is identically satisfied. Then, it can be shown that only three of the four coupling constants $c_i$ are independent. In fact, from Eq.() we find , \bq \lb{1.7} \Delta{\cal{L}}_{\phi} \equiv a^{\mu}a_{\mu} + \big(D_{\alpha}u_{\beta}\big)\big(D^{\alpha}u^{\beta}\big) - \big(D_{\alpha}u_{\beta}\big)\big(D^{\beta}u^{\alpha}\big) = 0. \eq Then, one can always add the term, \bq \lb{1.8} \Delta{S}_{\phi} = c_0 \int{\sqrt{|g|} \; d^{4}x \Delta{\cal{L}}_{\phi}}, \eq into $S_{\phi}$, where $c_0$ is an arbitrary constant. This is effectively to shift the coupling constants $c_i$ to ${c}_i'$, where \bq \lb{1.9} {c}_{1}' = c_1 + c_0,\; {c}_{2}' = c_2,\; {c}_{3}' = c_3 - c_0,\; {c}_{4}' = c_4 - c_0. \eq Thus, by properly choosing $c_0$, one can always set one of $c_{i}\; (i = 1, 3, 4)$ to zero. However, in the following we shall leave this possibility open. The variation of $S_{\phi}$ with respect to $\phi$ yields the khronon equation, \bqn \lb{1.11} D_{\mu} {\cal{A}}^{\mu} = 0, \eqn where , \bqn \lb{1.12} {\cal{A}}^{\mu} &\equiv& \frac{\left(\delta^{\mu}_{\nu} + u^{\mu}u_{\nu}\right)}{\sqrt{X}}\AE^{\nu},\nb\\ \AE^{\nu} &\equiv& D_{\gamma} J^{\gamma\nu} + c_4 a_{\gamma} D^{\nu}u^{\gamma},\nb\\ J^{\alpha}_{\;\;\;\mu} &\equiv& \big(V_0g^{\alpha\beta}g_{\mu\nu} + V_0 \delta^{\alpha}_{\mu}\delta^{\beta}_{\nu} + c_3 \delta^{\alpha}_{\nu}\delta^{\beta}_{\mu}\nb\\ && ~~~ - c_4 u^{\alpha}u^{\beta} g_{\mu\nu}\big)D_{\beta}u^{\nu}. \eqn \subsection{ Universal Horizons in Stationary and Asymptotically flat Spacetimes} In stationary and asymptotically flat spacetimes, there always exists a time translation Killing vector, $\zeta^{\mu}$, which is timelike asymptotically, \bq \lb{1.13a} \zeta^{\lambda}\zeta_{\lambda} < 0, \eq for $r \rightarrow \infty$. A {\em Killing horizon} is defined as the existence of a hypersurface on which the time translation Killing vector $\zeta^{\mu}$ becomes null, \bq \lb{1.10} \zeta^{\lambda} \zeta_{\lambda} = 0. \eq On the other hand, a {\em universal horizon} is defined as the existence of a hypersurface on which $\zeta^{\mu}$ becomes orthogonal to $u_{\mu}$, \bq \lb{1.16} u_{\lambda} \zeta^{\lambda} = 0. \eq Since $u_{\mu}$ is timelike globally, Eq.() is possible only when $\zeta_{\mu}$ becomes spacelike. This can happen only inside the apparent horizons, because only in that region $\zeta_{\mu}$ becomes spacelike. \subsection{Universal Horizons in Non-Stationary Spacetimes} To study the formation of universal horizons from gravitational collapse, we need first to generalize the above definition of the universal horizons to non-stationary spacetimes. For the sake of simplicity, in the rest of this paper we shall restrict ourselves only to spherical space-times, and its generalization to other spacetimes is straightforward. The metric for a specially symmetric space-time can be cost in the form, \bq \lb{2.1} ds^2 = g_{ij}dx^idx^j + {R}^2\left(x^i\right)d\Omega^2, \; (i, j = 0, 1), \eq in the spherical coordinates, $x^{\mu} = \left(x^0, x^1, \theta, \varpi\right), \; (\mu = 0, 1, 2, 3)$, where $d\Omega^2 = d\theta^2 + \sin^2\theta d\varpi^2$. The normal vector $n_{\mu}$ to the hypersurface $ {R} = C_0$ is given by, \bq \lb{2.3} n_{\mu} \equiv \frac{\partial({R} - C_0)}{\partial x^{\mu}} = \delta^{0}_{\mu} {R}_{0} + \delta^{1}_{\mu} {R}_{1}, \eq where $C_0$ is a constant and ${R}_{i} \equiv \partial{R} /\partial x^i$. Setting \bq \lb{2.4} \zeta^{\mu} = \delta^{0}_{\mu} {R}_{1} - \delta^{1}_{\mu} {R}_{0}, \eq we can see that $\zeta^{\mu}$ is always orthogonal to $n_{\mu}$, \bq \lb{2.4a} \zeta^{\lambda} n_{\lambda} = 0. \eq For spacetimes that are asymptotically flat there always exists a region, say, ${R} > {R}_{\infty}$, in which $n_{\mu}$ and $\zeta^{\mu} $ are, respectively, space- and time-like, that is, $\left. n_{\mu} n^{\mu}\right|_{{R} > {R}_{\infty}} > 0$ and $\left. \zeta^{\mu} \zeta_{\mu}\right|_{{R} > {R}_{\infty}} < 0$. An apparent horizon may form at ${R}_{AH}$, at which $n_{\mu} $ becomes null, \bq \lb{2.4b} \left. n_{\lambda} n^{\lambda} \right|_{{R} = {R}_{AH}} = 0, \eq where $ {R}_{AH} < {R}_{\infty} $. Then, in the internal region $ {R} < {R}_{AH}$, $n_{\mu} $ becomes timelike. Therefore, we have \bq \lb{2.4c} n_{\lambda} n^{\lambda} = \cases{ > 0, & ${R} > {R}_{AH}$, \cr = 0, & ${R} = {R}_{AH}$, \cr < 0, & ${R} < {R}_{AH}$. \cr} \eq Since Eq.() always holds, we must have \bq \lb{2.4d} \zeta_{\lambda} \zeta^{\lambda} = \cases{ < 0, & ${R} > {R}_{AH}$, \cr = 0, & ${R} = {R}_{AH}$, \cr > 0, & ${R} < {R}_{AH}$, \cr} \eq that is, $\zeta^{\mu}$ becomes null on the apparent horizon, and spacelike inside it. We define a {\em dynamical universal horizon} as the hypersurface at which \bq \lb{2.16} \left. u_{\lambda} \zeta^{\lambda}\right|_{{R} = {R}_{UH}} = 0. \eq Since $u_{\mu}$ is globally timelike, Eq.() is possible only when $\zeta_{\mu}$ is spacelike. Clearly, this is possible only inside the apparent horizons, that is, $ {R}_{UH} < {R}_{AH}$. In the static case, the apparent horizons defined above reduce to the Killing horizons, and the dynamical universal horizons defined by Eq.() are identical to those given by Eq.(). \section{Gravitational Collapse and Formation of Universal Horizons} \renewcommand{\theequation}{3.\arabic{equation}} \setcounter{equation}{0} Let us consider a collapsing star with a finite radius ${R}_{\Sigma}(\tau)$, where $\tau$ denotes the proper time of the surface of the star. To our current purpose, we simply assume that the space-time inside the star is described by the FRW flat metric, \bq \lb{4.1} ds^2_{-} =-dt^2+a^2(t)\left(dr^2+ r^2d\Omega^2\right),\; ({R} \le {R}_{\Sigma}(\tau)), \eq where ${R} =a(t)r$ is the geometric radius inside the collapsing star. From Eq.(), the normal vector $n_{\mu}$ to the hypersurface $ {R} = C_0$ takes the form, \bq n_{\mu} = \dot{a}(t)r \delta^{0}_{\mu}+ a(t)\delta^{1}_{\mu}, \eq where $\dot{a} \equiv da/dt$. Then, the corresponding vector $\zeta^{\mu}$ reads \bq \lb{4.3} \zeta^{\mu} =a(t) \delta^{0}_{\mu} -\dot{a}(t)r \delta^{1}_{\mu}. \eq According to Eq.() the apparent horizon locates at, \bq \lb{4.4} r_{AH} = - \frac{1}{\dot{a}(t)}. \eq Note that in the collapsing case we have $\dot{a} < 0$. The spacetime outside the collapsing star is vacuum. In the Einstein-aether theory , if we consider the case where the effects of the aether field is negligible, then the vacuum space-time will be that of the Schwarzschild, and in the ingoing Eddington-Finkelstein coordinates $(v, {R}, \theta,\varpi)$ the metric takes the form, \bq \lb{4.13} ds^2_{+}=-\left(1-\frac{2M}{{R}}\right)dv^2+2dv d{R}+ {R}^2d\Omega^2,\;\; ({R} \ge {R}_{\Sigma}(\tau)), \eq where $M$ denotes the total mass of the collapsing system, including that of the star surface, and has the dimension of length $L$, that is, $[M] = L$. The surface $\Sigma$ of the star can be parameterized as, \bqn \lb{4.14} r-r_\Sigma&=&0 \; (in\; V^-),\nb\\ {R}-{R}_\Sigma(v)&=&0 \; (in\; V^+), \eqn where $r_\Sigma$ is a constant, when we choose the internal coordinates $(t, r, \theta, \varpi)$ are comoving with the fluid of the collapsing star. On the surface of the collapsing star, the interior and exterior metrics reduce to \bqn \lb{4.15} ds_{-} ^2|_{r=r_\Sigma}&=&-dt^2+a^2(t)r_\Sigma^2d\Omega^2\nb\\ &=& ds_{+} ^2|_{{R}= {R}_\Sigma(v)}\nb\\ &=& -\left(1-\frac{2M}{R_\Sigma(v)}-2\frac{dR_\Sigma(v)}{dv}\right)dv^2 \nb\\ && + R_\Sigma(v)^2d\Omega^2\nb\\ &\equiv &-d\tau^2+R_\Sigma(\tau)^2d\Omega^2, \eqn where $R_\Sigma(\tau)$ is the geometric radius of the collapsing star, $v$ is a function of $\tau$, where $\tau$ denotes the proper time of the observers that are comoving with the collapsing surface of the star. In the current case, we have $\tau=t$. Then, we find \bqn \lb{4.16} R_\Sigma(\tau)=a(\tau)r_\Sigma,\nb\\ \left(1-\frac{2M}{R_\Sigma}\right)\dot{v}^2-2\dot{R}_\Sigma\dot{v}-1=0, \eqn where a dot denotes the derivative with respect to $\tau$. However, since in the present case we have $\tau = t$, so we shall not distinguish it from that with respect to $t$, used above. The extrinsic curvature tensor on the two sides of the surface defined by, \bqn \lb{4.17} K^\pm_{ab}=-n^\pm_\alpha\Big[\frac{\partial^{2}x^\alpha_\pm}{\partial\xi^a\partial\xi^b}+{\Gamma^\pm_{\beta \delta}}^\alpha\frac{\partial x^\beta_\pm}{\partial\xi^a}\frac{\partial x^\delta_\pm}{\partial\xi^b}\Big], \eqn has the following non-vanishing components $ between the one obtained here and that obtained in , because of the sign difference of the cross term $dvdR$ in the external metric ().} \bqn \lb{4.18} K^+_{\tau\tau}&=&-\frac{\ddot{v}}{\dot{v}}-\frac{M\dot{v}}{{R_\Sigma}^2},\nb\\ K^-_{\theta\theta}&=&sin^{-2}\theta K^-_{\varphi\varphi}=a(\tau)r_\Sigma,\\ K^+_{\theta\theta}&=&sin^{-2}\theta K^+_{\varphi\varphi}=(R_\Sigma-2M)\dot{v}-R_\Sigma\dot{R_\Sigma},\nb \eqn where $n^\pm_\alpha$ are the normal vectors defined in the two faces of the surface (), \bqn \lb{4.18a} n^{-}_\alpha &\equiv& \frac{\partial(r - r_{\Sigma})}{\partial x_{-}^{\alpha}} = \delta^{r}_{\alpha},\nb\\ n^{+}_\alpha &\equiv& \frac{\partial(R - R_{\Sigma}(v))}{\partial x_{+}^{\alpha}} = \delta^{R}_{\alpha} -\frac{dR_{\Sigma}(v)}{dv} \delta^{v}_{\alpha}. \eqn From the Israel junction conditions , \bqn \lb{4.19} \left[K_{ab}\right]^{-}-g_{ab}\left[K\right]^-=-8\pi\tau_{ab}, \eqn we can get the surface energy-momentum tensor $\tau_{ab}$, where $\left[K_{ab}\right]^{-} \equiv K^+_{ab}-K^-_{ab}, \; [K]^- \equiv g^{ab}[K_{ab}]^-$, and $g_{ab}$ can be read off from Eq.(), where $a, b = \tau, \theta, \varpi$. Inserting Eq.() into the above equation, we find that $\tau_{ab}$ can be written in the form \bqn \lb{4.20} \tau_{ab}=\sigma w_a w_b +\eta(\theta_a \theta_b+\varpi_a \varpi_b), \eqn where $w_a,\theta_a$ and $ \varpi_a$ are unit vectors defined on the surface of the star, given, respectively, by $w_a=\delta^\tau_a, \; \theta_a=R_\Sigma \delta^\theta_a, \varpi_a= R_\Sigma \sin\theta \delta^\varpi_a$, and \bqn \lb{4.21} \sigma&=&\frac{1}{4\pi r_\Sigma a}+\frac{\dot{a}}{4\pi a}+\frac{M\dot{v}}{2\pi r^2_\Sigma a^2}-\frac{\dot{v}}{4\pi r_\Sigma a},\\ \eta&=& \frac{\dot{v}}{8\pi r_\Sigma a}+\frac{\ddot{v}}{8\pi \dot{v}}-\frac{1}{8\pi r_\Sigma a}-\frac{\dot{a}}{8\pi a}-\frac{M\dot{v}}{8\pi r^2_\Sigma a^2},\nb \eqn here $\sigma$ is the surface energy density of the collapsing star, and $\eta$ its tangential pressure. Physically, they are often required to satisfy certain energy conditions, such as weak, strong and dominant , although in cosmology none of them seems necessarily to be satisfied . On the other hand, inside the collapsing star, the khronon can be parametrized as, \bqn \lb{4.6} u^\mu&=&\sqrt{1+a^2V^2} \delta^{\mu}_{0}+ V\delta^{\mu}_{1},\nb\\ u_\mu&=&-\sqrt{1+a^2V^2} \delta^{\mu}_{0}+ a^2V\delta^{\mu}_{1}, \eqn where $V = V(t, r)$ is determined by the khronon equation (), which now reduces to \bq \lb{4.5} {\cal{A}}^{r}_{,r} + \frac{2{\cal{A}}^r}{r}+ {\cal{A}}^{t}_{,t} + \frac{3\dot{a}(t){\cal{A}}^t}{a(t)}= 0, \eq where \bqn \lb{4.7} {\cal{A}}^t&=&c_{123}{\cal{A}}^t_1+c_{14}{\cal{A}}^t_2+c_{13}{\cal{A}}^t_3\nb\\ {\cal{A}}^{r} &=& \frac{\sqrt{1+a^2V^2}} {a^2V} {\cal{A}}^t,\;\;\; {\cal{A}}^{\theta} = {\cal{A}}^{\phi} = 0, \eqn with $c_{ab}\equiv c_a +c_b,\; c_{abc}\equiv c_a +c_b + c_c$, and \bqn \lb{4.7a} {\cal{A}}^t_1&=&\frac{V}{r^2(1+a^2V^2)}(2a^4V^5 ( -1+2r^2a\ddot{a})\nb \\ &&+r(r V''+ V'(2+r a^2\sqrt{1+a^2V^2}\dot{V}))\nb\\ &&+V(-2-3r^2\dot{a}^2+3r^2a\ddot{a}+5r^2a\sqrt{1+a^2V^2}\dot{a}V'\nb\\ &&+r^2a^4\dot{V}^2+2ra^2\sqrt{1+a^2V^2}(\dot{V}+r\dot{V'}))\nb\\ &&+a^2V^3(-4+r(-2r\dot{a}^2+4ra\sqrt{1+a^2V^2}\dot{a}V'\nb\\ &&+a(7r\ddot{a}+2a\sqrt{1+a^2V^2}(\dot{V}+r\dot{V'}))))\nb\\ &&+ra^4V^4(2V'+r(V''+a(5\dot{a}\dot{V}+a\ddot{V})))\nb\\ &&+ra^2V^2(4V'+r(2V''+a(7\dot{a}\dot{V}+a\ddot{V}))) ), \eqn \bqn \lb{4.7b} {\cal{A}}^t_2&=&-\frac{V}{r(1+a^2V^2)}(4rV\dot{a}^2+ra^5V^4(2V\ddot{a}+5\dot{a}\dot{V}) \nb\\ &&+2a^3V^2(2V^2\sqrt{1+a^2V^2}\dot{a}+2rV(\ddot{a} \nb\\ &&+\sqrt{1+a^2V^2}\dot{a}V')+5r\dot{a}\dot{V})+a(4V^2\sqrt{1+a^2V^2}\dot{a}\nb\\ &&+ r V(2\ddot{a}+5\sqrt{1+a^2V^2}\dot{a}V')+5r\dot{a}\dot{V})\nb\\ &&+ra^6V^4\ddot{V}+a^4V^2(4rV^3\dot{a}^2+V^2(2V'+r V'')\nb\\ &&+2V\sqrt{1+a^2V^2}(\dot{V}+r\dot{V'})+2r\ddot{V})\nb\\ &&+a^2(8rV^3\dot{a}^2+V^2(2V'+r V'')+V(rV^2\nb\\ &&+2\sqrt{1+a^2V^2}(\dot{V}+r\dot{V'}))\nb\\ &&+r(\sqrt{1+a^2V^2}V'\dot{V}+\ddot{V})) ),\\ \nb\\ {\cal{A}}^t_3&=&2V^2(1+a^2V^2)(\dot{a}^2-a\ddot{a}). \eqn Here a prime denotes the derivative with respect to $r$. It is found very difficult to solve Eq.() for any given coupling constants $c_i$. However, when $c_{14}=0$, we obtain a particular solution, \bqn \lb{4.11} V(t,r)=\frac{V_0 r}{a(t)}, \;\;\; a(t)=a_0e^{-Ht}, \eqn where $V_0, \; H$ and $a_0$ are integration constants with $H> 0, \; a_0 > 0$. It is remarkable to note that $c_{14} = 0$ corresponds to the case in which the speed of the khronon becomes infinitely large $c_{\phi}^2 = c_{123}/c_{14} \rightarrow \infty$, a case that was also studied in . From the definition of the dynamical universal horizon Eq.() and considering Eqs.(), () and (), we find that the collapse always forms a universal horizons inside the collapsing star, and its location is given by, \bq \lb{4.12} r_{UH}(t) = \sqrt{\frac{2}{\sqrt{V_0^4+4V_0^2 a_0^2 H^2 e^{-2H t}} -V_0^2}}. \eq From Eq.(), on the other hand, we find that, \bqn \lb{4.22} \dot{v}=\frac{-r^2_\Sigma a \dot{a}+\sqrt{r^2_\Sigma a^2+r^2_\Sigma a^2 \dot{a}^2-2a M r_\Sigma}}{2M-a r_\Sigma}. \eqn Substituting it together with $a(\tau)=a_0e^{-H\tau}$ back into Eqs.(), we obtain \bqn \lb{4.23} &\sigma&=\frac{1}{4\pi R_\Sigma}\left(1+{\cal{G}}\right),\nb\\ &\eta&= \frac{1}{8\pi R_\Sigma}\left(\frac{M-R_\Sigma -2H^2 R^3_\Sigma}{R_\Sigma {\cal{G}}} - 1\right), ~~~~~ \eqn where \bq \lb{4.23a} R_\Sigma = a_0 r_\Sigma e^{-H\tau},\;\;\; {\cal{G}} \equiv \sqrt{1+H^2 R^2_\Sigma -\frac{2M}{R_\Sigma }}. \eq Obviously, to have both $\sigma$ and $\eta$ real, we must assume that $R_\Sigma \ge R_\Sigma^{Min}$, where $ R_\Sigma^{Min}$ is a root of the equation, \bq \lb{4.23b} 1+H^2 R^2_\Sigma -\frac{2M}{R_\Sigma } = 0. \eq When the star collapses to the point $R_\Sigma\left(\tau_{Min}\right) = R_\Sigma^{Min}$, the tangential pressure diverges, whereby a space-time singularity (with a finite radius) is developed. This represents the end of the collapse, as the space-time beyond this moment is not extendable. It can be shown that the weak energy condition $\sigma\geq 0,\; \sigma+\eta\geq 0$ can always be satisfied by properly choosing the free parameters involved in the solution, before the formation of the universal horizon. In particular, from Eq.() we evan see that $\sigma$ is always non-negative, and \bqn \lb{4.23c} \sigma + \eta =\frac{1}{8\pi R_\Sigma {\cal{G}}}\left[{\cal{G}} - \left(\frac{3M}{R_\Sigma} -1\right)\right]. \eqn Thus, for $R_\Sigma \ge 3M$, the weak energy condition is always satisfied. When $R_\Sigma < 3M$, it is also satisfied, provided that ${\cal{G}} \ge {3M}/{R_\Sigma} -1$, or equivalently \bq \lb{4.23d} H^2R_\Sigma^2 + \frac{4M}{R_\Sigma} \ge \frac{9M^2}{R_\Sigma^2}. \eq Clearly, by properly choosing the free parameters involved in the solution, this condition can hold until the moment where the whole collapsing star is inside the universal horizon. However, at the end $ \tau = \tau_{Min}$ of the collapse this condition is necessarily violated, as can be seen from Eqs.() and (). Using the geometric radius $R=r a(t)=r a_0 e^{-H t}$ inside the collapsing star, we find that the apparent horizon given by Eq.() and the universal horizon given by Eq.() can be expressed as, \bqn \lb{4.24} R_{UH}&=&r_{UH} a(t) \nb\\ &=& a_0 e^{-H t}\sqrt{\frac{2}{\sqrt{V_0^4+4V_0^2 a_0^2 H^2 e^{-2H t}}-V_0^2}}, \nb\\ R_{AH}&=&r_{AH} a(t)= \frac{1}{H}. \eqn In Fig. we show one of the cases, in which the free parameters are chosen as $r_\Sigma=1, M = 1, a_0 = 3, V_0 = 0.6, H = 1.5$. Then, we find that the weak energy condition holds until the moment $t=0.624554$, at which we have $ R_{UH}= 1.24387 > R_\Sigma=1.1756$. That is, the weak energy condition holds all the way down to the moment when the whole star collapses inside the universal horizon, as it is illustrated clearly in Fig.. In this figure, three horizontal lines, $ R = 2M, \; 3M/2, \; R^{Min}_{\Sigma}$ are also plotted, where $R_{UH}^{\text{Sch.}} = 3M/2$ is the universal horizon in the Schwarzschild space-time [cf. Eq.()]. For the current choice of the free parameters, the universal horizon is not continuous. In fact, when the star collapses to the moment $t = t_o$, where $t_o$ is given by $R_{\Sigma}(t_o) = 3M/2$, the universal horizon jumps from $R_{UH}(t_o)$ to $3M/2$, as shown more clearly in Fig. . Physically, this is because that the surface shell of the collapsing star has non-zero mass. The collapse ends at $t = t^{Min}_{\Sigma}$, where $R_{\Sigma}( t^{Min}_{\Sigma}) = R^{Min}_{\Sigma}$, at which the surface pressure of the collapsing star becomes infinitely large. It is remarkable that $R^{Min}_{\Sigma}$ generically is different from zero, that is, the collapse generically forms a space-time singularity that has finite radius. The corresponding Penrose diagram is given in Fig. . In this figure the location of the event horizon denoted by the straight line $EH$ is also marked, although it can be penetrated by particles with sufficiently large velocities, and propagate to infinity, even they are initially trapped inside it. However, this is no longer the case when across the universal horizon. As explained above, once they are trapped inside the universal horizon, they cannot penetrate it and propagate to infinities, even they are moving with infinitely large velocities. As a result, an absolutely black region is (classically) formed from the gravitational collapse of a massive star, and this region is black, even in theories that allow instantaneous propagations! \section{Conclusions} In this paper, we have first generalized the definition of a stationary universal horizon to a dynamical one, by simply replacing Killing horizons by apparent ones. Then, we have constructed an analytical model that represents the gravitational collapse of a spherical symmetric star with finite thickness, and shown explicitly that dynamical universal horizons can be formed from such a ``realistic" gravitational collapse. Here ``realistic" is referred to as a gravitational collapse of a star with a finite thickness that satisfies at least the weak energy condition . To have the problem tractable, we have assumed that the star consists of an anisotropic and homogeneous perfect fluid and that outside the star the space-time is vacuum in the framework of the Einstein-aether theory . When the effects of the aether field is negligible, the vacuum space-time is uniquely described by the Schwarzschild solution . Even in this case, solving the khronon equation () inside the star is still very complicated. Instead, we have further assumed that the velocity of the the khronon is infinitely large, so the sound horizon of the khronon coincides with the universal horizon. Then, we have found that an analytical solution exists for the star made of the de Sitter universe, and shown explicitly how a dynamical universal horizon can be formed, as can be seen from Figs. - . Although such a constructed model serves our current purpose very well, that is, to show that universal horizons can be indeed formed from realistic gravitational collapse, it would be very interesting to consider cases without (some of) the above assumptions, specially the case in which the space-time outside of the star is not vacuum, so that the star may radiate, when it is collapsing. In addition, although in this paper we have considered the formation of the universal horizons only in the framework of the Einstein-aether theory, it is expected that our main conclusions should be true in other theories of gravity with broken LI, including the HL gravity . \section*{Acknowledgements} This work was done partly when A.W. was visiting the State University of Rio de Janeiro (UERJ). He would like to thank UERJ for hospitality. It was done also during the visit of M.T. to Baylor University. M.T. would like to express his gratitude to Baylor. This work is supported in part by Ci\^encia Sem Fronteiras, Grant No. A045/2013 CAPES, Brazil (A.W.); NSFC Grant No. 11375153, China (A.W.); CAPES, CNPq and FAPERJ Brazil (M.F.A.S.); CSC Grant No. 201208625022, China (M.T.); and Baylor Graduate Scholarship (X.W.). \begin{thebibliography}{nbound} \bibitem{Liberati13} A. Kostelecky and N. Russell, Rev. Mod. Phys. {\bf 83} 11 (2011) [arXiv:0801.0287v7, January 2014 Edition]. \bibitem{LZbreaking} D. Mattingly, Living Rev. Relativity, {\bf 8}, 5 (2005); S. Liberati, Class. Qnatum Grav. {\bf 30}, 133001 (2013). \bibitem{LACW} K. Lin, E. Abdalla, R.-G. Cai, and A. Wang, Inter. J. Mod. Phys. D{\bf 23}, 1443004 (2014); K. Lin, F.-W. Shu, A. Wang, and Q. Wu, Phys. Rev. D{\em in press} (2015) [arXiv:1404.3413]. \bibitem{QGs} S. Carlip, {\em Quantum Gravity in 2+1 Dimensions}, Cambridge Monographs on Mathematical Physics (Cambridge University Press, Cambridge, 2003); C. Kiefer, {\em Quantum Gravity} (Oxford Science Publications, Oxford University Press, 2007). \bibitem{Horava} P. Ho\v{r}ava, J. High Energy Phys. {\bf 0903}, 020 (2009); Phys. Rev. D{\bf 79}, 084008 (2009). \bibitem{Lifshitz} E.M. Lifshitz, Zh. Eksp. Toer. Fiz. {\bf 11}, 255 (1941); {\em ibid.}, {\bf 11}, 269 (1941). \bibitem{Visser} M. Visser, Phys. Rev. D{\bf 80}, 025011 (2009); {\em Power-counting renormalizability of generalized Ho\v{r}ava gravity}, arXiv:0912.4757. \bibitem{Stelle} K.S. Stelle, Phys. Rev. D{\bf 16}, 953 (1977). \bibitem{Collin04} J. Collins, A. Perez, D. Sudarsky, L. Urrutia, and H. Vucetich, Phys. Rev. Lett. {\bf 93}, 191301 (2004). \bibitem{PT14} M. Pospelov and C. Tamarit, J. High Energy Phys. {\bf 01} (2014) 048. \bibitem{ZWWS} T. Zhu, Q. Wu, A. Wang, and F.-W. Shu, Phys. Rev. D{\bf 84}, 101502(R) (2011); T. Zhu, F.-W. Shu, Q. Wu, and A. Wang, Phys. Rev. D{\bf 85}, 044053 (2012). \bibitem{LMWZ} K. Lin, S. Mukohyama, A. Wang, and T. Zhu, Phys. Rev. D{\bf 89}, 084022 (2014). \bibitem {Will} C.M. Will, {\em The Confrontation between General Relativity and Experiment}, Living Rev. Relativity, {\bf 9}, 3 (2006) [http://www.livingreviews.org/lrr-2006-3; arXiv:gr-qc/0510072]. \bibitem{ZHW} T. Zhu, Y.-Q. Huang, and A. Wang, Phys. Rev. D{\bf 87}, 084041 (2013); A. Wang, Q. Wu, W. Zhao, and T. Zhu, Phys. Rev. D{\bf 87}, 103512 (2013); T. Zhu, W. Zhao, Y.-Q. Huang, A. Wang, and Q. Wu, Phys. Rev. D{\bf 88}, 063508 (2013). \bibitem {cosmo} E. Komatsu, {\em et al.}, Astrophys. J. Suppl. {\bf 192}, 18 (2011); P. Ade et al., Planck Collaboration, Astron. $\&$ Astrophys., (2014). \bibitem{JK} S. Janiszewski and A. Karch, JHEP, {\bf 02}, 123 (2013); Phys. Rev. Lett. {\bf 110}, 081601 (2013). \bibitem{BPS} D. Blas, O. Pujolas, and S. Sibiryakov, Phys. Lett. B{\bf 688}, 350 (2010); J. High Energy Phys. {\bf 1104}, 018 (2011). \bibitem{EA} T. Jacobson and Mattingly, Phys. Rev. D{\bf 64}, 024028 (2001); T. Jacobson, Proc. Sci. QG-PH, {\bf 020} (2007). \bibitem{Jacob} T. Jacobson, Phys. Rev. D{\bf 81}, 101502 (R) (2010). \bibitem{Wang13} A. Wang, {\em On ``No-go theorem for slowly rotating black holes in Ho\v{r}ava-Lifshitz gravity}, arXiv:1212.1040. \bibitem {Yagi} K. Yagi, D. Blas, N. Yunes, and E. Barausse, Phys. Rev. Lett. {\bf 112}, 161101 (2014); K. Yagi, D. Blas, E. Barausse, and N. Yunes, Phys. Rev. D{\bf 89}, 084067 (2014). \bibitem{NM} R. Narayan and J.E. MacClintock, Mon. Not. R. Astron. Soc., {\bf 419}, L69 (2012). \bibitem{BS11} D. Blas and S. Sibiryakov, Phys. Rev. D{\bf 84}, 124043 (2011). \bibitem{UHs} E. Barausse, T. Jacobson, and T. Sotiriou, Phys. Rev. D{\bf 83}, 124043 (2011); B. Cropp, S. Liberati, and M. Visser, Class. Quantum Grav. {\bf 30}, 125001 (2013); M. Saravani, N. Afshordi, and R.B. Mann, Phys. Rev. D{\bf 89}, 084029 (2014); S. Janiszewski, A. Karch, B. Robinson, and D. Sommer, JHEP {\bf 04}, 163 (2014); C. Eling and Y. Oz, JHEP, {\bf 11}, 067 (2014); T. Sotiriou, I. Vega, and D. Vernieri, Phys. Rev. D{\bf 90}, 044046 (2014); J. Bhattacharyya and D. Mattingly, ``{\em Universal horizons in maximally symmetric spaces}," arXiv:1408.6479. \bibitem{BBM} P. Berglund, J. Bhattacharyya, and D. Mattingly, Phys. Rev. D{\bf 85}, 124019 (2012); Phys. Rev. Lett. {\bf 110}, 071301 (2013). \bibitem{LGSW} K. Lin, O. Goldoni, M.F. da Silva, and A. Wang, Phys. Rev. D {\em in press} (2015) [arXiv:1410.6678]. \bibitem{CLMV} B. Cropp, S. Liberati, A. Mohd, and M. Visser, Phys. Rev. D{\bf 89}, 064061 (2014). \bibitem{HE73} S.W. Hawking, G.F.R. Ellis, {\em The Large Scale Structure of Spacetime} (Cambridge University Press, Cambridge, 1973). \bibitem{Hayd} S.A. Hayward, Phys. Rev. {\bf D49}, 6467 (1994); Class. Quantum Grav. {\bf 17}, 1749 (2000). \bibitem{Wang} A. Wang, Phys. Rev. D{\bf 68}, 064006 (2003); Phys. Rev. D{\bf 72}, 108501 (2005); Gen. Relativ. Grav. {\bf 37}, 1919 (2005); Y. Wu, M.F.A. da Silva, N.O. Santos, and A. Wang, Phys. Rev. D{\bf 68}, 084012 (2003); A.Y. Miguelote, N.A. Tomimura, and A. Wang, Gen. Relativ. Grav. {\bf 36}, 1883 (2004); P. Sharma, A. Tziolas, A. Wang, and Z.-C. Wu, Inter. J. Mord. Phys. A{\bf 26}, 273 (2011). \bibitem{IC66} W. Israel, Nuovo Cimento, B 44 (1966) 1; B 48 (1967) 463 (E); A. Wang and N.O. Santos, Inter. J. Mod. Phys. A{\bf 25}, 1661 (2010). \bibitem{CS97} W.B. Bonnor, A.K.G. de Oliveira, and N.O. Santos, Phys. Rept. {\bf 181}, 269 (1989). \bibitem{Cosmo} L. Amendola and S. Tsujikawa, {\em Dark Energy: Theory and Observations} (Cambridge University Press, Cambridge, 2010). \end{thebibliography} |
1501.04135 | Title: Topology of density matrices
Abstract: We investigate topological properties of density matrices motivated by the
question to what extent phenomena like topological insulators and
superconductors can be generalized to mixed states in the framework of open
quantum systems. The notion of geometric phases has been extended from pure to
mixed states by Uhlmann in [Rep. Math. Phys. 24, 229 (1986)], where an emergent
gauge theory over the density matrices based on their pure-state representation
in a larger Hilbert space has been reported. However, since the uniquely
defined square root $\sqrt{\rho}$ of a density matrix $\rho$ provides a global
gauge, this construction is always topologically trivial. Here, we study a more
restrictive gauge structure which can be topologically non-trivial and is
capable of resolving homotopically distinct mappings of density matrices
subject to various spectral constraints. Remarkably, in this framework,
topological invariants can be directly defined and calculated for mixed states.
In the limit of pure states, the well known system of topological invariants
for gapped band structures at zero temperature is reproduced. We compare our
construction with recent approaches to Chern insulators at finite temperature.
Body: \title{Topology of density matrices} \author{Jan Carl Budich} \affiliation{Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria} \affiliation{Institute for Quantum Optics and Quantum Information, Austrian Academy of Sciences, 6020 Innsbruck, Austria} \author{Sebastian Diehl} \affiliation{Institute of Theoretical Physics, TU Dresden, D-01062 Dresden, Germany} \date{\today} \begin{abstract} We investigate topological properties of density matrices motivated by the question to what extent phenomena such as topological insulators and superconductors can be generalized to mixed states in the framework of open quantum systems. The notion of geometric phases has been extended from pure to mixed states by Uhlmann in [Rep. Math. Phys. {\bf{24}}, 229 (1986)], where an emergent gauge theory over the density matrices based on their pure-state representation in a larger Hilbert space has been reported. However, since the uniquely defined square root $\sqrt{\rho}$ of a density matrix $\rho$ provides a global gauge, this construction is always topologically trivial. Here, we study a more restrictive gauge structure which can be topologically non-trivial and is capable of resolving homotopically distinct mappings of density matrices subject to various spectral constraints. Remarkably, in this framework, topological invariants can be directly defined and calculated for mixed states. In the limit of pure states, the well known system of topological invariants for gapped band structures at zero temperature is reproduced. We compare our construction with recent approaches to Chern insulators at finite temperature. \end{abstract} \maketitle \section{Introduction} With the identification of geometric phases in quantum physics , an emergent classical gauge structure in elementary quantum mechanics has been revealed. Several fundamental discoveries of contemporary quantum physics such as the integer quantum Hall effect and, more recently, the periodic table of topological insulators and superconductors can be theoretically described as topological properties of such emergent gauge theories in the framework of Bloch bands: The local $U(n)$ gauge degree of freedom there consists in the choice of an arbitrary orthonormal basis of Bloch states $\lvert u_k^\alpha\rangle$ that span the projection $P(k)=\sum_{\alpha=1}^{n}\lvert u_k^\alpha\rangle\langle u_k^\alpha\rvert$ onto the $n$ occupied bands of an insulating band structure at lattice momentum $k$. The covariant derivative for such gauge theories has been constructed by Kato as early as 1950 in his proof of the adiabatic theorem of quantum mechanics . Topological invariants of this gauge structure distinguish homotopically distinct mappings $k\mapsto P(k)$, i.e., topologically inequivalent band structures. The consideration of generic symmetries refines this homotopy classification and leads to the periodic table of topological invariants (see Ref. for a recent rigorous discussion).\\ Real physical systems, however, are not perfectly isolated from their environment and are to be described by a mixed state density matrix $\rho$ rather than a pure state wave function. It is hence natural to ask whether the concept of geometric phases in quantum physics can be generalized to the realm of density matrices and their quantum mechanical evolution. This question has been answered in the affirmative in a series of papers by Uhlmann , who viewed the redundancy in the purification of density matrices, i.e., in the representation of $\rho$ as a pure state in a larger Hilbert space, as a gauge degree of freedom: $\rho=w w^\dag=(wU)(wU)^\dag$, where the Hilbert Schmidt operator $w$ denotes a purification of $\rho$ and $U$ is a unitary operator. In this context, the Uhlmann-connection defining parallel transport of Hilbert Schmidt operators along a path $t\mapsto \rho(t)$ of density matrices is given by the constraint \begin{align} \dot w^\dag w-w^\dag \dot w=0, \end{align} where the dot denotes the derivative with respect to $t$.\\ The purpose of our present work is to unravel the topological properties of various gauge structures over the space of quantum mechanical density matrices. Since the purification $w=\sqrt{\rho}$, uniquely defined for every density matrix $\rho$, provides a global gauge, the general purification scheme in terms of Hilbert Schmidt operators as considered by Uhlmann is always topologically trivial. However, we would like to point out that this scheme can also be applied to density matrices of pure states which then would also yield a topologically trivial gauge theory -- despite the fact that pure states give rise to a rich spectrum of topological phenomena as mentioned above. Thus, the fact that some topologically trivial gauge structure can be constructed over the space of density matrices by no means implies that there are no topological features to be discovered. In fact, more generally speaking, every topological gauge theory can be embedded into a topologically trivial theory with a larger gauge degree of freedom over the same base space . This raises the natural question whether a more restricted notion than general purification of density matrices in terms of Hilbert Schmidt operators could be employed to reveal topological aspects of families of density matrices $\rho(k)$ parameterized by a lattice momentum $k$.\\ {{\emph{Key results --}}} We investigate an ensemble of pure states (EOPS) scheme in terms of non-orthonormalized pure states $\lvert \tilde \psi^\alpha\rangle$ satisfying \begin{align} \rho=\sum_\alpha \lvert \tilde \psi^\alpha\rangle\langle \tilde \psi^\alpha\rvert \end{align} as a gauge structure over the space of density matrices $\rho$. This construction draws intuition from the entanglement theory of mixed states where, e.g., the entanglement of formation is defined in terms of similar ensembles of pure states. Starting from the spectral representation $\rho=\sum_\alpha p_\alpha \vert \psi_\alpha\rangle\langle \psi_\alpha\rvert$, a natural EOPS is $\lvert \tilde \psi_\alpha \rangle = \sqrt{p_\alpha} \lvert \psi_\alpha\rangle$. Under arbitrary unitary rotations of this frame, the pure states are no longer mutually orthogonal but still project down to the same density matrix $\rho$ via Eq. (). The resulting gauge theory of non-orthogonal frames over the density matrices is a generalization of the pure state case. For invertible density matrices without any further spectral assumptions, it is equivalent to Uhlmann's construction and thus topologically trivial. Interestingly, if assumptions on the spectral degeneracy of the density matrices are made, the present scheme can accommodate topologically non-trivial mixed states. Topologically inequivalent mappings $k\mapsto \rho(k)$ in this framework cannot be continuously deformed into each other without either violating the spectral assumptions or breaking possible protecting physical symmetries. Our construction formalizes the notion of a purity gap proposed in Refs. -- a purity gap closing is a level-crossing in $\rho$ and thus a violation of the underlying spectral assumptions that makes a change in the topology possible. Under non-equilibrium conditions, a physical system may be non-ergodic and may thus be described by a singular density matrix of rank $n<N$, where $N$ is the dimension of the Hilbert space. Also in this case, the EOPS gauge structure can become topologically non-trivial. The topological invariants associated with the proposed gauge structure are constructed in terms of Uhlmann's connection (see Eq. ()) without reference to the pure state case which is, however, correctly reproduced as a limiting case here. Finally, we discuss notable differences to recent publications where topological invariants for two-banded one-dimensional systems and Chern insulators at finite temperature, respectively, have been proposed that are not characteristic classes of a gauge theory. We illustrate with an explicit example that a two-dimensional system at finite temperature is in general not uniquely characterized by a single invariant of that kind.\\ {{\emph{Outline}--}} This article is structured as follows. In Section , we review how a gauge structure emerges for both pure and mixed states in elementary quantum mechanics. In Section , topological invariants for density matrices with various spectral assumptions are defined. We compare the present approach to the construction of Uhlmann phase winding numbers that have recently been proposed to classify thermal Chern insulators in Section . Concluding remarks are presented in Section . \section{Emergent gauge structures from pure to mixed states} \subsection{Adiabatic time evolution of pure states} A natural and conceptually simple scenario for the occurrence of geometric phases is the time evolution of a quantum mechanical system with a Hamiltonian $H(R(t))$ that depends adiabatically on time via some control parameters $R(t)$ . As a side remark, we note that the concept of geometric phases has been generalized to non-adiabatic and non-cyclic evolution. The non-degenerate ground state of $H(R)$ is denoted by $\lvert R\rangle$. Since the system cannot leave its instantaneous ground state when initially prepared in $\lvert {R(t_0)}\rangle$, the geometric constraint \begin{align} P(t)\lvert \Psi(t)\rangle=\lvert \Psi(t)\rangle \end{align} with $P(t)= \lvert{R(t)}\rangle\langle {R(t)}\rvert$ the projection onto the ground state, is imposed on the solution $\lvert \Psi(t)\rangle$ of the time dependent Schr\"odinger equation in the adiabatic limit. This constraint, reflecting the absence of any dynamical level transitions, implies the form $\lvert \Psi(t)\rangle=\text{e}^{i(\phi(t)-\phi_D(t))}\lvert {R(t)}\rangle$, where $\lvert {R(t)}\rangle$ is a family of ground states with a smooth relative phase, i.e., a gauge of ground states. $\phi_D(t)=\int_{t_0}^t \text{d} \tau E_0(\tau)$ with the ground state energy $E_0$ is called the dynamical phase. The additional phase factor $\text{e}^{i\phi(t)}$ of the state vector $\lvert\psi(t)\rangle=\text{e}^{i\phi_D(t)}\lvert \Psi(t)\rangle$ relative to the gauge $\lvert R(t)\rangle$ reveals a deep geometric ``principle of least effort" that nature employs in the adiabatic limit of quantum mechanics: $\lvert\psi(t)\rangle$ is the shortest path in Hilbert space that satisfies Eq. (), shortest as measured by $L=\int_{t_0}^t\text{d}\,\tau \sqrt{\langle \dot \psi(\tau) \vert \dot \psi(\tau)\rangle}$ with ``velocity'' $\vert \dot \psi(\tau)\rangle$, i.e., by the metric induced by the inner product in Hilbert space. To systematically construct this path, it is helpful to decompose its tangent vector into two orthogonal components as $\lvert \dot\psi(t)\rangle=P(t)\lvert \dot\psi(t)\rangle+(1-P(t))\lvert \dot\psi(t)\rangle$. The first component is called the vertical part $\lvert \dot \psi \rangle_V$ and generates an evolution that stays within the projection $P(t)$, i.e., a mere change of phase of $\lvert \psi(t)\rangle$. The second part in contrast generates an evolution perpendicular to $P(t)$ in the sense of the inner product and is consequently called the horizontal part $\lvert \dot \psi \rangle_H$. From $\lVert \dot \psi\rVert^2=\lVert\dot \psi_V\rVert^2+\lVert \dot \psi_H\rVert^2$, it is easy to see that the tangent vector to the shortest path must be purely horizontal, i.e., \begin{align} P(t)\lvert \dot\psi(t)\rangle=0, \end{align} or, equivalently for non-degenerate ground states, $\langle \psi(t)\vert \dot \psi(t)\rangle=0$. This parallel-transport prescription immediately determines \begin{align} \phi(t)=i\int_{t_0}^t \text{d}\,\tau \langle{R(\tau)}\rvert \frac{d}{d\tau}\lvert{R(\tau)}\rangle. \end{align} The quantity $\mathcal A_{\frac{d}{dt}}(R(t))=\langle{R(t)}\rvert \frac{d}{dt}\lvert{R(t)}\rangle$ transforms under a gauge transformation $\lvert R\rangle\rightarrow \text{e}^{i\chi(R)}\lvert R\rangle$ as $\mathcal A_{\frac{d}{dt}}\rightarrow \mathcal A_{\frac{d}{dt}}+i\frac{d}{dt}\chi(R(t))$, i.e., as a $U(1)$ gauge field. The geometric phase $\phi^\gamma=i\int_\gamma \mathcal A ~(\text{mod } 2\pi)$ associated with a loop $\gamma$ in parameter space is gauge invariant. This construction can be immediately generalized to $n$-fold degenerate ground states where the projection $P(t)$ in Eq. () becomes \begin{align} P(t)=f(t) f^\dag(t) = \sum_{j=1}^n \lvert {R^j(t)}\rangle \langle {R^j(t)}\rvert \end{align} with the ground state manifold basis or frame $f = (\lvert {R^1}\rangle, ... , \lvert {R^n}\rangle)$. We note the independence of the projection under $U(n)$ basis transformations $f \to f U$. Thus, instead of a mere phase factor, the local gauge degree of freedom is then a $U(n)$ basis transformation on the ground state manifold. In this case, the geometric phase or, in mathematical terms, the $U(1)$ holonomy ${e}^{i\phi^\gamma}$ is replaced by the $U(n)$ holonomy \begin{align} U^\gamma=\mathcal T \text{e}^{-\int_\gamma \mathcal A}, \end{align} where $\mathcal A^{jl}=\langle R^j\rvert d\lvert R^l\rangle$ is the non-Abelian gauge field or connection and $\mathcal T$ is the time ordering operator along the cyclic path $\gamma$. We finally note that the connection can be expressed in a manifestly gauge invariant or basis independent form as $\mathcal A_K=-[(dP),P]$ . In this case, the geometric phase in the ground state manifold for a loop $\gamma$ starting at time $t_0$ is described by the propagator of the Schr\"odinger equation in the adiabatic limit, or the holonomy operator \begin{align} U^\gamma_K= P(t_0)\mathcal T \text{e}^{-\int_\gamma \mathcal A_K}P(t_0). \end{align} $U^\gamma$ as occurring in Eq. () is then just the gauge (basis) dependent representation matrix of $U^\gamma_K$. \\ \subsection{From geometric phases to topological band structures} Let us consider an insulating band structure in $d$ spatial dimensions with $n$ occupied Bloch states $\lvert u_k^\alpha\rangle$ below the Fermi energy that span the projection $P(k)=\sum_{\alpha=1}^{n}\lvert u_k^\alpha\rangle\langle u_k^\alpha\rvert$ at lattice momentum $k$. If we identify the energy gap relevant for the adiabatic approximation with the band gap of the insulator and the parameter manifold with the Brillouin zone (BZ) in which the lattice momentum is defined, a gauge structure can be defined in complete analogy to Section . The projection $P(k)$ defines an $n$-plane in the $N$-dimensional Hilbert space defined at every momentum $k$ in the BZ, where $N$ is the total number of bands (occupied and empty) of the model system. $P(k)$ is thus by definition a point on the Grassmann manifold $G_n(\mathbb C^N)=U(N)/(U(n)\times U(N-n))$. Topological invariants of this gauge structure then distinguish topologically inequivalent mappings $k\mapsto P(k)$, that is homotopically inequivalent mappings from the BZ torus $T^d$ to $G_n(\mathbb C^N)$. For the case of particle number conserving insulators without additional symmetries outlined here, an integer invariant named the $m$-th Chern number is defined for even spatial dimensions $d=2m$ while all insulators are equivalent in odd spatial dimensions. Physical symmetries imply constraints on the form of the mapping $k\mapsto P(k)$. For example the anti-unitary time reversal symmetry $T$ yields $P(k)=TP(-k)T^{-1}$. Mappings (band structures) that would be topologically equivalent when breaking the relevant symmetries may be topologically inequivalent when maintaining the symmetries. This phenomenon is commonly referred to as symmetry protection of a topological state. Taking also into account superconducting band structures as well as all symmetries considered by Altland and Zirnbauer significantly refines the system of topological invariants and leads to a pattern which has been coined the periodic table of topological insulators and superconductors . In this work, we would like to address the question to what extent such invariants can be generalized to lattice translation-invariant quantum many body systems in a mixed state. While the relevant physical symmetries can readily be generalized to the realm of density matrices, the definition of topological invariants for mixed states from a gauge structure over the space of density matrices has not been discussed so far. However, a topologically trivial gauge structure for density matrices has been discovered by Uhlmann as we will review now. Later on (see Section ), a more restrictive gauge structure that can be topologically non-trivial will be derived from Uhlmann's general construction. \subsection{Parallel transport and geometric phases for density matrices} A density matrix $\rho$ is a positive semi-definite operator with unit trace on a Hilbert space $\mathcal H$. Here, we consider $\text{dim} \mathcal H=N<\infty$. For the application to gapped band structures that we have in mind here, $N$ plays the role of the total number of bands of a model system. $\rho$ can be represented as a pure state $\lvert \psi\rangle$ on an extended Hilbert space $\mathcal H_A\otimes \mathcal H_B$ with $\mathcal H_A \simeq \mathcal H_B \simeq \mathcal H$ such that $\text{Tr}\left[O\otimes\mathbb{1}\lvert \psi\rangle\langle \psi\rvert\right]=\text{Tr}_A\left[O\rho\right]$ for any operator $O$ acting in $\mathcal H$, a prescription referred to as state purification. Equivalently, the purification can be represented in terms of a Hilbert Schmidt operator, i.e., a $N\times N$ matrix $w$ that satisfies \begin{align} \rho =\text{Tr}_B\left[\lvert \psi\rangle\langle \psi\rvert\right] = w w^\dag, \end{align} where the trace over the auxiliary Hilbert space $\mathcal H_B$ becomes a matrix multiplication. This description contains a redundancy since under $w\rightarrow w U$ with $U\in U(N)$, $\rho\rightarrow wUU^\dag w^\dag=\rho$ is unchanged. We note a formal analogy to the basis independence of the projection \eqref{eqn:proj}. Furthermore, the inner product $(w,v)=\text{Tr}[w^\dag v]$ for matrices again (cf. Section ) defines a natural means to measure the length $L=\int_{t_0}^{t}\text{d}\,\tau \sqrt{(\dot w,\dot w)}$ of a path $t\mapsto w(t)$ of Hilbert Schmidt operators that purify a path $t\mapsto \rho(t)$ in the sense of Eq. (). The shortest possible path can again be obtained by decomposing the tangent vector $\dot w$ into a vertical part $\dot w_V$ and a horizontal part $\dot w_H$ with the help of the inner product. $\dot w$ is purely horizontal for the shortest path. Vertical vectors can be defined in terms of Eq. () as tangent vectors to curves $s\mapsto w(s)$ that project to the same $\rho=ww^\dag$ for all $s$. Vertical vectors are hence of the form $\dot w_V=\left.\frac{d}{ds}wU(s)\right|_{s=0}=wu$ with $U(s)=\text{e}^{us}\in U(N)$. Horizontal vectors $\dot w_H$ are defined as orthogonal to all vertical vectors, i.e., $\text{Tr}[\dot w_H^\dag wu]=0$ for all $U(N)$-generators $u=-u^\dag$. Adding the hermitian conjugate to this condition yields $\text{Tr}[u(\dot w_H^\dag w-w^\dag \dot w_H)]=0$. This can be true for every antihermitian $u$ only if the second antihermitian factor is zero by itself, i.e., the tangent vector $\dot w$ is horizontal if , \begin{align*} \dot w^\dag w-w^\dag \dot w=0. \end{align*} Eq. (), repeated here for convenience, is the density matrix analog of Eq. (). Let us now construct the $U(N)$-gauge field associated with this parallel transport prescription, in a form that is amenable to practical calculations. For a given $\rho=\sum_j p_j \lvert j\rangle\langle j\rvert $, the uniquely defined square root $\sqrt{\rho}=\sum_j \sqrt{p_j}\lvert j\rangle\langle j\rvert$ defines a generic purification $w=\sqrt{\rho}$. In fact every purification can by means of a polar decomposition be written as \begin{align} w=\sqrt{\rho} U \text{ with } U\in U(N) \end{align} in some analogy to the polar decomposition of a complex number. Let us consider a loop $\gamma:t\mapsto \rho(t),~t\in[0,T]$ in the space of density matrices. We denote by $t\mapsto w(t)=\sqrt{\rho(t)}U(t)$ the parallel-transport of an arbitrary initial purification $w(0)=\sqrt{\rho(0)}U(0)$ along this loop $\gamma$ and call the geometric phase $H_U^\gamma=U(T)U(0)^\dag$ its Uhlmann holonomy. With $t_i=i T/M,~ M\in \mathbb N$, we can express $H_U^\gamma$ as \begin{align} H_U^\gamma=\lim_{M\rightarrow \infty}\prod_{i=1}^M U(t_i)U^\dag(t_{i-1}), \end{align} where the product is ordered from right to left with increasing $i$. We now wish to obtain the analog of Eq. \eqref{eqn:WilZee} in an explicit form. To this end, we first obtain the parallel-transported $U(t)$ in Eq. \eqref{eqn:wrhoU} at infinitesimally neighbouring points in time. Eq. () can be shown to be equivalent to \begin{align} w^\dag(t+\epsilon) w(t)=w^\dag(t)w(t+\epsilon) \ge 0 \end{align} to leading order in $\epsilon$; in particular, the hermitian matrix $w^\dag(t+\epsilon) w(t)$ is positive semi-definite. Defining $V=w^\dag(t+\epsilon)w(t)=U^\dag(t+\epsilon)\sqrt{\rho(t+\epsilon)}\sqrt{\rho(t)}U(t)\ge 0$ we immediately verify \begin{align*} \sqrt{\rho(t+\epsilon)}\sqrt{\rho(t)}=H U(t+\epsilon)U^\dag(t) \end{align*} with the semi-positive hermitian matrix $H=U(t+\epsilon)V(t)U^\dag(t+\epsilon)$. Hence, the singular value decomposition \begin{align*} \sqrt{\rho(t+\epsilon)}\sqrt{\rho(t)}=L DR^\dag \end{align*} with unitary $L,R$ yields $H=L D L^\dag$ and, more importantly , \begin{align} U(t+\epsilon)U^\dag(t)=L R^\dag. \end{align} With this recipe the connection $\mathcal A_U=-\dot U(t) U^\dag(t)=\lim_{\epsilon\rightarrow 0}\frac{1}{\epsilon}(1-U(t+\epsilon)U^\dag(t))$ can be computed explicitly. The desired holonomy then reads as \begin{align} H_U^\gamma = \mathcal T \text{e}^{-\int_\gamma \mathcal A_U}. \end{align} We note that $\mathcal A_U$ is invariant under a rescaling $\rho(t)\rightarrow \lambda(t) \rho(t)$ with a strictly positive real function $\lambda(t)>0$, which affects the singular values $D$ alone. Therefore, the normalization of the density matrix is not relevant for geometrical and topological considerations. \section{Topological aspects of mixed states} In order to discuss topological features of emergent gauge theories for mixed states, it is natural to view the above decomposition of tangent vectors into vertical and horizontal components as a connection on a principle fiber bundle (PFB) (see e.g., Ref. ). It is well known that the PFB of general Hilbert-Schmidt operators is topologically trivial since $\rho\mapsto\sqrt{\rho}$ is a global section (gauge). At first sight, this may seem discouraging. On the other hand, even the special case of pure states can be viewed as a topologically trivial gauge structure in terms of Hilbert Schmidt operators. In contrast, it is well known that the gauge structure for pure states discussed in Section can very well be topologically non-trivial, phenomena such as topological insulators being striking ramifications of this possibility. More concretely, the frame bundle of orthonormal frames $f_R$ spanning the projection $P(R)$ can be topologically non-trivial. However, it can always be viewed as a sub-bundle of the topologically trivial Hilbert-Schmidt bundle of arbitrary quadratic matrices $w(R)$ that satisfy $P(R)=w(R)w^\dag(R)$. In other words, the level of description determines whether topological aspects of pure states can be resolved. More generally speaking, every gauge theory (PFB) can be viewed as a sub-bundle of a topologically trivial gauge theory (PFB) with a bigger gauge degree of freedom but over the same base manifold. As for mixed state density matrices key questions are thus whether topologically non-trivial features exist and, if so, to identify a gauge structure which is able to reveal their topological content. Here we consider several physically motivated constraints under which a topologically non-trivial PFB of EOPS will be defined. \subsection{Triviality and structure of the Hilbert-Schmidt bundle} Let us first review the case of arbitrary invertible $N \times N$ density matrices $\rho$ as considered by Uhlmann , i.e., strictly positive matrices without any further constraints such as normalization of the trace. The space $M_N(N)$ of all Hilbert Schmidt operators $w$ purifying such density matrices via Eq. () is then simply given by the group of all invertible matrices $GL(\mathbb C,N)$. This is because $\rho=ww^\dag$ is strictly positive and has the non-vanishing determinant $\lvert \text{det} (w) \rvert^2$ if and only if $w\in GL(\mathbb C,N)$. $M_N(N)=GL(\mathbb C,N)$ is the total space of the Hilbert-Schmidt bundle which projects via $\Pi: w\mapsto ww^\dag$ onto the base manifold of strictly positive matrices $D_N(N)=GL(\mathbb C,N)/U(N)$. The quotient form of $D_N(N)$ is rooted in the fact that any regular matrix $M$ can be uniquely written as $M= \sqrt{MM^\dag}U, ~U\in U(N)$. For the same reason, $w$ with $ww^\dag=\rho$, can be uniquely represented as $w=\sqrt{\rho}U,~U\in U(N)$. Hence the fiber over $\rho$ is given by $G_\rho=\left\{\sqrt{\rho}U:U\in U(N)\right\}$, which is manifestly isomorphic to $U(N)$. $GL(\mathbb C,N)\overset{\Pi}{\rightarrow}D_N(N)$ thus defines a PFB. Due to the existence of the global gauge or section $\rho\rightarrow \sqrt{\rho}$, the PFB is topologically trivial, i.e., \begin{align} GL(\mathbb C,N)=GL(\mathbb C,N)/U(N)\times U(N). \end{align} \\ Now we consider singular density matrices, i.e., semi-positive $N\times N$ matrices with rank $n<N$. A naive analog of the regular case with fibers $G_\rho=\left\{\sqrt{\rho}U:U\in U(N)\right\}$ where $\rho$ now has rank $n$ does not give a PFB. The reason is that the right-action $w=\sqrt{\rho}U$ of $U(N)$ is no longer free, i.e., many $U\in U(N)$ give the same $w=\sqrt{\rho}U$ due to the $N-n$-dimensional null-space of $\sqrt{\rho}$. Simply restricting the $U$ to $U(n)$ acting on the support of $\sqrt{\rho}$ does not resolve this issue since this action would not be transitive, i.e., unable to reach all $w$ with $w w^\dag=\rho$. In Ref. , it has been shown that $M_N(n)\overset{\Pi}{\rightarrow}D_N(n),~n<N$ with the same projection $\Pi: w\mapsto ww^\dag$ still defines a PFB. However, by the same argument as before, namely by the existence of the global section $\rho \mapsto \sqrt{\rho}$, also these bundles are topologically trivial. \subsection{Ensemble of pure states gauge structure} Instead of considering general Hilbert Schmidt operators, we define a EOPS bundle in terms of $n$ non-orthonormalized pure states $\lvert \tilde \psi^\alpha\rangle$ that project onto a given $N\times N$ density matrix $\rho$ of rank $n$ as (Eq. () is repeated for convenience) \begin{align*} \rho=\sum_{\alpha=1}^{n} \lvert \tilde \psi^\alpha\rangle\langle \tilde \psi^\alpha\rvert \end{align*} as a gauge structure over the space of density matrices $\rho$. Instead of viewing $\rho$ as a single pure state in a larger Hilbert space, EOPS in this scheme are plausible ensembles of pure states in the system Hilbert space. A generic choice $\lvert \tilde \psi_\alpha \rangle = \sqrt{p_\alpha} \lvert \psi_\alpha\rangle$ along these lines is obtained from the spectral representation $\rho=\sum_\alpha p_\alpha \vert \psi_\alpha\rangle\langle \psi_\alpha\rvert$. Under arbitrary unitary rotations \begin{align} \lvert\tilde \varphi_\alpha\rangle= \lvert \tilde\psi_{\beta}\rangle U_{\beta\alpha}, \end{align} the density matrix $\rho$ obtained from the projection () is unchanged. However, the pure states $\lvert \tilde \varphi_\alpha\rangle$ are no longer mutually orthogonal. The matrix representation \begin{align} w=\left(\lvert \tilde \psi_1\rangle,\ldots,\lvert \tilde \psi_n\rangle,0,\ldots,0\right) \end{align} with $N-n$ zero columns defines a natural embedding of the space $P_N(n)$ of non-orthonormal $n$-frames into the space $M_N(n)$ of Hilbert Schmidt operators of rank $n$ . $U$ as occurring in Eq. () defines the local $U(n)$ gauge degree of freedom of the PFB $P_N(n)\overset{\Pi}{\rightarrow}D_N(n)$. For regular density matrices ($n=N$), it is easy to see that the space of all EOPS is identical to the space of all Hilbert Schmidt operators satisfying $\rho=ww^\dag$, i.e., $P_N(N)=M_N(N)$. Thus, for invertible density matrices with no further spectral constraints, the EOPS scheme is equivalent to the purification by pure state representation in a larger Hilbert space. From this observation, we immediately conclude that the EOPS bundle is topologically trivial for unconstrained regular density matrices. In the following, we will consider several constraints under which the EOPS bundle can become topologically non-trivial and explicitly construct the corresponding topological invariants. \subsection{Gauge structure of spectrally constrained density matrices} {\emph{Non-degenerate density matrices --}} The situation becomes more interesting if we impose certain conditions on the spectrum of the density matrix. Let us first consider a regular {\emph{non-degenerate}} density matrix $\rho=\sum_\alpha p_\alpha \lvert \psi_\alpha\rangle \langle \psi_\alpha\rvert$, where we can now without loss of generality assume $p_1>p_2>\ldots>p_N$. For the hermitian operator $\rho$ with non-denerate eigenvalues, the projections $P_\alpha =\lvert \psi_\alpha\rangle \langle \psi_\alpha\rvert$ are mutually orthogonal and it is natural to order the columns $\sqrt{p_\alpha} \lvert\psi_\alpha\rangle$ of a Hilbert Schmidt representation () with descending size of the spectral weights $p_\alpha$. This ordering will only be maintained under a subgroup of gauge transformations () that are a direct sum of $N$ $U(1)$-transformations $\text{e}^{i\phi_\alpha}$ acting on the rays of eigenstates associated with the eigenvalues $p_\alpha$. Eq. () then simplifies to \begin{align} \lvert \tilde \varphi_\alpha\rangle=\text{e}^{i\phi_\alpha}\lvert \tilde\psi_\alpha\rangle,\quad\alpha=1,\ldots,N. \end{align} The EOPS bundle thus constrained consists of $N$ $U(1)$-bundles $P_N(1)\overset{\Pi}{\rightarrow}D_N(1)$ which are subject to the constraint \begin{align} \sum_{\alpha=1}^{N}\lvert \psi_\alpha\rangle\langle \psi_\alpha\rvert=\mathbb 1. \end{align} While the individual $U(1)$-bundles can be topologically non-trivial, Eq. () enforces a zero sum rule for their topological invariants. The reason for this is analogous to the pure state case and can be intuitively understood as follows. The individual subspaces may exhibit a topologically non-trivial winding as a function of the control parameters (e.g. lattice momentum) in the total Hilbert space. This total space serves as fixed reference and thus by definition does not exhibit any "net-winding". Hence, the parameter dependences of the individual subspaces have to compensate each other in order to span the resolution of the identity () of the total embedding space at every point in parameter space. The spectral constraint forbidding any level degeneracy can be somewhat relaxed by assuming that $m$ subsets of cardinality $n_1,\ldots,n_m$ with $N=\sum_{j=1}^{m}n_j$ of levels are allowed to be degenerate, but still have distinct eigenvalues from levels outside of their subset. Such constraints simply result in a restricted gauge group of the form $U(n_1)\times U(n_2)\times \ldots \times U(n_m)$. Again, the individual $U(n_j)$-structures may be topologically non-trivial but obey a zero sum rule due to Eq. (). {\emph{Purity gaps --}} The notion of a purity gap can be rationalized in this framework. To this end, we consider a spectral constraint along the lines of the above discussion where the $n$ largest eigenvalues of $\rho$ belong to one subset but are by assumption not degenerate with the $(n+1)$-th largest eigenvalue of $\rho$. Under these circumstances we can uniquely define a pure state projection $P=\sum_{\alpha=1}^{n}\lvert \psi_\alpha\rangle\langle \psi_\alpha\rvert$ which is closest to the mixed state defined by $\rho$. In Refs. , the topological invariant of $\rho$ has been defined as the invariant of the corresponding pure state projection $P$. Below, we directly define topological invariants for mixed states in terms of the Uhlmann connection () restricted to the subspace associated with the $n$ largest eigenvalues of $\rho$. Their value is equal to the one obtained by performing an adiabatic deformation into a pure state along the lines of Refs. but, quite remarkably, no reference to pure states is required here. A purity gap closing means that the $n$-th largest eigenvalue and the $(n+1)$-th largest eigenvalue of $\rho$ become degenerate, a singular-point at which the spectral constraint is violated and the topological invariant associated with the subspace of the $n$ largest eigenvalues is not well defined. As a consequence, this subspace can exchange topological charge with the eigenspace associated to the $(n+1)$-th largest eigenvalue and the topological invariant for the subspace of the $n$ largest eigenvalues may have changed when the purity gap reopens -- a topological phase transition by purity gap closing.\\ {\emph{Singular density matrices --}} A physical system under non-equilibrium conditions may not be ergodic and can thus be described by a singular density matrix, i.e., a density matrix of rank $n<N$. This constraint can be viewed as special case of the above discussion of spectral constraints: A subset of $n$ eigenvalues $p_\alpha$ is non-zero and hence not degenerate with the complement of $N-n$ non-zero eigenvalues. The $U(n)$ EOPS bundle associated with the non-zero subspace can be topologically non-trivial. \subsection{Topological invariants} The gauge structure of spectrally constrained density matrices defined in Section allows for the definition of topological invariants directly in terms of the Uhlmann connection (see Section ), i.e., without reference to pure states. To this end, we consider a subset of $n<N$ eigenvalues (eigenvectors) labeled by $j=1,\ldots,n$ that are by assumption non-degenerate with (orthogonal to) the $N-n$ remaining eigenvalues (eigenvectors) of $\rho$. We denote by $\mathcal P=\sum_{j=1}^n \lvert \psi_j\rangle\langle \psi_j \rvert$ the projection onto the rank $n$ density matrix $\hat \rho=\sum_j \lvert \tilde \psi_j\rangle\langle \tilde \psi_j \rvert$, i.e., $\hat \rho=\mathcal P \rho \mathcal P$. Eq. () shows how the corresponding $U(n)$ EOPS bundle can be embedded into the general PFB of Hilbert Schmidt operators considered by Uhlmann. This allows us to construct a curvature on the constrained $U(n)$ EOPS bundle which is directly inherited from Uhlmann's general construction. By general definition, a curvature $\mathcal F_{\mu\nu}$ is the geometric phase per area associated with the parallel transport around and infinitesimal parallelogram $\gamma_{\mu\nu}$ spanned by the vectors $\delta_\mu \hat e_\mu,\delta_\nu \hat e_\nu$ in momentum space (or more generally in parameter space). To obtain the curvature of the constrained EOPS bundle, we have to project Uhlmann's geometric phase $H_U^{\gamma_{\mu\nu}}$ (see Eq. ()) associated with the infinitesimal loop $\gamma_{\mu\nu}$ for $\hat \rho$ onto the $n$-dimensional subspace of interest by virtue of $\mathcal P$. We denote by $\hat H_U^{\gamma_{\mu\nu}}$ the representation matrix of the geometric phase in this subspace, i.e., $\left(\hat H_U^{\gamma_{\mu\nu}}\right)_{ij}=\langle \psi_i\rvert H_U^{\gamma_{\mu\nu}} \lvert \psi_j\rangle$. With these definitions, the curvature of the $U(n)$ EOPS bundle is given by \begin{align} \hat{\mathcal F}_{\mu\nu}=\lim_{\delta\rightarrow 0}\frac{1-\hat H_U^{\gamma_{\mu\nu}}}{\delta_\mu\delta_\nu}, \end{align} where $\delta \rightarrow 0$ is shorthand for $(\delta_\mu,\delta_\nu)\rightarrow (0,0)$. In particular, the gauge invariant Abelian curvature reads as \begin{align} \hat{\mathcal F}_{\mu\nu}^A= \text {Tr} \hat{\mathcal F}_{\mu \nu}=-\lim_{\delta\rightarrow 0}\frac{\log\det(\hat H_U^{\gamma_{\mu\nu}})}{\delta_\mu\delta_\nu}, \end{align} where $\log\det(\hat H_U^{\gamma_{\mu\nu}})$ is $i$ times the argument of the Abelian geometric phase factor. In terms of this curvature, standard topological invariants for mixed states with spectral constraints can be defined. For example, if the constrained rank $n$ density matrix $\hat \rho(k)$ is parameterized by a lattice momentum $k$ in a two-dimensional Brillouin zone (BZ), the first Chern number is given by \begin{align} \mathcal C=\int_{\text{BZ}}\frac{i\hat{\mathcal F}^A}{2\pi}=\int_{\text{BZ}}\text{d}^2k\,\frac{i\hat{\mathcal F}^A_{xy}}{2\pi}. \end{align} The generalization to other invariants and the implementation of physical symmetries is straight forward and analogous to the pure state case. As a prominent example, we would like to mention the case of time reversal symmetry as a protecting symmetry (see Section for a discussion of the pure state analog). Instead of the projection onto the occupied bands, the density matrix $\rho(k)$ itself then obeys the symmetry $\mathcal T\rho(k)\mathcal T^{-1}=\rho(-k)$ with the anti-unitary time reversal operator $\mathcal T$. As in the pure state case, this symmetry carries over to arbitrary Uhlmann-holonomies and allows for the definition of topological invariants in analogy to Ref. . \section{Uhlmann phase winding numbers} In two recent back-to-back publications , a complementary approach towards the definition of topological invariants for mixed states has been reported, in particular for Chern insulators in thermal states. Their approach is based on the so called Uhlmann phase $\phi_U^\gamma$ associated with a loop $\gamma$ in momentum space, defined as \begin{align} \text{e}^{i\phi_U^\gamma}&=\text{Tr}\left[w(0)^\dag w(T)\right]=\text{Tr}\left[\rho(0)U(T)U(0)^\dag\right]\nonumber\\&=\text{Tr}\left[\rho(0)H_U^\gamma\right], \end{align} where $\gamma$ is traversed between $t=0$ and $t=T$, $w(t)=\sqrt{\rho(t)}U(t)$ is a parallel transport in the sense of Uhlmann's connection () and the Uhlmann holonomy $H_U^\gamma$ has been defined in Eq. (). By construction, $\phi_U^\gamma$ bears some analogy with the Abelian Berry phase of a pure state defined as $\text{e}^{i\phi^\gamma}=\langle \psi(0)\vert \psi(T)\rangle$ (cf. Section ). However, there is one crucial difference in the mathematical structure of $\phi^\gamma$ and $\phi_U^\gamma$: While $\text{e}^{i \phi^\gamma}$ is a $U(1)$-holonomy as is clear from the explicit representation in terms of an integral over a $U(1)$-gauge field (see Eq. ()), $\text{e}^{i\phi_U^\gamma}$ does in general {\emph{not}} have this property -- in contrast to the full non-Abelian Uhlmann holonomy $H_U^\gamma$. This observation is not just a minor technical point but can have drastic consequences. The key feature of a $U(1)$-holonomy is its additive group structure, i.e., $\phi^{\gamma_{12}}=\phi^{\gamma_{1}} + \phi^{\gamma_2}~(\text{mod} 2\pi)$, where $\gamma_{12}$ denotes the concatenation of the loops $\gamma_1$ and $\gamma_2$. It is this very property that allows us to go from an infinitesimal geometric phase as measured locally by a curvature (see Eq. ()) to a quantized global topological invariant represented as an integral over the entire parameter space (see, e.g., the definition of the Chern number in Eq. ()). Again making use of this additive group structure of the ordinary Berry phase, the Chern number $\mathcal C$ can be represented as the change of the Berry phase $\phi^{k_x}=\int_{S^1}\text{d}k_y\,\mathcal A_{y}(k_x,k_y)$ associated with a $k_y$-circle in the BZ at fixed $k_x$. This reformulation is achieved by dividing the BZ torus into $k_y$-rings $R_{k_x}$ of infinitesimal width $[k_x,k_x+dk_x]$ around $k_x$ and then using the Stokes theorem to transform the area integral of the curvature over these rings into a line integral of the connection $\mathcal A$ along their boundaries, i.e., $\int_{R_{k_x}}\text{d}^2k \mathcal F_{xy}=\int_{S^1}\text{d}k_y\left[\mathcal A_y(k_x+dk_x,k_y)-\mathcal A_y(k_x,k_y)\right]=\left(\frac{\partial \phi^{k_x}}{\partial k_x}\right)dk_x$. Explicitly, this procedure leads to the following well known representation of the Chern number \begin{align} \mathcal C = \frac{1}{2\pi}\int_{S^1} \left(\frac{\partial \phi^{k_x}}{\partial k_x}\right)\, \text{d} k_x, \end{align} where the circle $S^1$ is around the $k_x$ loop of the BZ. Equivalently, the role of $k_x$ and $k_y$ could be exchanged here, of course. This is again because of the holonomy group structure of the Berry phase which allows us to divide the BZ into stripes in an arbitrary direction. The basic idea of Refs. is to take the right hand side of Eq. () but to replace the ordinary Berry phase by the Uhlmann phase as defined in Eq. (), i.e., \begin{align} \mathcal C_U = \frac{1}{2\pi}\int_{S^1} \left(\frac{\partial \phi^{k_x}_U}{\partial k_x}\right)\, \text{d} k_x. \end{align} We note that Ref. actually discusses several constructions in terms of the spectrum of the ``holonomy matrix" $\rho(0)H_U^\gamma$. Since the absence of a holonomy group structure which is at the heart of our present discussion also pertains to $\rho(0)H_U^\gamma$, this distinction is not of central importance here as will be addressed more explicitly in our example below. As both $k_x$ in the BZ and the phase $\phi_U^{k_x}$ are defined on a circle, the mapping $k_x \mapsto \phi_U^{k_x}$ is characterized by an integer quantized winding number which is exactly measured by Eq. (). However, since the Uhlmann phase does not have an additive group structure, $\mathcal C_U$ cannot be represented as the integral of a curvature over the 2D BZ. It is hence not immediately clear to what extent $\mathcal C_U$, technically being a 1D winding number, can be seen as a unique property of the 2D system under investigation. In particular, it is not clear that $\mathcal C_U$ as defined in Eq. () is equal to $\tilde{\mathcal C}_U$ obtained from $\mathcal C_U$ by exchanging $k_x$ and $k_y$ in all calculations, i.e. \begin{align} \tilde{\mathcal C}_U=\frac{1}{2\pi}\int_{S^1} \left(\frac{\partial \phi^{k_y}_U}{\partial k_y}\right)\, \text{d} k_y. \end{align} Here, we demonstrate that indeed $\mathcal C_U\ne \tilde{\mathcal C}_U$ can occur for a Chern insulator the Hamiltonian of which is not invariant under the exchange of $k_x$ and $k_y$. To this end, we consider a two-banded fermionic Chern insulator on a 2D square lattice with unit lattice constant defined by the Bloch Hamiltonian \begin{align} &H(k)=\vec d(k)\cdot \vec \sigma=\sum_{j=1}^3 d^j(k)\sigma_j,\nonumber\\ &d^1(k)=\sin(k_x),~d^2(k)=3\sin(k_y),\\ &d^3(k)=1-\cos(k_x)-\cos(k_y)\nonumber, \end{align} where $\sigma_j$ are Pauli matrices. Note the anisotropy factor of $3$ in $d^2(k)$ which is crucial here. For the {\emph{two-banded}} model (), the Uhlmann connection and the Uhlmann phase can be readily calculated along the lines of Ref. . Explicitly, the Uhlmann connection $\mathcal A_U$ in this special case reads as \begin{align} \mathcal A_{U,\mu} = -\left[(\partial_{k_\mu}\sqrt{\rho(k)}),\sqrt{\rho(k)}\right]. \end{align} At $T=0$, Eq. () concurs with the Kato connection (see Eq. ()) for the pure state case and the ordinary Chern number () reads as \begin{align*} \mathcal C=\frac{1}{4\pi}\int_{\text{BZ}}\text{d}^2k\,\left (\hat d(k)\cdot [(\partial_{k_x} \hat d(k))\times(\partial_{k_y} \hat d(k))]\right) \end{align*} with $\hat d=\frac{\vec d}{\lvert \vec d\rvert}$. Explicit calculation yields $\mathcal C=-1$. Also, both $\mathcal C_U$ and $\tilde{\mathcal C}_U$ trivially concur with $\mathcal C$ since the Uhlmann phase reduces to the Berry phase for pure states such that Eq. (), Eq. (), and Eq. () become equivalent. At finite $T$, the system is in a thermal state defined by $\rho(k)=\frac{1}{Z}\text{e}^{-\beta H(k)}$ with $Z=\text{Tr}[\text{e}^{-\beta H(k)}]$. Using the simplified form () of the Uhlmann connection, the direct calculation of Uhlmann holonomies () for arbitrary loops in momentum space is straightforward which in turn allows the direct calculation of $\mathcal C_U$ and $\tilde{\mathcal C}_U$ from Eq. () and Eq. (), respectively. In the infinite temperature limit, this calculation becomes trivial as $\rho(k)=\mathbb 1$ independent of $k$ and hence $\mathcal A_U=0$ in Eq. (), resulting in $\mathcal C_U=\tilde{\mathcal C}_U=0$. Hence, the Uhlmann phase winding numbers have to jump from $-1$ to $0$ at some temperature. Numerically performing this calculation for finite temperature, we find that both $\mathcal C_U$ and $\tilde{\mathcal C}_U$ jump to zero at finite critical temperatures $\beta_c =\frac{1}{T_c}$ and $\tilde \beta_c =\frac{1}{\tilde T_c}$, respectively. Remarkably, for the present model, $\beta_c\ne \tilde \beta_c$, i.e., $\mathcal C_U$ and $\tilde{\mathcal C}_U$ do not concur at all temperatures. Numerically, we find $\beta_c=0.874$ and $\tilde \beta_c=1.32$. This discrepancy is illustrated In Fig. , where we visualize the winding numbers $\mathcal C_U$ (left panel) and $\tilde{\mathcal C}_U$ (right panel) for $\beta=\frac{1}{T}=1.3$, i.e., $\beta_c<\beta<\tilde \beta_c$. Clearly, $\mathcal C_U=1$ since $\phi_U^{k_x}$ monotonously increases, interrupted by a jump from $\pi$ to $-\pi$ at $k=0$, by $2\pi$ as $k_x$ completes the loop from $k_x=-\pi$ to $k_x=\pi$. In contrast, $\phi_U^{k_y}$ does not reach all values in $[-\pi,\pi]$ which implies $\tilde{ \mathcal C}_U=0$. On a more detailed note, we would like to point out that, for the parameters chosen here, $\mathcal C_U$ is in the well defined regime in the sense of Ref. . This is because the ``holonomy matrix" $\rho(k_x,k_{y0})H_U^{\gamma_{k_x,k_{y0}}}$ of a $k_y$-loop with footpoint $k_{y0}$ at fixed $k_x$ is gapped for all $k_x,k_{y0}$, i.e., the difference of the absolute values of its eigenvalues is finite and bounded from below by $\Delta= 0.268$. Comparing the construction in Refs. to our present analysis in Section , we would like to emphasize that the Chern number defined through Eq. () and Eq. () does not exhibit similar finite temperature transitions as $\mathcal C_U$ and $\tilde{\mathcal C}_U$ but can, for the model (), be uniquely defined as $-1$ for the larger eigenvalue of the density matrix for all finite temperatures. At infinite temperature, in contrast, the density matrix does not obey any spectral constraints in the sense discussed in Section , rendering all topological invariants trivial.\\ \section{Concluding remarks} We have discussed how several assumptions regarding the spectrum of a family of density matrices can lead to a topologically non-trivial gauge structure. In this framework topological invariants that are protected by these spectral assumptions have been defined for mixed states. Protected here means that topologically inequivalent mappings from a parameter space into the density matrices can be continuously deformed into each other only if the underlying spectral assumptions are violated. Identification of the parameter space with the Brillouin zone of a lattice translation invariant system provides one way of generalizing topological band structure invariants to the realm of mixed states. A non-trivial example where going beyond pure states is crucial to obtain a topologically non-trivial Chern insulator as a steady state in the framework of a non-equilibrium open quantum system dynamics has been reported in Ref. . Additional physical symmetries refining this system of topological invariants can be considered in analogy to the pure state case. The topological invariants defined here, being gauge invariant properties of the density matrix, are in principle experimentally accessible via state tomography. However, their relation to natural observables such as response functions is not yet conclusively understood. First progress along these lines has been reported in Ref. . As for two-banded Chern insulators such as the toy model in Eq. (), any statistical mixture of bands with opposite Chern number will certainly cause deviations from the quantized Hall conductance. In other situations where the physical ramification of the topological invariant is related to a half-integer quantized polarization, a statistical mixture may still exhibit a quantization.\\ \section{Acknowledgements} {Support from the ERC \je{Synergy Grant} UQUAM, and the START Grant No. Y 581-N16 is gratefully acknowledged.} \bibliographystyle{apsrev} \begin{thebibliography}{7} \bibitem{Berry} M.\ V.\ Berry, Proc. R. Soc. London A {\bf{392}}, 45 (1984). \bibitem{Simon} B.\ Simon, Phys. Rev. Lett. {\bf{51}}, 2167 (1983). \bibitem{WilZee} F.\ Wilczek, and A. Zee, Phys. Rev. Lett. {\bf{52}}, 2111 (1984). \bibitem[{\ {Klitzing et~al.}(1980)\ {Klitzing, Dorda, and Pepper}}]{Klitzing1980} \bibinfo{author}{\ {K.~v.} \ {Klitzing}}, \bibinfo{author}{\ {G.}~\ {Dorda}}, \ {and} \bibinfo{author}{\ {M.}~\ {Pepper}}, \bibinfo{journal}{Phys. Rev. Lett.} \textbf{\bibinfo{volume}{45}}, \bibinfo{pages}{494} (\bibinfo{year}{1980}). \bibitem[{\ {Laughlin}(1981)}]{Laughlin1981} \bibinfo{author}{\ {R.~B.} \ {Laughlin}}, \bibinfo{journal}{Phys. Rev. B} \textbf{\bibinfo{volume}{23}}, \bibinfo{pages}{5632} (\bibinfo{year}{1981}). \bibitem{TKNN1982} D.\ J.\ Thouless, M.\ Kohmoto, M.\ P.\ Nightingale, and M.\ den Nijs, Phys.\ Rev.\ Lett.\ {\bf 49}, 405 (1982). \bibitem[{ {Kohmoto}(1985)}]{Kohmoto1985} \bibinfo{author}{ {M.}~ {Kohmoto}}, \bibinfo{journal}{Annals of Physics} \textbf{\bibinfo{volume}{160}}, \bibinfo{pages}{343} (\bibinfo{year}{1985}). \bibitem{Schnyder2008} A.\ P.\ Schnyder, S.\ Ryu, A.\ Furusaki, and A.\ W.\ W.\ Ludwig, Phys.\ Rev.\ B {\bf 78}, 195125 (2008). \bibitem{Kitaev2009} A.\ Kitaev, AIP Conference Proceedings {\bf 1134}, 22 (2009). \bibitem{Ryu2010} S.\ Ryu, A.\ P.\ Schnyder, A.\ Furusaki, and A.\ W.\ W.\ Ludwig, New J.\ Phys.\ {\bf 12}, 065010 (2010). \bibitem{HasanKane} M.\ Z.\ Hasan and C.\ L.\ Kane, Rev.\ Mod.\ Phys.\ {\bf 82}, 3045 (2010). \bibitem{XLReview} X.-L.\ Qi and S.-C.\ Zhang, Rev.\ Mod.\ Phys.\ {\bf 83}, 1057 (2011). \bibitem{Kato1950} T.\ Kato, J. Phys. Soc. Japan \emph{5}, 435 (1950). \bibitem{Altland} {A. Altland and M.\ R.\ Zirnbauer, Phys.\ Rev.\ B {\bf 55}, 1142 (1997).} \bibitem{KennedyZirnbauer} R.\ Kennedy, M.\ R.\ Zirnbauer, arXiv:1409.2537 (2014). \bibitem{Uhlmann1986} A. Uhlmann, Rep. Math. Phys. {\bf{24}}, 229 (1986). \bibitem{Uhlmann1991} A. Uhlmann, Lett. Math. Phys. {\bf{21}}, 229 (1991). \bibitem{Uhlmann1993} A. Uhlmann, Rep. Math. Phys. {\bf{33}}, 253 (1993). \bibitem{note2} The mathematical background of this statement is that for every smooth vector bundle, there is a ``complement" such that the direct sum of the two gives a trivial bundle over the same base manifold. \bibitem{EntanglementFormation} C.\ H.\ Bennett, D.\ P.\ DiVincenzo, J.\ A.\ Smolin, and W.\ K.\ Wootters, Phys.\ Rev.\ A {\bf{54}}, 3824 (1996). \bibitem{DiehlTopDiss} S.~Diehl, E.~Rico, M.~A. Baranov, and P.~Zoller, Nature Physics {\bf{7}}, 971 (2011). \bibitem{BardynTopDiss} C.-E.\ Bardyn, M.\ A.\ Baranov, C.\ V.\ Kraus, E.\ Rico, A.\ Imamoglu, P.\ Zoller, S.\ Diehl, New J. Phys. {\bf 15}, 085001 (2013). \bibitem{Delgado1D} O.\ Viyuela, A.\ Rivas, M.\ A.\ Martin-Delgado, Phys.\,Rev.\, Lett.\, {\bf{112}}, 130401 (2014). \bibitem{Arovas2014} Z.\ Huang, D.\ P.\ Arovas, Phys.\,Rev.\, Lett.\, {\bf{113}}, 076407 (2014). \bibitem{Delgado2014} O.\ Viyuela, A.\ Rivas, M.\ A.\ Martin-Delgado, Phys.\,Rev.\, Lett.\, {\bf{113}}, 076408 (2014). \bibitem{QAH} \bibinfo{author}{{F.~D.~M. Haldane}}, \bibinfo{journal}{Phys. Rev. Lett.} \textbf{\bibinfo{volume}{61}}, \bibinfo{pages}{2015} (\bibinfo{year}{1988}). \bibitem{note1} $R(t)$ may be thought of as external electric or magnetic fields for example. A finite spacing between the energy levels at all times is assumed to define the adiabatic time scale. \bibitem{AAphase} Y.\ Aharonov, J.\ Anandan, Phys.\ Rev.\ Lett.\ {\bf {58}}, 1593 (1987). \bibitem{Pphase} S.\ Pancharatnam, Proc. Indian Acad. Sci. A {\bf{44}}, 247 (1956). \bibitem{PphasePRL} J.\ Samuel and R.\ Bhandari, Phys.\ Rev.\ Lett.\ {\bf{60}}, 2339 (1988). \bibitem{Chern1946} S.\ S.\ Chern, The Annals of Mathematics, {\bf 47} (1), 85 (1946). \bibitem{Nakahara} M.\ Nakahara, \emph{Geometry, Topology and Physics}, CRC Press (2003). \bibitem{DittmannRudolph1992} J.\ Dittmann, G.\ Rudolph, Journal of Geometry and Physics {\bf{10}}, 93 (1992). \bibitem{Prodan2011} E.\ Prodan, Phys.\ Rev.\ B {\bf{83}}, 235115 (2011). \bibitem{Huebner1993} M.\ H\"ubner, Phys. Lett. A {\bf{179}}, 226 (1993). \bibitem{DissCI} J.\ C.\ Budich, P.\ Zoller, S.\ Diehl, Phys. Rev. A {\bf{91}}, 042117 (2015). \bibitem{Avron2012} J.\, E.\, Avron, M.\, Fraas, G.\, M.\, Graf, Journal of Statistical Physics, {\bf{148(5)}}, 800 (2012). \end{thebibliography} |
1501.04136 | Title: Dynamical Localization in Molecular Alignment of Kicked Quantum Rotors
Abstract: The periodically $\delta$-kicked quantum linear rotor is known to experience
non-classical bounded energy growth due to quantum dynamical localization in
angular momentum space. We study the effect of random deviations of the kick
period in simulations and experiments. This breaks the energy and angular
momentum localization and increases the rotational alignment, which is the
analog of the onset of Anderson localization in 1-D chains.
Body: \title{Dynamical Localization in Molecular Alignment of Kicked Quantum Rotors} \author{A. Kamalov$^{1,2\: \ddagger}$} \author{D.W. Broege$^{1,3}$} \author{P.H. Bucksbaum$^{1,2,3}$} \affiliation{$^{1}$Stanford PULSE Institute, SLAC National Accelerator Laboratory, 2575 Sand Hill Rd., Menlo Park, CA 94025, USA} \affiliation{$^{2}$Department of Physics, Stanford University, 450 Serra Mall, Stanford, CA, 94305, USA} \affiliation{$^{3}$Department of Applied Physics, Stanford University, 450 Serra Mall, Stanford, CA, 94305, USA} \begin{abstract} The periodically $\delta$-kicked quantum linear rotor is known to experience non-classical bounded energy growth due to quantum dynamical localization in angular momentum space. We study the effect of random deviations of the kick period in simulations and experiments. This breaks the energy and angular momentum localization and increases the rotational alignment, which is the analog of the onset of Anderson localization in 1-D chains. \begin{description} \item[PACS numbers] \verb+33.80.-b,+ \verb+05.45.Mt,+ \verb+72.15.Rn,+ \verb+64.60.Cn+ \end{description} \end{abstract} \maketitle A consequence of quantum mechanics is a set of localization mechanisms that are inexplicable in a classical picture without state quantization. Earlier studies of the periodic $\delta$-kicked rotor found strong contrast between the energy growths of the classical and quantum rotors with subsequent kicks: The classical rotor exhibited unbounded energy growth, but the quantum rotor was shown to have bounded growth despite no limitation in the excitation bandwidth. This result was explained by showing the periodically kicked rotor to be mathematically analogous to Anderson localization in the 1-D tight binding model , leading to localization within the quantum rotor's angular momentum state space . The localization mechanism in the quantum rotor, popularly referred to as dynamical localization, is of significant interest in the quantum chaos community but has never before been experimentally studied in a true quantum linear rotor. Dynamical localization has been studied within the context of an approximately analogous system consisting of ultracold atoms subject to periodically applied potentials created by standing waves . The ultracold atom approach culminated in multiple demonstration of dynamical localization . A previous experimental study using true quantum rotors by Zhdanovich et all has utilized a periodic pulse train to study energy transitions across a limited number of rotation states. Zhdanovich's work demonstrated isotope selectivity which relied on localization within the unselected isotope's angular momentum states, but did not itself show localization. This report is the first study that includes experimental evidence of the dynamical localization mechanism within the true quantum rotor. We investigate the impact of deviation from periodic kicking on the localization of angular momentum population distributions of quantum rotors. Simulations of 64 pulse trains show that the introduction of non-uniform separation, or spacing disorder, to the pulse time separations causes the rotor to become more oriented along the $z$-axis, the axis of kicking. This corresponds to increased angular momentum in the $x-y$ plane and the delocalization of the rotor's angular momentum state population distribution. Experiments agree with our simulations. Our findings display the dynamical localization process expected in periodically kicked quantum linear rotors. The alignment of the quantum linear rotor $^{14}$N is studied when it is subject to multiple laser-induced impulsive kicks of the sample. Molecular alignment is the spatial anisotropy of the molecule's angular orientation . Substantial literature has been published on the alignment of linear molecules after the application of a multi-pulse trains with uniform pulse separation , but there has been no experimental investigation of linear molecules using pulse sequences with non-uniform time separation. We perform a numerical quantum propagation simulation to study the molecular alignment of $^{14}$N using pulse trains consisting of 64 pulses. A split-step operator method is performed to calculate an effective Hamiltonian in the interaction picture. The effective Hamiltonian acts upon independent density matrices representing states with different directional angular momentum, $m$. Step sizes of 1 fs or 10 fs are used depending on whether the electric field is present or not. The pulse temporal envelope is approximated using a cosine squared pulse profile. We use the rotational constant $B = 1.998 cm^{-1}$ and centrifugal distortion constant $D = 5.737 \cdot 10^{-6} cm^{-1}$ to simulate the quantum rotor. We consider multiple extended pulse trains with varying spacing disorder in the temporal spacings and plot the results in fig. . We display both population alignment $<cos^{2} \theta >_t$, which is the time averaged value of field-free molecular alignment; and the average rotational energy of the sample, $<J^{2} / 2I>$, as functions of the number of pulses applied. The population alignment is an experimentally accessible observable, while the energy is the traditionally analyzed quantity in studies focused on dynamical localization. The results indicate that the two quantities are correlated. A discussion of the relationship between population alignment and energy is presented later in this report. \begin{figure*}[t] \begin{tabular}{@{}c@{}} \\ \end{tabular} \caption{Simulated population alignment and energy for extended pulse trains with different levels of spacing disorder. The train with no disorder is a pulse train with uniform $\tau = 1.159 T_{rev} = \frac{2\pi + 1}{2\pi}T_{rev}$. The periodic disorder train consists of a repeating series with three spacings of $\tau = 1.159T_{rev}$ and a fourth spacing of $\tau = 1.183T_{rev}$. The aperiodic disorder train is similar to the periodic sequence, except every fourth spacing is a randomly generated value within the range $\tau = [1.159T_{rev}, 1.207T_{rev}]$. The full disordered train uses random values within the range $\tau = [1.159T_{rev}, 1.207T_{rev}]$ for all spacings. The train that excites $^{14}$N with kicks on revival is periodic with all $\tau = T_{rev}$; it is plotted here for reference only. The rotational energy of $^{14}$N is plotted for comparison to theoretical studies of dynamical localization. The relationship between these two values is not one-to-one but shows matching behaviors, especially in terms of boundedness. The color maps indicate the J-state population distribution as a function of kick count for each of the applied pulse trains. The peak intensity used for this simulation is 1.2 $\cdot 10^{14} \frac{W}{cm^{2}}$. $T_{rev}$ is the well known $^{14}$N alignment revival time of 8.38 ps.} \end{figure*} The result in fig. shows that the population alignment and energy growth are highly dependent on pulse train periodicity. The population alignment results for the ordered train and the train utilizing a periodic spacing disorder show very limited growth with increasing pulse count. The aperiodic pulse trains with randomly generated pulse spacings do not show clear limitations in growth. These trains yield higher population alignment, meaning that the orientation of $^{14}$N becomes more anisotropic and increasingly favors the z-axis with subsequent kicks. The inset in fig. shows there is clear evidence that disorded-induced effects are visible with eight pulses at the chosen intensity. The fully disordered train used in fig. approaches the highest accessible population alignment value. The maximum time averaged population alignment for $\delta$-kicked low temperature rotors approaches 0.5. For finite laser pulses, bandwidth limitations set an upper bound on which $J$ states may be occupied using Raman transitions. The simulations in fig. consider states up to $J = 20$ in accord with previous studies . We note that our pulse centered at $800$ nm with $40$ nm bandwidth has an energy cutoff of J $\approx 80$. Our room temperature gas has a reduced population alignment limit because of significant occupation of states with $m \neq 0$. The population alignment in our system begins at the isotropic value of $0.33$ and is limited to a maximum value of $\approx 0.40$. The choice of $\tau$ in fig. for the periodically kicked rotor is made to be selective of the regime. Rotational wavepackets made with Raman transitions experience either wavepacket amplification or partial annihilation for many rational values of $\frac{T_{rev}}{\tau}$ , where $T_{rev} = 8.38$ ps, the field-free molecular alignment revival period of $^{14}$N. We avoid these regimes with a proper irrational choice of $\frac{T_{rev}}{\tau}$ for the uniformly spaced pulse train, placing the system in a regime where energy gain is bounded for uniformly spaced trains. Our choice of $\tau = \frac{2\pi + 1}{2\pi}T_{rev}$ is also in adherence to previous work that considered first order transitions when using exciting pulses with finite bandwidth. Fig. includes the $\tau = T_{rev}$ case for comparison. The alignment oscillations for this pulse sequence have been surveyed elsewhere and are caused by centrifugal distortion within a real rotor. The results of the simulation were tested in experiments that introduced kicks with spacing disorder. We create an eight pulse train utilizing a triple nested interferometric pulse stacker to split a single 800 nm, 70 fs pulse originating from a 3 W 1 kHz Ti:Sapph source into eight pulses with estimated peak focal irradiance of $1.2\cdot 10^{14} \frac{W}{cm^2}$ as described previously . The nested interferometer arms can be adjusted to provide several different patterns of disorder. Our data shows that four pulses are adequate to initiate dynamical localization. The pulses are focused into a fused silica cell containing a constant flow of dry nitrogen gas. The kick-induced molecular alignment causes a time dependent optical birefringence in a circularly polarized 400 nm low-intensity overlapping probe pulse with a variable delay. The optical birefringence alters the probe's polarization which is measured using time-gated fast photodiodes placed downstream of a polarizer. The polarization components are measured for two cases in rapid succession using an optical chopper: with and without the pump beam. The normalized difference is proportional to the molecular alignment for a specified delay time between the pump and probe beams. The field free molecular alignment signal is averaged across one revival period after each kick to remove time-dependence and yield population alignment. ~1400 shots are taken for each time dependent alignment value for each of 125 evenly spaced probe delay values. Probe delay values for which the kerr effect is present are not considered. A second delay stage controls the interferometric spacing between the first and second set of four kicks. The two delay stages constantly cycle across the four choices of $\epsilon$ and the eight revival periods after individual kicks to evenly disperse beam drift effects. A more detailed discussion of the experimental setup and measurement procedure is in Cryan et al . We compare the measured population alignment of $^{14}$N excited by an eight pulse train with uniform spacing $\tau$, to $^{14}$N excited by a similar train with a modest amount of time spacing-disorder introduced into the pulse train. The pulse train is set to a value $\tau$ in accord with our previous discussio. The measurement is then repeated after the time separation between the fourth and fifth pulses is changed to $\tau + \epsilon$. A non-zero value of $\epsilon$ thus represents disorder in the pulse spacing. We plot our results in fig. . The results for all measured irrational choices of $\frac{T_{rev}}{\tau}$ show qualitative agreement with simulation results in fig. , confirming that the introduction of any non-zero $\epsilon$ causes an increase in the population alignment. Our results are a consequence of dynamical localization within the quantum rotor. The dynamical localization phenomenon corresponds to a predicted localization of rotation state populations within the periodically kicked quantum rotor. We have shown that cases of periodic kicking lead to a slowing of growth in the sample's population alignment in agreement with the predicted onset of localization within the angular momentum state population distribution. Disordered kicking is shown to cause greater change within the population alignment than periodic kicking: this strongly implies that there is greater rotation state population redistribution within the aperiodically kicked rotor, and hence that any localization phenomenon seen for periodic kicking is either absent or significantly weaker in the aperiodically kicked regime. The increased population alignment for disordered kicking means the observed increase in growth within angular momentum space is inherently accompanied by an increase of the molecular orientation along the z-axis. Simulation results of fig. also show that cases of disordered but periodic kicking exhibit localized behavior similar to that seen in the strictly periodically kicked rotor. This is not surprising - the link between Anderson localization and dynamical localization relies on the presence of a periodic Hamiltonian and not necessarily a strictly periodic kicking potential. The measurement was also performed for a choice of $\tau$ close to a rational fraction of $T_{rev}$. Fig. -f shows data in which ordered kicking does not yield a clear minimum in population alignment. For this set of parameters, the fifth pulse is expected to overlap with the remnants of the recurring half revival created by the first pulse when $\epsilon = 200$ fs. The fifth pulse causes annihilation of the remainder of the initial wavepacket when $\epsilon$ is tuned to the proper spacing, causing a minimum in the population alignment. This rotational wavepacket annihilation was originally demonstrated with two pulses in previous experimental reports . The data in fig. is taken at $\tau$ that do not risk interference from wavepacket annihilation. We conclude with a discussion of the relationship between population alignment and angular momentum state populations. The raw data that was used in this work is the molecular alignment signal acquired by measuring the birefringence of a rotational wavepacket as outlined above. The field-free molecular alignment signal is known to have the form \begin{align} <\psi(t)|cos^{2} \theta |& \psi(t)> = \\ &\sum^{\infty}_{J,m} \alpha^{m}_{J}|a^{m}_{J}|^{2} \nonumber + 2 Re[\sum^{\infty}_{J,m}\beta^{m}_{J}a^{m}_{J+2}a^{m*}_{J}e^{i(2J+3)t}] \end{align} where $a^{m}_{J}$ is the coefficient of the state $|J,m>$ and the coefficients $\alpha^{m}_{J}$ and $\beta^{m}_{J}$ were initially published by Leibscher . For our purposes, we reproduce the values \begin{align} \alpha^{m}_{J} = \frac{1}{2J+1}[\frac{(J+1)^{2} - m^{2}}{2J+3} + \frac{J^2 - m^2}{2J-1}]. \end{align} By averaging the molecular alignment signal across $T_{rev}$ between two applied pulses, we discard the second summation term in Eq. and obtain the population alignment value. The values of $\alpha^{m}_{J}$ increase monotonically with increasing $J$ for cases where $m$ is held constant, with the exception of $m = 0$. We note that our simulation and experiment were performed at room temperature, for which only 12$\ This work provides experimental evidence supporting dynamical localization and also shows how disorder may be used to optimize population alignment. This is the first measurement to encompass a broad range of rotation states that demonstrates effects of dynamical localization as the localization regime is approached by making the pump pulse train strictly periodic. One impact of dynamical localization is the ability to selectively excite specific linear rotor isotopes of a sample . Within air, for example, periodic kicking of $\tau = T_{rev}$ the revival period of $^{14}$N would excite the J-state distribution within Nitrogen while having a bounded effect on Oxygen due to localization within the angular momentum space of oxygen. We have shown an increase in population alignment is attainable by the introduction of disorder in pulse spacings. A theoretical study of alignment of SO$_{2}$ shows a similar increase in alignment for disordered kicks, although Pabst's work places all pulse kicks prior to any initial alignment peak. These results prompt further investigation into the effects of non-periodic pulse trains on molecular alignment. In conclusion, we have studied the effect of non-periodic rotational impulses delivered to a quantum linear rotor. We find that introducing disorder into the spacings between individual pulses of the pulse train results in an increased population alignment of the sample. The increase of population alignment corresponds to an increase in the sample's rotational energy and average angular momentum value, $J$, while conserving the directional angular momentum $m$. The difference in population alignment is explained by the presence of dynamical localization for periodic pulse train kicks, which is no longer present for disordered pulse trains. We acknowledge Johannes Flo$\ss{}$ and Ilya Averbukh for stimulating discussions on the subject of dynamical localization. James Cryan is thanked for his preliminary work on periodically kicked $^{14}$N, which paved much of the way for the work reported. This research is supported through the Stanford PULSE Institute at the SLAC National Accelerator Laboratory by the U.S. Department of Energy, Office of Basic Energy Sciences. \vspace{4mm}\noindent $^{\ddagger}$ Corresponding author. \\ akamalov@stanford.edu \bibliography{biblio} \bibliographystyle{unsrtnat} |
1501.04138 | Title: Ricci Curvature of the Internet Topology
Abstract: Analysis of Internet topologies has shown that the Internet topology has
negative curvature, measured by Gromov's "thin triangle condition", which is
tightly related to core congestion and route reliability. In this work we
analyze the discrete Ricci curvature of the Internet, defined by Ollivier, Lin,
etc. Ricci curvature measures whether local distances diverge or converge. It
is a more local measure which allows us to understand the distribution of
curvatures in the network. We show by various Internet data sets that the
distribution of Ricci cuvature is spread out, suggesting the network topology
to be non-homogenous. We also show that the Ricci curvature has interesting
connections to both local measures such as node degree and clustering
coefficient, global measures such as betweenness centrality and network
connectivity, as well as auxilary attributes such as geographical distances.
These observations add to the richness of geometric structures in complex
network theory.
Body: \includepdf[pages=1-last]{Infocom2015-ricciCurvature.pdf} |
1501.04145 | Title: A twistor approach to the Kontsevich complexes
Abstract: This is a note to revisit interesting results of H. Esnault, C. Sabbah, M.
Saito and J.-D. Yu on the Kontsevich complexes from the viewpoint of mixed
twistor D-modules. We explicitly describe the V-filtration of the mixed twistor
D-modules and their relative de Rham complexes, associated to some meromorphic
functions. We explain how such descriptions imply the results on the Kontsevich
complexes.
Body: \section{Introduction} \subsection{Mixed twistor $\nbigd$-modules} The theory of mixed twistor $\nbigd$-modules was developed by C. Sabbah and the author (, , , , ). Very roughly, mixed twistor $\nbigd$-modules are holonomic $\nbigd$-modules equipped with mixed twistor structure. We have the standard $6$-operations on the derived category of algebraic mixed twistor $\nbigd$-modules on complex algebraic manifolds, which are compatible with the standard $6$-operations for algebraic holonomic $\nbigd$-modules. For any algebraic function $f$ on a complex algebraic manifold $Y$, the associated algebraic flat bundle $(\nbigo_Y,d+df)$ is naturally enhanced to an algebraic pure twistor $\nbigd$-module on $Y$. As a result, we can say that many holonomic $\nbigd$-modules are naturally enhanced to mixed twistor $\nbigd$-modules. One of general issues is to describe such mixed twistor $\nbigd$-modules as explicitly as possible. Once we have an explicit description of a mixed twistor $\nbigd$-module, we might have a chance to relate it with a more concrete object, and to apply a general theory of mixed twistor $\nbigd$-modules for the study of the object. \paragraph{$\nbigr$-modules} Let us recall the concept of $\nbigr$-modules which is one of the ingredients to formulate mixed twistor $\nbigd$-modules. Let $\cnum_{\lambda}$ denote just a complex line with the coordinate $\lambda$. For any complex manifold $Y$, let $\nbigr_Y$ denote the sheaf of algebras on $\cnum_{\lambda}\times Y$ obtained as the subalgebra of the sheaf of holomorphic differential operators $\nbigd_{\cnum_{\lambda}\times Y}$ generated by $\lambda p_{\lambda}^{\ast}\Theta_Y$. Here, $p_{\lambda}:\cnum_{\lambda}\times Y\lrarr Y$ denotes the projection, and $\Theta_Y$ denotes the tangent sheaf of $Y$. When we are given a local coordinate system $(x_1,\ldots,x_n)$, then $\lambda\del_{x_i}$ are denoted by $\deldel_{x_i}$ or $\deldel_i$. Mixed twistor $\nbigd$-modules $\nbigt$ on $Y$ are formulated as a pair of $\nbigr_Y$-modules $\nbigm_i$ $(i=1,2)$, a sesqui-linear pairing $C$ of $\nbigm_1$ and $\nbigm_2$, and a weight filtration $W$, satisfying some conditions. (See and for more details on sesqui-linear pairings, weight filtrations, and the conditions.) In this paper, $\nbigm_2$ is called the underlying $\nbigr_Y$-module of $\nbigt$. Note that $\Xi_{\DR}(\nbigm_2):= \iota_1^{-1}\bigl(\nbigm_2/(\lambda-1)\nbigm_2\bigr)$ is naturally a $\nbigd_Y$-module, where $\iota_1:\{1\}\times Y\lrarr \cnum\times Y$ is the inclusion. We call $\Xi_{\DR}(\nbigm_2)$ the $\nbigd$-module underlying $\nbigt$. Then, we reword the general issue as follows: We would like to describe the $\nbigr$-modules underlying mixed twistor $\nbigd$-modules as explicitly as possible. \subsection{Main result} In this paper, we study the $\nbigr$-modules underlying the mixed twistor $\nbigd$-module associated to algebraic meromorphic functions. More precisely, let $X$ be a smooth complex projective manifold with a morphism $f:X\lrarr \proj^1$. Let $D$ be a hypersurface of $X$ such that $f^{-1}(\infty)\subset D$. We assume that $D$ is normal crossing. We put $X^{(1)}:=\cnum_{\tau}\times X$ and $D^{(1)}:=\cnum_{\tau}\times D$. We obtain the meromorphic function $\tau f$ on $(X^{(1)},D^{(1)})$. We have the holonomic $\nbigd_{X^{(1)}}$-module $\nbigm:=\nbigo_{X^{(1)}}(\ast D^{(1)})\,v$ with the flat connection $\nabla$ given by $\nabla v=v\,d(\tau f)$. It is naturally enhanced to a mixed twistor $\nbigd$-module $\nbigt_{\ast}(\tau f,D^{(1)})$. We have the underlying $\nbigr$-module $\nbigl_{\ast}(\tau f,D^{(1)})$. We obtain the $\nbigr_{X}(\ast\tau)$-module $\nbigmtilde:=\nbigl_{\ast}(\tau f,D^{(1)})(\ast\tau)$. We consider the sheaf of subalgebras $\lefttop{\tau}V_0\nbigr_{X^{(1)}}$ in $\nbigr_{X^{(1)}}$ generated by $\nbigo_{\cnum_{\lambda}\times X^{(1)}}$ and $\lambda p_{\lambda}^{\ast}\Theta_{X^{(1)}}(\log \tau)$, where $\Theta_{X^{(1)}}(\log\tau)$ denote the sheaf of vector fields which are logarithmic along $\tau=0$. By a general theory of mixed twistor $\nbigd$-modules, $\nbigmtilde$ is uniquely equipped with an increasing sequence of coherent $\lefttop{\tau}V_0\nbigr_{X^{(1)}}$-submodules $U_{\alpha}\nbigmtilde$ $(\alpha\in\real)$ with the following property: \begin{itemize} \item $\bigcup U_{\alpha}\nbigmtilde=\nbigmtilde$. \item For any $\alpha\in\real$, we have $\epsilon>0$ such that $U_{\alpha}\nbigmtilde=U_{\alpha+\epsilon}\nbigmtilde$. \item We have $\tau U_{\alpha}\nbigmtilde =U_{\alpha-1}\nbigmtilde$ and $\deldel_{\tau} U_{\alpha}\nbigmtilde \subset U_{\alpha+1}\nbigmtilde$ for any $\alpha\in\real.$ \item The induced endomorphisms $\tau\deldel_{\tau}+\lambda\alpha$ on $\Gr^{U}_{\alpha}\nbigmtilde:= U_{\alpha}\nbigmtilde/U_{<\alpha}\nbigmtilde$ are nilpotent for any $\alpha\in\real$. Here, we set $U_{<\alpha}\nbigmtilde:= \bigcup_{\beta<\alpha}U_{\beta}\nbigmtilde$. \item $\Gr^{U}_{\alpha}\nbigmtilde$ are strict, i.e., flat over $\nbigo_{\cnum_{\lambda}}$. \end{itemize} Note that the $V$-filtrations of $\nbigr$-modules underlying mixed twistor $\nbigd$-modules are characterized by a more involved condition. Because $\nbigmtilde$ is naturally equipped with the action of $\lambda^2\del_{\lambda}$, the condition is simplified as above. We shall describe $U_{\alpha}\nbigmtilde$ explicitly. Then, we shall show that their de Rham complex relative to $X^{(1)}\lrarr \cnum_{\tau}$ is quasi-isomorphic to an explicitly given family of complexes, called Kontsevich complexes. \paragraph{Kontsevich complexes} Let us recall the concept of Kontsevich complexes. Put $P:=f^{\ast}(\infty)$ as an effective divisor. The reduced divisor is denoted by $P_{\reduced}$. The multiplication of $df$ induces a morphism $\Omega^k_X(\log D) \lrarr \Omega^{k+1}_X(\log D)\otimes\nbigo_X(P)$. The inverse image of $\Omega^{k+1}_X(\log D) \subset \Omega^{k+1}_X(\log D)\otimes\nbigo_X(P)$ by $df$ is denoted by $\Omega_f^k$. The multiplication of $df$ induces a morphism $df:\Omega_f^k\lrarr \Omega_f^{k+1}$. The exterior derivative induces $d:\Omega_f^k\lrarr \Omega_f^{k+1}$. For any $0\leq\alpha<1$, we set $\Omega^{k}_f(\alpha):= \Omega^{k}_f\otimes \nbigo_X([\alpha P])$. Here, for real numbers $a_i$ and reduced hypersurfaces $H_i$ $(i=1,\ldots,N)$, we set $[a_i]:=\max\{n\in\seisuu\,|\,n\leq a_i\}$ and $\bigl[\sum_{i=1}^{N} a_i H_i\bigr] :=\sum_{i=1}^N [a_i]\,H_i$. We have the induced operators $d:\Omega_f^k(\alpha)\lrarr \Omega_f^{k+1}(\alpha)$ and $df:\Omega_f^k(\alpha)\lrarr \Omega_f^{k+1}(\alpha)$. For any $(\lambda,\tau)\in\cnum^2$, we have the derivative $\lambda d+\tau df: \Omega_f^k(\alpha) \lrarr \Omega_f^{k+1}(\alpha)$. They satisfy the integrability condition $(\lambda d+\tau df)\circ(\lambda d+\tau df)=0$. Thus, we obtain complexes $\bigl( \Omega_f^{\bullet}(\alpha), \lambda d+\tau df \bigr)$. They are called Kontsevich complexes. We clearly have the family version of the complexes. Let $q_X:\cnum_{\lambda}\times X^{(1)}\lrarr X$ denote the projection. We set $\Omegatilde_f^{j}:= \lambda^{-j}q_X^{\ast}\Omega_f^j$. Then, we have the derivative $d+\lambda^{-1}\tau\,df: \Omegatilde_f^j\lrarr \Omegatilde_f^{j+1}$ satisfying $(d+\lambda^{-1}\tau\,df)\circ(d+\lambda^{-1}\tau\,df)=0$. \begin{theorem}[Theorem ] For $0\leq\alpha<1$, $\bigl( \Omegatilde_f^{\bullet}(\alpha),d+\lambda^{-1}\tau\,df \bigr)$ is naturally quasi-isomorphic to the de Rham complex of $U_{\alpha}\nbigmtilde$ relative to the projection $X^{(1)}\lrarr \cnum_{\tau}$. \end{theorem} This theorem reveals a generalized Hodge theoretic property of the Kontsevich complexes, and provides a bridge between the theory of mixed twistor $\nbigd$-modules and the study of Kontsevich complexes. We shall show that it is useful by giving an alternative proof of the interesting results of H. Esnault, M. Kontsevich, C. Sabbah, M. Saito and J.-D. Yu, which we will explain in the next subsection. \subsection{An application} The following theorem was conjectured by M. Kontsevich. It was proved by H. Esnault, C. Sabbah and J.-D. Yu . Some interesting cases were proved by Kontsevich and M. Saito. (See and the appendix of .) \begin{theorem} The dimension of the hypercohomology groups $\hyperh^i\bigl( X, \Omega_f^{\bullet}(\alpha), \lambda d+\tau df \bigr)$ are independent of $(\lambda,\tau)\in\cnum^2$ and any $0\leq \alpha<1$. \end{theorem} We set $\gbigx:=\proj^1_{\tau}\times X$. Let $p_i$ denote the projection of $\gbigx$ onto the $i$-th components. We identify $\nbigo_{\proj^1}(1)$ with $\nbigo_{\proj^1}(\{\infty\})$. The constant function $1$ and the coordinate function $\tau$ are naturally regarded as sections of $\nbigo_{\proj^1}(1)$. We consider the bundles $\gbigk^k_f(\alpha):= p_1^{\ast}\nbigo_{\proj^1}(k) \otimes p_2^{\ast}\Omega^k_{f}(\alpha)$. We have the relative differential operators \[ p_2^{\ast}(d)+\tau p_2^{\ast}(df): \gbigk_f^k(\alpha) \lrarr \gbigk_f^{k+1}(\alpha). \] Thus, we obtain a complex of sheaves $\gbigk_f^{\bullet}(\alpha)$ on $\gbigx$. We have the decreasing filtration $F^{\bullet}$ of the complex $\gbigk_f^{\bullet}(\alpha)$ defined as follows: \begin{equation} F^j\gbigk_f^i(\alpha):= \left\{ \begin{array}{ll} 0 & (i<j)\\ \gbigk_f^i(\alpha) & (i\geq j) \end{array} \right. \end{equation} By Theorem , we obtain a vector bundle $\nbigk^i_f(\alpha):= \hyperr^ip_{1\ast}\bigl( \gbigk^{\bullet}_f(\alpha) \bigr)$ on $\proj^1$. Let $F^j\nbigk^i_f(\alpha)$ denote the image of $\hyperr^ip_{1\ast}\bigl( F^j\gbigk_f^{\bullet}(\alpha) \bigr)$. We have \[ \Gr_F^j\nbigk^i_f(\alpha) \simeq H^{i-j}\bigl(X,\Omega_f^j(\alpha)\bigr) \otimes \nbigo_{\proj^1}(j). \] Hence, $F^{\bullet}$ is the Harder-Narasimhan filtration of $\nbigk^i_f(\alpha)$. Moreover, Sabbah and Yu proved the following theorem in . \begin{theorem} Let $0\leq\alpha<1$. \begin{description} \item[\rm(i)] The bundle $\nbigk^i_f(\alpha)$ is equipped with a naturally induced meromorphic connection $\nabla$ such that \begin{equation} \nabla\cdot \nbigk^i_f(\alpha) \subset \nbigk^i_f(\alpha) \otimes\Omega^1_{\proj^1}\bigl(\{0\}+2\{\infty\}\bigr). \end{equation} Note that the condition {\rm()} implies that $\nabla F^j\nbigk_f^i(\alpha)\subset F^{j-1}\nbigk_f^i(\alpha) \otimes \Omega^1_{\proj^1}\bigl(\{0\}+2\{\infty\}\bigr)$. \item[\rm(ii)] Let $\Res_0(\nabla)$ denote the endomorphism of $\nbigk_f^i(\alpha)_{|0}$ obtained as the residue. Then, any eigenvalue $\beta$ of $\Res_0(\nabla)$ is a rational number such that $-\alpha\leq\beta<-\alpha+1$. \end{description} \end{theorem} Let $\EE_{\beta}\nbigk_f^i(\alpha)_{|0}$ denote the generalized eigen space of $\Res_0(\nabla)$ corresponding to $\beta$. By setting $U_{\gamma}\nbigk_f^i(\alpha)_{|0} =\bigoplus_{-\beta\leq \gamma} \EE_{\beta}\nbigk_f^i(\alpha)_{|0}$, we obtain an increasing filtration $U_{\bullet}\nbigk_f^i(\alpha)_{|0}$ indexed by $\alpha-1<\gamma\leq\alpha$. Note that $\Gr^U_{\gamma}\nbigk_f^i(\alpha)_{|0}$ is naturally isomorphic to $\EE_{-\gamma}\nbigk_f^i(\alpha)_{|0}$. The filtration $F$ on $\nbigk_f^i(\alpha)_{|0}$ naturally induces filtrations on $\Gr^U_{\gamma}\nbigk_f^i(\alpha)_{|0}$, which are also denoted by $F$. Because $\Res_0(\nabla)$ preserves the filtration $U$, we have the induced endomorphisms $\Gr^U_{\gamma}\Res_0(\nabla)$ of $\Gr^U_{\gamma}\nbigk_f^i(\alpha)_{|0}$. Let $N_{\gamma}$ denote the nilpotent part of $\Gr^U_{\gamma}\Res_0(\nabla)$. Then, Sabbah and Yu also proved the following in . \begin{theorem} For any $\alpha-1<\gamma\leq\alpha$, $N_{\gamma}$ gives a strict morphism \[ (\Gr^U_{\gamma}\nbigk^i_{f}(\alpha)_{|0},F) \lrarr (\Gr^U_{\gamma}\nbigk^i_{f}(\alpha)_{|0},F[-1]), \] where $F[-1]$ denotes the filtration defined by $F[-1]^a:=F^{a-1}$. Namely, we have $N_{\gamma}\bigl( F^j\Gr^U_{\gamma}\nbigk_f^i(\alpha)_{|0} \bigr) =\Image(N_{\gamma}) \cap F^{j-1}\Gr^U_{\gamma}\nbigk_f^i(\alpha)_{|0}$. \end{theorem} In \S, we shall explain how to deduce Theorems , and from Theorem and general results for mixed twistor $\nbigd$-modules. We remark that our argument is based on a generalized Hodge theory and some concrete computations of $V$-filtrations. Hence, eventually, it is not completely different from those in . But, the argument in this paper looks more direct. The author thinks that it would be desirable to have many ways to explain the theorems because the complexes $(\Omega_f^{\bullet}(\alpha),\lambda d+\tau df)$ and the bundles $\nbigk^i_f$ seem quite basic. He also hopes that this paper might also be useful to explain how the general theory of twistor $\nbigd$-modules could be applied to the study of specific objects. \paragraph{Outline of the paper} In \S, we study the $\nbigd$-module associated to the meromorphic function as in \S. In particular, we explicitly describe the relative de Rham complexes of their $V$-filtration along $\tau$ (Propositions and ), which is essentially a simpler version of Theorem . Then, we explain how to deduce a part of Theorem . We could include the computations for $\nbigd$-modules in \S to the computations for $\nbigr$-modules in \S after minor modifications. But, the author expects that it would be useful to give an explanation in this simpler situation. In \S, we study the $\nbigr$-module in \S, and we revisit Theorems , and from the viewpoint of mixed twistor $\nbigd$-modules. After the preliminaries in \S--, we explain in \S the main theorem (Theorem ) and how we deduce Theorems , and from general results of mixed twistor $\nbigd$-modules. In \S, we give a description of the $V$-filtration of the $\nbigr_{X^{(1)}}(\ast\tau)$-module $\nbigmtilde$ (Theorem ). Besides the argument in \S for $\nbigd$-modules, we need an additional task to check the strictness. Then, we establish Theorem in \S. \paragraph{Acknowledgements} The author thanks the referee for valuable suggestions to improve this paper. This note is written to understand the intriguing work of H. Esnault, C. Sabbah, M. Saito and J.-D. Yu on the Kontsevich complexes , , . I thank Sabbah for sending an earlier version of and for discussions on many occasions. I also thank him for clarifying the statement of Theorem . I am grateful to Esnault for some discussions and for her kindness. I thank M.-H. Saito for his kindness and support. I thank A. Ishii and Y. Tsuchimoto for their constant encouragement. This study was partially supported by the Grant-in-Aid for Scientific Research (C) (No. 22540078), the Grant-in-Aid for Scientific Research (C) (No. 15K04843), the Grant-in-Aid for Scientific Research (A) (No. 22244003), the Grant-in-Aid for Scientific Research (S) (No. 24224001), the Grant-in-Aid for Scientific Research (S) (No. 17H06127) and the Grant-in-Aid for Scientific Research (S) (No. 16H06335), Japan Society for the Promotion of Science. \section{The case of $\nbigd$-modules} \subsection{Meromorphic flat bundles} We continue to use the setting in \S. We set $X^{(1)}:=\cnum_{\tau}\times X$. Here, $\cnum_{\tau}$ is just an affine line with a coordinate $\tau$. We use the notation $D^{(1)}$, $P^{(1)}$, etc., with a similar meaning. For any complex manifold $Y$ with a hypersurface $Y_1$, let $\nbigo_Y(\ast Y_1)$ denote the sheaf of meromorphic functions on $Y$ whose poles are contained in $Y_1$. For any $\nbigo_Y$-module $M$, let $M(\ast Y_1):=M\otimes_{\nbigo_Y}\nbigo_Y(\ast Y_1)$. If $Y_1$ is given as $\{f=0\}$ for a holomorphic function $f$, we also use the notation $\nbigo_Y(\ast f)$ and $M(\ast f)$. We shall consider the meromorphic flat bundle $\nbigm:=\nbigo_{X^{(1)}}(\ast D^{(1)})\,v$ with \[ \nabla v=v\,d(\tau f) \] of rank one, where $v$ denotes a global frame. The corresponding $\nbigd_{X^{(1)}}$-module is also denoted by $\nbigm$. By the isomorphism $\nbigo_{X^{(1)}}(\ast D^{(1)}) \simeq \nbigm$; $1\longleftrightarrow v$, the connection $\nabla$ is identified with the connection $d+d(\tau f)$ on $\nbigo_{X^{(1)}}(\ast D^{(1)})$. \subsubsection{Local coordinate systems} When we study $\nbigm$ locally around a point $(\tau, Q)\in D_{\reduced}^{(1)}$, we shall use a holomorphic local coordinate neighbourhood $(U,x_1,\ldots,x_n)$ of $Q$ in $X$ with the following property: \begin{itemize} \item $P_{\reduced}\cap U=\bigcup_{i=1}^{\ell_1}\{x_i=0\}$, $H\cap U=\bigcup_{i=\ell_1+1}^{\ell}\{x_i=0\}$, $f_{|U}=\prod_{i=1}^{\ell_1}x_i^{-k_i}$ for $k_i>0$ $(i=1,\ldots,\ell_1)$. \end{itemize} We set $k_i:=0$ $(i=\ell_1+1,\ldots,\ell)$, and the tuple $(k_i\,|\,i=1,\ldots,\ell)\in\seisuu^{\ell}$ is denoted by $\veck$. For any $\vecm\in\seisuu^{\ell}$, we set $x^{\vecm}:=\prod_{i=1}^{\ell}x_i^{m_i}$. In particular, we have $f=x^{-\veck}$ on $U$. We set $U^{(1)}:=\cnum_{\tau}\times U$ which is equipped with a coordinate system $(\tau,x_1,\ldots,x_n)$. We have $\del_i v=-k_i \tau fx_i^{-1}\,v$ for $i=1,\ldots,\ell$, and $\del_i v=0$ for $i=\ell+1,\ldots,n$. We also have $\del_{\tau}v=f\,v$. We set $\vecdelta:=(1,\ldots,1)\in\seisuu^{\ell}$. We have \[ \tau\del_{\tau}(x^{-\vecdelta+\vecm}v) =\tau fx^{-\vecdelta+\vecm}v, \quad \del_ix_i(x^{-\vecdelta+\vecm}v) =(m_i-k_i\tau f)x^{-\vecdelta+\vecm}v. \] Hence, for $i=1,\ldots,\ell_1$, we have \begin{equation} (\tau\del_{\tau}+k_i^{-1}\del_ix_i) (x^{-\vecdelta+\vecm}v) =(m_i/k_i)x^{-\vecdelta+\vecm} v. \end{equation} When we are given a real number $\alpha$, we set $[\alpha k_i]:= \max\bigl\{n\in\seisuu\,\big|\,n\leq \alpha k_i\bigr\}$, and $[\alpha\veck]:= \bigl( [\alpha k_i]\,\big|\, i=1,\ldots,\ell \bigr)\in\seisuu^{\ell}$. \subsection{$V$-filtration along $\tau$} Let $\pi:X^{(1)}\lrarr X$ denote the projection. We naturally regard $\pi^{\ast}\nbigd_{X}$ as the subalgebra of $\nbigd_{X^{(1)}}$. Let $\lefttop{\tau}V_0\nbigd_{X^{(1)}}\subset \nbigd_{X^{(1)}}$ denote the sheaf of subalgebras generated by $\tau\del_{\tau}$ over $\pi^{\ast}\nbigd_{X}$. We shall construct $\lefttop{\tau}V_{0}\nbigd_{X^{(1)}}$-submodules $U_{\alpha}\nbigm$ $(\alpha\in\real)$ of $\nbigm$, and we shall prove that the filtration $U_{\bullet}\nbigm$ is a $V$-filtration of $\nbigm$. \vspace{.1in} For $0\leq \alpha\leq 1$, we have an $\nbigo_{X^{(1)}}$-submodule $\nbigo_{X^{(1)}}\bigl(D^{(1)}+[\alpha P^{(1)}]\bigr)\,v \subset\nbigm$. We set \[ U_{\alpha}\nbigm:= \pi^{\ast}\nbigd_X\cdot \bigl( \nbigo_{X^{(1)}}(D^{(1)}+[\alpha P^{(1)}])\,v \bigr) \subset\nbigm. \] \begin{lemma} Outside $\{\tau=0\}\cap P^{(1)}$, we have $U_{\alpha}\nbigm=\nbigm$ for $0\leq\alpha\leq 1$. \end{lemma} \pf The claim is clear outside $D^{(1)}$. Let $(\tau_0,Q)$ be any point of $D^{(1)}$. We use a coordinate system as in \S. Suppose $Q\not\in P_{\reduced}$. For any $\vecm\in\seisuu^{\ell}_{\geq 0}$, we have $\del_i(x^{-\vecdelta-\vecm}v) =-(m_i+1)x^{-\vecdelta-\vecm}x_i^{-1}v$. Hence, we easily obtain \[ \pi^{\ast}\nbigd_X\cdot\bigl( \nbigo_{X^{(1)}}(D^{(1)}) \bigr)v=\nbigo_{X^{(1)}}(\ast D^{(1)})v \] on a neighbourhood of $(\tau_0,Q)$. Suppose $Q\in P_{\reduced}$ and $\tau_0\neq 0$. For any $\vecm\in\seisuu^{\ell}_{\geq 0}$, we have $\del_i(x^{-\vecdelta-\vecm}v) =-(1+m_i+k_i\tau x^{-\veck})x^{-\vecdelta-\vecm}x_i^{-1}v$. Then, we easily obtain $\nbigo(\ast D^{(1)})v \subset U_{\alpha}\nbigm$. \qed \begin{lemma} We have $\del_{\tau}U_0\nbigm \subset U_1\nbigm$ and $\tau U_1\nbigm \subset U_0\nbigm$. \end{lemma} \pf We have $\del_{\tau}(gx^{-\vecdelta}v) =(\del_{\tau}g)x^{-\vecdelta}v +gx^{-\vecdelta-\veck}v \in U_1\nbigm$. Hence, we have $\del_{\tau}U_0\nbigm \subset U_1\nbigm$. Let us prove $\tau U_1\nbigm\subset U_0\nbigm$. We have only to check it locally around any point of $P^{(1)}_{\reduced}$ by using a coordinate system as in \S. We have \[ \tau \bigl( gx^{-\vecdelta-\veck}v \bigr) =gx^{-\vecdelta}(\tau f)v = -k_i^{-1}\del_{i}(x_i gx^{-\vecdelta}v) +k_i^{-1}(\del_ig)x^{-\vecdelta}x_iv \in U_0\nbigm. \] Thus, we are done. \qed \vspace{.1in} If $\alpha\leq 0$, we take the integer $n$ such that $0\leq \alpha+n<1$, and we set $U_{\alpha}\nbigm:= \tau^nU_{\alpha+n}\nbigm$. For $\alpha\geq 1$, we define $U_{\alpha}\nbigm:= \sum_{\beta+n\leq\alpha} \del_{\tau}^nU_{\beta}\nbigm$. If $\alpha\leq \alpha'$, we have $U_{\alpha}\nbigm\subset U_{\alpha'}\nbigm$. By the construction, for any $\alpha\in\real$, we have $\epsilon>0$ such that $U_{\alpha}\nbigm=U_{\alpha+\epsilon}\nbigm$. We define $U_{<\alpha}\nbigm:= \bigcup_{\beta<\alpha} U_{\beta}\nbigm$. \begin{proposition} $U_{\bullet}\nbigm$ is a $V$-filtration of $\nbigm$ along $\tau$ indexed by the rational numbers with the standard order (up to shift of the degree by $1$). More precisely, the following holds: \begin{itemize} \item $U_{\alpha}\nbigm$ are coherent $\lefttop{\tau}V_0\nbigd_{X^{(1)}}$-modules such that $\bigcup_{\alpha} U_{\alpha}\nbigm=\nbigm$. \item We have $\tau U_{\alpha}\nbigm \subset U_{\alpha-1}\nbigm$ and $\del_{\tau} U_{\alpha}\nbigm \subset U_{\alpha+1}\nbigm$. \item $\tau\del_{\tau}+\alpha$ is nilpotent on $U_{\alpha}\nbigm/U_{<\alpha}\nbigm$. Indeed, we have $\bigl( \tau\del_{\tau}+\alpha \bigr)^{\dim X+1}=0$ on $U_{\alpha}\nbigm/U_{<\alpha}\nbigm$. \end{itemize} \end{proposition} \pf We divide the claim into several lemmas. \begin{lemma} We have a natural action of $\lefttop{\tau}V_0\nbigd_{X^{(1)}}$ on $U_{\alpha}\nbigm$. \end{lemma} \pf It is enough to prove $\tau\del_{\tau}\bigl( U_{\alpha}\nbigm \bigr) \subset U_{\alpha}\nbigm$. We have only to consider the case $0\leq\alpha\leq 1$. We have only to check it locally around any point of $P^{(1)}_{\reduced}$. We use a local coordinate system as in \S. We set $\vecp:=[\alpha\veck]$. We have $\nbigo_{X^{(1)}}(D^{(1)}+[\alpha P^{(1)}]) v =\nbigo_{X^{(1)}}\,x^{-\vecdelta-\vecp}v$ on $U^{(1)}$. We have \[ \tau\del_{\tau} (x^{-\vecdelta-\vecp}v) +(p_i/k_i)x^{-\vecdelta-\vecp}v =-k_i^{-1}\del_i(x_ix^{-\vecdelta-\vecp}v). \] Then, the claim is clear. \qed \begin{lemma} $\lefttop{\tau}V_0\nbigd_{X^{(1)}}$-modules $U_{\alpha}\nbigm$ are coherent. \end{lemma} \pf They are pseudo-coherent over $\nbigo_{X^{(1)}}$ and locally finitely generated over $\lefttop{\tau}V_0\nbigd_{X^{(1)}}$. Hence, they are coherent over $\lefttop{\tau}V_0\nbigd_{X^{(1)}}$. (See .) \qed \vspace{.1in} Let $[\tau\del_{\tau}]$ denote the endomorphism of $U_{\alpha}\nbigm/U_{<\alpha}\nbigm$ induced by $\tau\del_{\tau}$. \begin{lemma} For any $0<\alpha\leq 1$, $[\tau\del_{\tau}]+\alpha$ is nilpotent. Indeed, $\bigl([\tau\del_{\tau}]+\alpha\bigr)^{\dim X}=0$ on $U_{\alpha}\nbigm/U_{<\alpha}\nbigm$. \end{lemma} \pf Because the support of the sheaf $U_{\alpha}\nbigm/U_{<\alpha}\nbigm$ is contained in $P^{(1)}_{\reduced}$, we have only to check the claim locally around any point of $P^{(1)}_{\reduced}$ by using a coordinate system as in \S. We set $\vecp=(p_i):=[\alpha\veck]$. If $\alpha=p_i/k_i$, we have $\bigl( \tau\del_{\tau}+\alpha \bigr)(x^{-\vecdelta-\vecp}v) =-k_i^{-1}\del_i(x_ix^{-\vecdelta-\vecp}v)$. Let $S:=\{i\,|\,\alpha=p_i/k_i\}$. Then, we have \[ \bigl(\tau\del_{\tau}+\alpha\bigr)^{|S|} (x^{-\vecdelta-\vecp}v) =\prod_{i\in S}(-k_i^{-1}\del_i) \Bigl( x^{-\vecdelta-\vecp} \prod_{i\in S}x_i\cdot v\Bigr) \in U_{<\alpha}\nbigm. \] We also have the following equality for any holomorphic function $g$: \[ (\tau\del_{\tau}+\alpha)(gx^{-\vecdelta-\vecp}v) =g\cdot(\tau\del_{\tau}+\alpha)(x^{-\vecdelta-\vecp}v) +(\tau\del_{\tau}g)(x^{-\vecdelta-\vecp}v). \] We have the following for any $i$ with $k_i\neq 0$: \begin{multline} k_i \tau(\del_{\tau}g)x^{-\vecdelta-\vecp}v = -\del_i\Bigl( x_i(\del_{\tau}g)x^{-\vecdelta-\vecp+\veck}v \Bigr) +(\del_i\del_{\tau}g)x^{-\vecdelta-\vecp+\veck}x_iv \\ +(\del_{\tau}g)(k_i-p_i)x^{-\vecdelta-\vecp+\veck}v \in U_{<\alpha}\nbigm. \end{multline} Then, we can easily deduce the claim of the lemma. \qed \begin{lemma} If $N\geq \dim X+1$, we have $(\tau\del_{\tau})^{N}U_0\nbigm \subset \tau U_{<1}\nbigm$. \end{lemma} \pf We have only to check the claim locally around any point of $P^{(1)}_{\reduced}$ by using a coordinate system as in \S. Set $\vecdelta_1:= (\overbrace{1,\ldots,1}^{\ell_1},0,\ldots,0)\in\seisuu^{\ell}$. We have $(\tau\del_{\tau})^{\ell_1}(x^{-\vecdelta}v) =\prod_{i=1}^{\ell_1} \bigl(-k_i^{-1}\del_i\bigr)\cdot x^{-(\vecdelta-\vecdelta_1)}v$. Hence, we have \begin{multline} (\tau\del_{\tau})^{\ell_1+1} (x^{-\vecdelta}v) =\prod_{i=1}^{\ell_1}(-k_i^{-1}\del_i) (\tau f x^{-(\vecdelta-\vecdelta_1)}v) \\ =\tau \prod_{i=1}^{\ell_1}(-k_i^{-1}\del_i) (x^{-\vecdelta-(\veck-\vecdelta_1)}v) \in \tau U_{<1}\nbigm. \end{multline} For any section $s$ of $U_0\nbigm$ and any holomorphic function $g$, we have $\tau\del_{\tau}(gs) =\tau(\del_{\tau}g)\,s +g\,\tau\del_{\tau}s$, and $\tau(\del_{\tau}g)s\in \tau U_{<1}(\nbigm)$. Then, we can deduce the claim of the lemma. \qed \begin{lemma} We have $\sum_{j=0}^{\infty}\del_{\tau}^jU_1\nbigm=\nbigm$. \end{lemma} \pf Let $\nbigm'$ denote the left hand side. We clearly have $\nbigm'\subset\nbigm$. Let $\vecm\in\seisuu_{\geq 0}^{\ell}$. For $i=\ell_1+1,\ldots,\ell$, we have $\del_{i}(x^{-\vecdelta-\vecm}v) =-(m_i+1)x^{-\vecdelta-\vecm}x_i^{-1}v$. We also have $\del_{\tau}(x^{-\vecdelta-\vecm}v) =x^{-\vecdelta-\vecm-\veck}v$. Then, we obtain $\nbigm=\nbigo_{X^{(1)}}(\ast D^{(1)}) \subset\nbigm'$. \qed \vspace{.1in} We obtain Proposition from Lemmas --. \qed \begin{remark} We set $V_{\alpha}\nbigm:=U_{\alpha+1}\nbigm$, then $-\del_{\tau}\tau-\alpha$ is nilpotent on $V_{\alpha}/V_{<\alpha}$. \end{remark} \subsection{Primitive expression} \subsubsection{Primitive expression for sections of $\nbigo_{X^{(1)}}(\ast D^{(1)})$} Let $(U,x_1,\ldots,x_n)$ be a holomorphic coordinate neighbourhood of $X$ as in \S. We set $U^{(1)}:=\cnum_{\tau}\times U$. We set $Y:=\{\tau=x_1=\cdots =x_{\ell}=0\}\subset U^{(1)}$. We shall consider local sections $\nbigo_{X^{(1)}}(\ast D^{(1)})$ on a small neighbourhood of $Y$. Such a section $s$ has the unique Laurent expansion: \[ s=\sum_{\vecm\in\seisuu^{\ell}} \sum_{j\in\seisuu_{\geq 0}} h_{\vecm,j}x^{\vecm}\tau^j. \] Here, $h_{\vecm,j}$ are holomorphic functions on $Y$. Note that if $N>0$ is sufficiently large, depending on $s$, we have $h_{\vecm,j}=0$ unless $m_i\geq -N$ $(i=1,\ldots,\ell)$. We also have the following unique expression: \[ s=\sum_{\vecm\in\seisuu^{\ell}} \sum_{j\in\seisuu_{\geq 0}} h^{(1)}_{\vecm,j}x^{\vecm}\cdot(\tau f)^j. \] Here, $h^{(1)}_{\vecm,j}$ are holomorphic functions on $Y$. Indeed, we have $h^{(1)}_{\vecm,j}=h_{\vecm-j\veck,j}$. \begin{definition} A section $s$ of $\nbigo_{X^{(1)}}(\ast D^{(1)})$ on a neighbourhood of $Y$ is called $(\vecm,j)$-primitive if we have $s=gx^{\vecm}(\tau f)^j$ for a section $g$ of $\nbigo_{X^{(1)}}$ with $g_{|Y}\neq 0$. \end{definition} \begin{definition} A primitive expression of a section $s$ of $\nbigo_{X^{(1)}}(\ast D^{(1)})$ on a neighbourhood of $Y$ is a decomposition \[ s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j} \] where $\nbigs\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ is a finite subset, and each $s_{\vecm,j}$ is $(\vecm,j)$-primitive. \end{definition} \begin{lemma} Any section $s$ of $\nbigo_{X^{(1)}}(\ast D^{(1)})$ on a neighbourhood of $Y$ has a primitive expression. \end{lemma} \pf We have an expression $s=\sum_{(\vecm,j)\in\nbigt} g_{\vecm,j}x^{\vecm}\tau^j$ for a finite subset $\nbigt\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ and holomorphic functions $g_{\vecm,j}$ with $g_{\vecm,j|Y}\neq 0$. Then, we have $s=\sum_{(\vecm,j)\in\nbigt} g_{\vecm,j}x^{\vecm+j\veck} (\tau f)^j$. Each $g_{\vecm,j}x^{\vecm+j\veck}(\tau f)^j$ is $(\vecm+j\veck,j)$-primitive. \qed \vspace{.1in} We consider the partial order on $\seisuu^{\ell}$ defined by $(a_i)\leq (b_i) \stackrel{\rm def}{\Longleftrightarrow} a_i\leq b_i\,(\forall i)$. For any $T\subset\seisuu^{\ell}$, let $\min(T)$ denote the set of the minimal elements in $T$ with respect to the partial order. We also use a similar partial order on $\seisuu^{\ell}\times\seisuu_{\geq 0}$. Let $\pi:\seisuu^{\ell}\times\seisuu_{\geq 0} \lrarr \seisuu^{\ell}$ denote the projection. \begin{lemma} Let $\nbigt\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ be a finite subset. Suppose $\sum_{(\vecm,j)\in\nbigt} g_{\vecm,j}x^{\vecm}(\tau f)^j=0$ on a neighbourhood of $Y$. \begin{itemize} \item For any $\vecm\in\min\pi(\nbigt)$ and any $j\in\seisuu_{\geq 0}$, we have $g_{\vecm,j|Y}=0$. \item For any $(\vecm,j)\in\min(\nbigt)$, we have $g_{\vecm,j|Y}=0$. \end{itemize} \end{lemma} \pf For any $(\vecm,j)$, we have the Laurent expansion \[ g_{\vecm,j}=\sum_{\vecn\in\seisuu_{\geq 0}^{\ell}} \sum_{q\in\seisuu_{\geq 0}} g_{\vecm,j;\vecn,q}x^{\vecn}\tau^q. \] Here, $g_{\vecm,j;\vecn,q}$ are holomorphic functions on $Y$. We have \[ 0=\sum_{(\vecm,j)\in\nbigt} \sum_{\vecn\in\seisuu_{\geq 0}^{\ell}} \sum_{q\geq 0} g_{\vecm,j;\vecn,q} x^{\vecm+\vecn+q\veck} (\tau f)^{j+q}. \] If $\vecm\in\min\pi(\nbigt)$, the coefficient of $x^{\vecm}(\tau f)^j$ is $g_{\vecm,j;0,0}=g_{\vecm,j|Y}$. Hence, we obtain $g_{\vecm,j|Y}=0$. If $(\vecm,j)\in\min(\nbigt)$, the coefficient of $x^{\vecm}(\tau f)^j$ is $g_{\vecm,j;0,0}=g_{\vecm|Y}$. Hence, we obtain $g_{\vecm,j|Y}=0$. \qed \begin{corollary} Let $s$ be a non-zero section of $\nbigo_{X^{(1)}}(\ast D^{(1)})$ on a neighbourhood of $Y$ with a primitive expression $s=\sum_{(\vecm,j)\in\nbigs} g_{\vecm,j}x^{\vecm}(\tau f)^j$. Then, the following holds: \begin{itemize} \item The set $\min\pi(\nbigs)$ is well defined for $s$. For any $\vecm\in\min\pi(\nbigs)$ and any $j\in\seisuu_{\geq 0}$, $g_{\vecm,j|Y}$ is well defined for $s$. \item The set $\min(\nbigs)$ is well defined for $s$. For any $(\vecm,j)\in\min(\nbigs)$, $g_{\vecm,j|Y}$ is well defined for $s$. \qed \end{itemize} \end{corollary} \subsubsection{Subsheaf $\nbigm_0^{\alpha}$} For $0\leq \alpha\leq 1$ and for $N\in\seisuu_{\geq 0}$, we define \[ G_N\nbigm_0^{\alpha}:= \sum_{j=0}^{N} \nbigo_{X^{(1)}}\bigl(D^{(1)}+[\alpha P^{(1)}]\bigr) (\tau f)^jv \subset\nbigm. \] We set $\nbigm_0^{\alpha}:= \bigcup_{N\geq 0}G_N\nbigm_0^{\alpha}$. Let $V_0\nbigd_X$ denote the subalgebra of $\nbigd_X$ generated by $\nbigo_X$ and the logarithmic tangent sheaf $\Theta_X(\log D)$ of $X$ with respect to $D$. We naturally regard $\pi^{\ast}V_0\nbigd_X$ as a subsheaf of $\nbigd_{X^{(1)}}$. \begin{lemma} We have $\nbigm_0^{\alpha} =\pi^{\ast}V_0\nbigd_X\cdot\Bigl( \nbigo_{X^{(1)}}\bigl(D^{(1)}+[\alpha P^{(1)}]\bigr) \,v \Bigr)$. \end{lemma} \pf We have only to check the claim locally around any point of $P^{(1)}_{\reduced}$. We use a local coordinate system as in \S. We set $\vecp=(p_i):=[\alpha\veck]$. Because $\del_ix_i\bigl( vx^{-\vecdelta-\vecp}(\tau f)^j \bigr) =-(p_i+jk_i)v\,x^{-\vecdelta-\vecp}(\tau f)^j -k_i v x^{-\vecdelta-\vecp} (\tau f)^{j+1}$, we easily obtain the claim of the lemma. \qed \vspace{.1in} By the construction, we have $U_{\alpha}\nbigm =\pi^{\ast}\nbigd_X\cdot\nbigm_0^{\alpha}$ for any $0\leq\alpha\leq 1$. \subsubsection{Primitive expression for sections of $U_{\alpha}\nbigm$ $(0\leq\alpha\leq 1)$} We use the notation in \S. Let $0\leq\alpha\leq 1$. Set $\vecp:=[\alpha\veck]\in\seisuu^{\ell}$. For any $\vecn\in\seisuu^{\ell}_{\geq 0}$, we set $\del^{\vecn}:=\prod_{i=1}^{\ell}\del_i^{n_i}$ and $|\vecn|=\sum n_i$. For any $\vecm\in\seisuu^{\ell}$, we have the unique decomposition $\vecm=\vecm_+-\vecm_-$ such that $\vecm_{\pm}\in\seisuu_{\geq 0}^{\ell}$ and $\{i\,|\,m_{+,i}\neq 0\}\cap \{i\,|\,m_{-,i}\neq 0\}=\emptyset$. Any section $s$ of $U_{\alpha}\nbigm$ on a neighbourhood of $Y$ has an expression \[ s=\sum_{\vecn\in\seisuu^{\ell}_{\geq 0}} \del^{\vecn}s_{\vecn} \] as an essentially finite sum, where $s_{\vecn}$ are sections of $\nbigm_0^{\alpha}$. Here, ``essentially finite'' means there exists a finite subset $T\subset\seisuu^{\ell}_{\geq 0}$ such that $s_{\vecn}=0$ unless $\vecn\in T$. \begin{definition} Let $(\vecm,j)\in\seisuu^{\ell}\times\seisuu_{\geq 0}$. A section $s$ of $U_{\alpha}\nbigm$ on a neighbourhood of $Y$ is called $(\vecm,j)$-primitive if $s=\del^{\vecm_-}(g\vecx^{-\vecp-\vecdelta+\vecm_+}(\tau f)^jv)$ for a holomorphic function $g$ with $g_{|Y}\neq 0$, i.e., $g\vecx^{\vecm_+}(\tau f)^j$ is $(\vecm_+,j)$-primitive as a section of $\nbigo_{X^{(1)}}(\ast D^{(1)})$. \qed \end{definition} \begin{definition} Let $s$ be a non-zero section of $U_{\alpha}\nbigm$ on a neighbourhood of $Y$. A primitive expression of $s$ is a decomposition \[ s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j} \] where $\nbigs\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ is a finite set, and $s_{\vecm,j}$ are $(\vecm,j)$-primitive sections of $\nbigm_0^{\alpha}$. \end{definition} This kind of expressions have been used in the study of pure twistor $\nbigd$-modules , for example. \begin{lemma} Any non-zero section $s$ of $U_{\alpha}\nbigm$ on a neighbourhood of $Y$ has a primitive expression. \end{lemma} \pf We have the following expression as an essentially finite sum: \[ s=\sum_{\vecn\in\seisuu^{\ell}_{\geq 0}} \sum_{\vecq\in\seisuu_{\geq 0}^{\ell}} \sum_{j\in\seisuu_{\geq 0}} \del^{\vecn} \Bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^jv \Bigr). \] Here, ``essentially finite'' means that there exists a finite subset $T\subset \seisuu^{\ell}_{\geq 0} \times\seisuu_{\geq 0}^{\ell} \times\seisuu_{\geq 0}$ such that $g_{\vecn,\vecq,j}=0$ unless $(\vecn,\vecq,j)\in T$. We consider the following claim. \begin{description} \item[$(P_a)$:] If $s$ has an expression $s=\sum_{|\vecn|\leq a} \sum_{\vecq,j} \del^{\vecn} \bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^jv \bigr)$ as an essentially finite sum, then $s$ has a primitive expression $s=\sum_{(\vecm,j)\in\nbigs} s_{\vecm,j}$ such that $|\vecm_-|\leq a$ for any $\vecm\in\pi(\nbigs)$. \end{description} If $a=0$, the claim is given by Lemma . We prove $(P_a)$ by assuming $(P_{a-1})$. Note that if $q_i>0$, we have \begin{multline} \del_ix_i\bigl( g x^{-\vecdelta-\vecp+\vecq}x_i^{-1} (\tau f)^jv \bigr) = \Bigl( x_i\del_ig +(-p_i+q_i-1-k_ij)g \Bigr)x^{-\vecdelta-\vecp+\vecq}x_i^{-1}(\tau f)^jv \\ -k_igx^{-\vecdelta-\vecp+\vecq}x_i^{-1} (\tau f)^{j+1}v. \end{multline} By using (), $s$ has an expression \[ s=\sum_{\vecn,\vecq,j} \del^{\vecn}\bigl( g^{(1)}_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^jv \bigr) \] with the following property: \begin{itemize} \item $g^{(1)}_{\vecn,\vecq,j}=0$ unless $|\vecn|\leq a$. \item If $|\vecn|=a$ and $g^{(1)}_{\vecn,\vecq,j}\neq 0$, we have $\{i\,|\,n_i\neq 0,q_i\neq 0\} =\emptyset$ and $g^{(1)}_{\vecn,\vecq,j|Y}\neq 0$. \end{itemize} By applying $(P_{a-1})$ to the lower term $\sum_{|\vecn|<a}\sum_{\vecq,j} \del^{\vecn}\bigl( g^{(1)}_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^jv \bigr)$, we obtain a primitive expression of $s$ as desired. \qed \vspace{.1in} Suppose that we are given a finite set $\nbigs\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ and sections $g_{\vecm,j}$ of $\nbigo_{X^{(1)}}$ on a neighbourhood of $Y$ for $(\vecm,j)\in\nbigs$, such that \begin{equation} \sum_{(\vecm,j)\in\nbigs} \del^{\vecm_-} \Bigl( g_{\vecm,j}x^{-\vecdelta-\vecp+\vecm_+} (\tau f)^jv \Bigr)=0. \end{equation} \begin{lemma} We have $g_{\vecm,j|Y}=0$ for any $\vecm\in\min\pi(\nbigs)$ and $j\in\seisuu$. We also have $g_{\vecm,j|Y}=0$ for any $(\vecm,j)\in\min\nbigs$. \end{lemma} \pf In general, we have the following equality for any section $g$ of $\nbigo_{X^{(1)}}$ on a neighbourhood of $Y$: \begin{multline} \del_i\bigl(gx^{-\vecdelta-\vecp+\vecn}(\tau f)^jv\bigr)= \Bigl( x_i\del_ig -(1+p_i-n_i+jk_i)g \Bigr) x^{-\vecdelta-\vecp+\vecn}x_i^{-1} (\tau f)^jv \\ -k_igx^{-\vecdelta-\vecp+\vecn}x_i^{-1} (\tau f)^{j+1}v. \end{multline} Hence, we have the following expression: \[ \del^{\vecm_-} \bigl(g_{\vecm,j}x^{-\vecdelta-\vecp}x^{\vecm_+} (\tau f)^jv \bigr) =\sum_{0\leq k\leq |\vecm_-|} h_{\vecm,j,k}x^{-\vecdelta-\vecp+\vecm} (\tau f)^{j+k}v, \] where $h_{\vecm,j,k}$ are sections of $\nbigo_{X^{(1)}}$ on a neighbourhood of $Y$ such that $h_{\vecm,j,k|Y}= C_{\vecm,j,k}\cdot g_{\vecm,j|Y}$ for some $C_{\vecm,j,k}\in\rnum$. Because $\{i\,|\,m_{+,i}\neq 0,m_{-,i}\neq 0\}=\emptyset$, we have $C_{\vecm,j,0}\neq 0$. By (), we have the following in $\nbigo_{X^{(1)}}(\ast D^{(1)})$: \[ \sum_{(\vecm,j)\in\nbigs} \sum_{0\leq k\leq |\vecm_-|} h_{\vecm,j,k}x^{-\vecdelta-\vecp+\vecm} (\tau f)^{j+k} =0. \] Take $\vecm\in\min\pi(\nbigs)$. According to Lemma , for any $p\geq 0$, we have \[ \sum_{j+k=p} C_{\vecm,j,k}g_{\vecm,j|Y}=0. \] We obtain $g_{\vecm,j|Y}=0$ by an ascending induction on $j$. Take $(\vecm,j)\in\min\nbigs$. By Lemma , we obtain $h_{\vecm,j,0|Y} =C_{\vecm,j,0}g_{\vecm,j|Y}=0$. Hence, we obtain $g_{\vecm,j|Y}=0$. \qed \vspace{.1in} \begin{corollary} Let $s$ be a section of $U_{\alpha}\nbigm$ on a neighbourhood of $Y$ with a primitive expression \[ s=\sum_{(\vecm,j)\in\nbigs} \del^{\vecm_-}\bigl( g_{\vecm,j}x^{-\vecdelta-\vecp+\vecm_+} (\tau f)^jv \bigr). \] \begin{itemize} \item The set $\min\pi(\nbigs)$ is well defined for $s$. For any $\vecm\in\min\pi(\nbigs)$ and $j\in\seisuu_{\geq 0}$, $g_{\vecm,j|Y}$ is well defined for $s$. \item The set $\min\nbigs$ is well defined for $s$. For any $(\vecm,j)\in\min(\nbigs)$, $g_{\vecm,j|Y}$ is well defined for $s$. \qed \end{itemize} \end{corollary} \subsection{Quasi-isomorphism of complexes} \subsubsection{Statements} We set $\nbigm_0^{\alpha}(-D^{(1)}):= \nbigm_0^{\alpha}\otimes \nbigo_{X^{(1)}}(-D^{(1)})$. The action of $\pi^{\ast}V_0\nbigd_{X}$ on $\nbigm_0^{\alpha}(-D^{(1)})$ naturally induces a complex $\nbigm_0^{\alpha}(-D^{(1)}) \otimes \Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$. We have a natural inclusion of complexes: \begin{equation} \nbigm_0^{\alpha}(-D^{(1)})\otimes \Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \lrarr U_{\alpha}\nbigm\otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}. \end{equation} We shall prove the following proposition in \S. \begin{proposition} For $0\leq\alpha\leq 1$, the morphism {\rm()} is a quasi-isomorphism. \end{proposition} We set $\Omega_f^k(\alpha):= \Omega_f^k\otimes\nbigo([\alpha P])$. We set $\Omega_{f,\tau}^k(\alpha):= \pi^{\ast}\Omega_f^k(\alpha)$. In the case $\alpha=0$, we also use the symbols $\Omega_f^k$ and $\Omega_{f,\tau}^k$. We obtain a complex $\Omega^{\bullet}_{f,\tau}(\alpha)$ with the differential given by $d+\tau\,df$. We have a natural morphism of complexes induced by the correspondence $1\longmapsto v$: \begin{equation} \Omega^{\bullet}_{f,\tau}(\alpha) \lrarr \nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}). \end{equation} We shall prove the following proposition in \S. \begin{proposition} For $0\leq \alpha\leq 1$, the morphism {\rm()} is a quasi-isomorphism. \end{proposition} Let $p_2:X^{(1)}\lrarr\cnum_{\tau}$ denote the projection. \begin{corollary} $\hyperr^ip_{2\ast}\bigl( \Omega_f^{\bullet}(\alpha) \bigr)$ $(0\leq \alpha<1)$ is a locally free $\nbigo_{\cnum_{\tau}}$-module equipped with a naturally induced logarithmic connection $\nabla$. Any eigenvalue $\beta$ of $\Res_0(\nabla)$ is a rational number such that $-\alpha\leq \beta< -\alpha+1$. \end{corollary} \pf By the construction, it is easy to see that $\hyperr^ip_{2\ast}\bigl( \Omega_f^{\bullet}(\alpha) \bigr)$ is a coherent $\nbigo_{\cnum_{\tau}}$-module. By the above propositions, $\hyperr^ip_{2\ast}\bigl( \Omega_f^{\bullet}(\alpha) \bigr)$ is equipped with the action of the differential operator $\tau\del_{\tau}$. The restriction $\hyperr^ip_{2\ast}\bigl( \Omega_f^{\bullet}(\alpha) \bigr)_{|\cnum_{\tau}^{\ast}}$ is a $\nbigd_{\cnum_{\tau}^{\ast}}$-module. Because it is coherent as an $\nbigo_{\cnum_{\tau}^{\ast}}$-module, it is a locally free $\nbigo_{\cnum_{\tau}^{\ast}}$-module. Although we may finish the proof by applying the functoriality of the $V$-filtration, let us recall the proof in this easy situation with the argument in \S3.2 of . The multiplication of $\tau^n$ $(n\geq 0)$ induces an isomorphism of $\pi^{\ast}\nbigd_X$-modules $U_{\alpha}\nbigm \simeq U_{\alpha-n}\nbigm$ because $\alpha<1$. We have the exact sequence of $\pi^{\ast}\nbigd_X$-modules: \[ 0\lrarr U_{\alpha-1}\nbigm \lrarr U_{\alpha}\nbigm \lrarr U_{\alpha}\nbigm/U_{\alpha-1}\nbigm \lrarr 0. \] It induces the following exact sequence of coherent $\nbigo_{\cnum_{\tau}}$-modules: \begin{multline} \hyperr^ip_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \lrarr \hyperr^ip_{2\ast} \bigl( U_{\alpha}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \\ \lrarr \hyperr^ip_{2\ast} \bigl( (U_{\alpha}\nbigm/U_{\alpha-1}\nbigm) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \stackrel{\varphi}{\lrarr} \hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \end{multline} Suppose that $\varphi\neq 0$, and we will deduce a contradiction. For $N\geq 1$, the image of the morphism \[ \hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-N}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \lrarr \hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \] is equal to the image of the multiplication of $\tau^{N-1}$ on $\hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$. Hence, by applying Nakayama's lemma to $\hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$, there exists $N_0$ such that the image of the composite of the following morphisms is non-zero: \begin{multline} \hyperr^ip_{2\ast} \bigl( (U_{\alpha}\nbigm/U_{\alpha-1}\nbigm) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \stackrel{\varphi}{\lrarr} \hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \lrarr \\ C:= \Cok\Bigl( \hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-N_0}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \lrarr \hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr) \Bigr) \end{multline} For any $\beta\in\real$, $\tau\del_{\tau}+\beta$ on $\hyperr^ip_{2\ast}\bigl( \Gr^U_{\beta}(\nbigm) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$ is nilpotent. Hence, the eigenvalues of the endomorphism $\tau\del_{\tau}$ of $\hyperr^ip_{2\ast} \bigl( U_{\beta}\nbigm/U_{\gamma}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$ are contained in $\closedopen{-\beta}{-\gamma} =\bigl\{ -\beta\leq y<-\gamma \bigr\}$. In particular, the eigenvalues of the endomorphism $\tau\del_{\tau}$ of $\hyperr^ip_{2\ast} \bigl( U_{\alpha}\nbigm/U_{\alpha-1}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$ are contained in $\closedopen{-\alpha}{-\alpha+1}$. Because $C$ is contained in $\hyperr^{i+1}p_{2\ast} \bigl( U_{\alpha-1}\nbigm/U_{\alpha-N_0}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$, the eigenvalues of the endomorphism $\tau\del_{\tau}$ on $C$ are contained in $\closedopen{-\alpha+1}{-\alpha+N_0}$. Hence, the image of () is $0$, and we arrived at a contradiction, i.e., $\varphi=0$. It implies that the multiplication of $\tau$ on $\hyperr^ip_{2\ast} \bigl( U_{\alpha}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$ is injective. Thus, we obtain that $\hyperr^ip_{2\ast} \bigl( U_{\alpha}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$ is locally free at $\tau=0$. Moreover, we have $\hyperr^ip_{2\ast} \bigl( U_{\alpha}\nbigm \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)_{|\tau=0} \simeq \hyperr^ip_{2\ast} \bigl( (U_{\alpha}\nbigm\big/U_{\alpha-1}\nbigm) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}} \bigr)$, and the eigenvalues of $\tau\del_{\tau}$ are contained in $\closedopen{-\alpha}{-\alpha+1}$. \qed \subsubsection{Proof of Proposition } We have only to check the claim around any point $(\tau_0,Q)$ of $D^{(1)}$. We use a coordinate system $(U,x_1,\ldots,x_n)$ around $Q$ as in \S. Set $\vecp:=[\alpha\veck]$. Let us consider the case $\tau_0\neq 0$. We have $\nbigm_0^{\alpha}(-D^{(1)})= \nbigo_{X^{(1)}}(\ast P_{\reduced}^{(1)}) v$ and $U_{\alpha}\nbigm =\nbigm =\nbigo_{X^{(1)}}(\ast D^{(1)})\,v$ around $(\tau_0,Q)$. For $\ell_1\leq p\leq \ell$, we set $S(p):=\{(0,\ldots,0)\}\times\seisuu^{p-\ell_1}_{\geq 0} \times\{(0,\ldots,0)\} \subset\seisuu^{\ell_1}\times\seisuu^{p-\ell_1} \times\seisuu^{\ell-p} =\seisuu^{\ell}$. We put $H^{(1)}_{\leq p}:=\bigcup_{\ell_1+1\leq i\leq p}\{x_i=0\}$ and $H^{(1)}_{>p}:=\bigcup_{p<i\leq \ell}\{x_i=0\}$ on the neighbourhood of $(\tau_0,Q)$. We consider \[ \nbigm^{\leq p}:= \sum_{\vecn\in S(p)} \del^{\vecn} \nbigm_0^{\alpha} =\nbigo_{X^{(1)}}(H^{(1)}_{> p}) (\ast P^{(1)}_{\reduced}) (\ast H_{\leq p}^{(1)}) v. \] We have $\nbigm^{\leq \ell_1}=\nbigm_0^{\alpha}$ for any $\alpha$. \begin{lemma} If $p\geq \ell_1+1$, the complexes $x_p\nbigm^{\leq p-1} \stackrel{\del_p}{\lrarr} \nbigm^{\leq p-1}$ and $\nbigm^{\leq p} \stackrel{\del_p}{\lrarr} \nbigm^{\leq p}$ are quasi-isomorphic with respect to the inclusion. \end{lemma} \pf For $j\in\seisuu_{\geq 0}$, we consider the following: \[ \lefttop{p}F_j\nbigm^{\leq p} :=\sum_{\substack{\vecn\in S(p)\\ n_p\leq j}} \del^{\vecn} \nbigm_0^{\alpha} =\nbigo_{X^{(1)}}(H^{(1)}_{> p}) (\ast P^{(1)}_{\reduced}) (\ast H^{(1)}_{\leq p-1}) x_p^{-j-1}v. \] We have $\lefttop{p}F_0\nbigm^{\leq p} =\nbigm^{\leq p-1}$. We have the morphisms of sheaves $\del_p: \lefttop{p}F_j\nbigm^{\leq p} \lrarr \lefttop{p}F_{j+1}\nbigm^{\leq p}$. We can easily check the following by direct computations. \begin{itemize} \item If $j\geq 1$, the induced morphisms of $\nbigo_{X^{(1)}}/x_p\nbigo_{X^{(1)}}$-modules $\lefttop{p}F_j\big/\lefttop{p}F_{j-1} \lrarr \lefttop{p}F_{j+1}\big/\lefttop{p}F_{j}$ are isomorphisms. \item The morphism $\lefttop{p}F_0\nbigm^{\leq p} =\nbigm^{\leq p-1} \lrarr \lefttop{p}F_1/\lefttop{p}F_0$ is a surjection, and the kernel is $x_p\nbigm^{\leq p-1}$. \end{itemize} Then, the claim of the lemma follows. \qed \vspace{.1in} By Lemma , the following inclusion of the complexes of sheaves are quasi-isomorphisms: \begin{multline} \Bigl( \nbigm^{\leq p-1}(-H^{(1)}_{>p-1}) \stackrel{a_1}{\lrarr} \nbigm^{\leq p-1}(-H^{(1)}_{>p-1})\cdot dx_p/x_p \Bigr) \lrarr \\ \Bigl( \nbigm^{\leq p}(-H^{(1)}_{>p}) \stackrel{a_2}{\lrarr} \nbigm^{\leq p}(-H^{(1)}_{>p})\cdot dx_p \Bigr). \end{multline} Here, $a_i$ $(i=1,2)$ are induced by the exterior derivative in the $x_p$-direction. Hence, the following inclusions of the complexes of sheaves are quasi-isomorphisms: \begin{equation} \nbigm^{\leq p-1}(-H^{(1)}_{>p-1}) \otimes \Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log H^{(1)}_{>p-1}) \lrarr \nbigm^{\leq p}(-H^{(1)}_{>p}) \otimes \Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log H^{(1)}_{>p}). \end{equation} We have $\nbigm^{\leq \ell}(-H^{(1)}_{>\ell}) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log H^{(1)}_{>\ell}) =\nbigm\otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}$, and \begin{multline} \nbigm^{\leq \ell_1}(-H^{(1)}_{>\ell_1}) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log H^{(1)}_{>\ell_1}) =\nbigm_0^{\alpha}(-H^{(1)}) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log H^{(1)}) \\ =\nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}). \end{multline} Hence, we are done in the case $\tau_0\neq 0$. \vspace{.1in} Let us consider the case $\tau_0=0$. For $0\leq p\leq \ell$, we regard $\seisuu^{p}= \seisuu^{p} \times\{\overbrace{(0,\ldots,0)}^{\ell-p}\} \subset \seisuu^{\ell}$. We set \[ U_{\alpha}\nbigm^{\leq p}:= \sum_{\vecn\in\seisuu^p_{\geq 0}} \del^{\vecn}\nbigm_0^{\alpha}, \quad\quad \lefttop{p}F_jU_{\alpha}\nbigm^{\leq p} =\sum_{\substack{ \vecn\in\seisuu^p_{\geq 0}\\ n_p\leq j }} \del^{\vecn}\nbigm_0^{\alpha}. \] We have $U_{\alpha}\nbigm^{\leq \ell} =U_{\alpha}\nbigm$ and $\lefttop{p}F_0U_{\alpha}\nbigm^{\leq p} =U_{\alpha}\nbigm^{\leq p-1}$. We consider the following maps \[ \del_p: \lefttop{p}F_jU_{\alpha}\nbigm^{\leq p} \lrarr \lefttop{p}F_{j+1}U_{\alpha}\nbigm^{\leq p}. \] The following lemma is easy to see by Corollary . \begin{lemma} Let $s$ be a section of $U_{\alpha}\nbigm$ on a neighbourhood of $Y$ with a primitive expression \[ s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}. \] Then, $s$ is a section of $\lefttop{p}F_jU_{\alpha}\nbigm^{\leq p}$ if and only if we have $m_i\geq 0$ $(i>p)$ and $m_p\geq -j$ for any $\vecm\in\min\pi(\nbigs)$. \qed \end{lemma} \begin{lemma} If $j\geq 1$, the following induced morphism of sheaves is an isomorphism: \[ \frac{\lefttop{p}F_jU_{\alpha}\nbigm^{\leq p}} {\lefttop{p}F_{j-1}U_{\alpha}\nbigm^{\leq p}} \stackrel{\del_p}{\lrarr} \frac{\lefttop{p}F_{j+1}U_{\alpha}\nbigm^{\leq p}} {\lefttop{p}F_{j}U_{\alpha}\nbigm^{\leq p}} \] \end{lemma} \pf It is surjective by construction. Let $s$ be a non-zero section of $\lefttop{p}F_jU_{\alpha}\nbigm^{\leq p}$ on a neighbourhood of $Y$ with a primitive decomposition $s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}$ such that $\del_ps$ is also a section of $\lefttop{p}F_jU_{\alpha}\nbigm^{\leq p}$. We set $s':=\sum_{m_p=-j}s_{\vecm,j}$ and $s'':=\sum_{m_p>-j}s_{\vecm,j}$. Because $\del_ps''\in \lefttop{p}F_{j}U_{\alpha}\nbigm^{\leq p}$, we obtain $\del_ps'\in\lefttop{p}F_jU_{\alpha}\nbigm^{\leq p}$. If $s'$ is non-zero, $\del_ps'=\sum_{m_p=-j} \del_ps_{\vecm,j}$ is a primitive expression of $\del_ps'$. We obtain $\del_ps'\not\in \lefttop{p}F_{j}U_{\alpha}\nbigm^{\leq p}$, and thus we arrive at a contradiction. Hence, $s'=0$, i.e., $s\in \lefttop{p}F_{j-1}U_{\alpha}\nbigm^{\leq p}$. \qed \begin{lemma} The kernel of the following induced surjection is $x_pU_{\alpha}\nbigm^{\leq p-1}$: \[ U_{\alpha}\nbigm^{\leq p-1} \stackrel{\del_p}{\lrarr} \frac{\lefttop{p}F_1U_{\alpha}\nbigm^{\leq p}} {U_{\alpha}\nbigm^{\leq p-1}} \] \end{lemma} \pf Let $s$ be a section of $U_{\alpha}\nbigm^{\leq p-1}$. We take a primitive expression $s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}$. We set $s':=\sum_{m_p=0}s_{\vecm,j}$ and $s'':=\sum_{m_p>0}s_{\vecm,j}$. We have $s''\in x_pU_{\alpha}\nbigm^{\leq p-1}$. Because \[ \del_{p}\cdot x_p\cdot \bigl( g\cdot x^{-\vecdelta-\vecp}(\tau f)^jv \bigr) =\bigl( x_p\del_p(g) -(p_p+k_pj)g \bigr)x^{-\vecdelta-\vecp}(\tau f)^jv -k_pg\cdot x^{-\vecdelta-\vecp}(\tau f)^{j+1}v, \] we have $\del_ps''\in U_{\alpha}\nbigm^{\leq p-1}$. Hence, we have $\del_ps'\in U_{\alpha}\nbigm^{\leq p-1}$. If $s'\neq 0$, $\del_ps'=\sum_{m_p=0}\del_ps_{\vecm,j}$ is a primitive expression of $s'$. We obtain $\del_ps'\not\in U_{\alpha}\nbigm^{\leq p-1}$, and we arrive at a contradiction. Hence, we have $s'=0$, i.e., $s\in U_{\alpha}\nbigm^{\leq p-1}x_p$. \qed \vspace{.1in} By Lemma and Lemma , for $1\leq p\leq \ell$, the inclusions of the complexes \[ \bigl( x_pU_{\alpha}\nbigm^{\leq p-1} \stackrel{\del_p}{\lrarr} U_{\alpha}\nbigm^{\leq p-1} \bigr) \lrarr \bigl( U_{\alpha}\nbigm^{\leq p} \stackrel{\del_p}{\lrarr} U_{\alpha}\nbigm^{\leq p} \bigr) \] are quasi-isomorphisms. For $0\leq p \leq \ell$, we set $D^{(1)}_{>p}:=\bigcup_{i=p+1}^{\ell}\{x_i=0\}$ on the neighbourhood of $(0,Q)$. We obtain that the following inclusions of the complexes of sheaves are quasi-isomorphisms: \begin{equation} U_{\alpha}\nbigm^{\leq p-1}(-D^{(1)}_{>p-1}) \otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}(\log D^{(1)}_{>p-1}) \lrarr \\ U_{\alpha}\nbigm^{\leq p}(-D^{(1)}_{>p}) \otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}(\log D^{(1)}_{>p}) \end{equation} We have $U_{\alpha}\nbigm^{\leq \ell}(-D^{(1)}_{>\ell}) \otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}(\log D^{(1)}_{>\ell}) =U_{\alpha}\nbigm\otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}$. We also have \[ U_{\alpha}\nbigm^{\leq 0}(-D^{(1)}_{>0}) \otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}(\log D^{(1)}_{>0}) =\nbigm^{\alpha}_0(-D^{(1)})\otimes \Omega_{X^{(1)}/\cnum_{\tau}}^{\bullet}(\log D^{(1)}). \] Hence, Proposition is proved. \qed \subsubsection{Proof of Proposition } We have only to check the claim around any point of $P^{(1)}$. We use the coordinate system as in \S. Set $\vecp:=[\alpha\veck]$. For any non-negative integer $N$, we set $G_N\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \bigr) :=\sum_{j=0}^N \nbigo_{X^{(1)}}x^{-\vecp} (\tau f)^jv$. We set \[ G_N\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr) := G_N\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \bigr) \otimes \Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}). \] The derivative $d$ of the complex $\nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^{\bullet}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$ induces \begin{multline} d: G_N\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr) \lrarr \\ G_{N+1}\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr). \end{multline} We also set $G_{-1}\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \otimes \Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr) :=\Omega^k_{f,\tau}(\alpha) v= \Omega_{f,\tau}^{k}x^{-\vecp}v$. Let $N\geq 0$. Take a section \[ \omega =\sum_{j=0}^N \omega_j x^{-\vecp}(\tau f)^j\cdot v \in G_N\bigl( \nbigm^{\alpha}_0(-D^{(1)})\otimes \Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr), \] where $\omega_j\in \Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$. Suppose that \[ d\omega \in G_{N}\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr). \] Then, $\tau df\wedge \omega_N\,(\tau f)^Nx^{-\vecp} v$ is a section of $G_N\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \bigr) \otimes \Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$. \begin{lemma} We have $df\wedge\omega_N \in \Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$, i.e., $\omega_N$ is a section of $\Omega^k_{f,\tau}$. \end{lemma} \pf Let $s$ be a local section of $\nbigo_{X^{(1)}}$ such that $(\tau f)^{N+1}s \in \sum_{j=0}^N \nbigo_{X^{(1)}} (\tau f)^{j}$. We obtain that $\tau^{N+1}s \in \nbigo_{X^{(1)}}f^{-1}$, and $s\in\nbigo_{X^{(1)}}f^{-1}$. There exists a holomorphic function $t$ such that $s=tf^{-1}$. We obtain $(\tau f)^{N+1}s=(\tau f)^{N} \tau t \in \nbigo_{X^{(1)}}(\tau f)^{N}$. Note that $(df/f)\wedge \omega_N\cdot (\tau f)^{N+1}$ is a section of $\sum_{j=0}^N \Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \cdot (\tau f)^{j}$. By the above argument, we obtain that $(df/f)\wedge \omega_N\cdot (\tau f)^{N+1}$ is a section of $\Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) (\tau f)^{N}$. Hence, $\tau df\wedge\omega_N$ is a section of $\Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$. Then, we obtain that $df\wedge\omega_N$ is a section of $\Omega^{k+1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$. \qed \begin{lemma} We have an expression $\omega_N= (df/f)\wedge \kappa_1 +f^{-1}\kappa_2$, where $\kappa_1$ and $\kappa_2$ are local sections of $\Omega^{k-1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$ and $\Omega^{k}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)})$, respectively. \end{lemma} \pf Let $Q$ be any point of $P_{\reduced}$. Let $U$ be a neighbourhood of $Q$ in $X$. The complex $\bigl(\Omega^{\bullet}_X(\log D),df/f\bigr)$ is acyclic on $U$ because $df/f$ is a nowhere vanishing section of $\Omega^1_{X}(\log D)$ on $U$. If $U$ is sufficiently small, we can take decompositions $\Omega_X^k(\log D) =\nbigb^k\oplus\nbigc^k$ such that the multiplication of $df/f$ induces an isomorphism $\nbigc^k\simeq\nbigb^{k+1}$. Then, the claim of the lemma follows. \qed \vspace{.1in} If $N\geq 1$, we have $f^{-1}\kappa_2(\tau f)^N =\tau \kappa_2(\tau f)^{N-1}$. Hence, we have \[ \omega-d\bigl(\kappa_1(\tau f)^{N-1}x^{-\vecp}v\bigr) \in G_{N-1}\bigl( \nbigm_0^{\alpha}(-D^{(1)}) \otimes\Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr). \] We also have $\kappa_1(\tau f)^{N-1}x^{-\vecp}v \in G_{N-1}$. Let $\omega$ be a local section of $G_N\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$ such that $d\omega$ is a local section of \[ G_{-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr). \] By applying the previous argument successively, we can find a local section $\tau$ of \[ G_{N-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k-1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr) \] such that $\omega-d\tau$ is a local section of $G_{-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$. As a consequence, we have the following. \begin{itemize} \item If a local section $\omega$ of $G_N\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^k_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$ satisfies $d\omega=0$, we can find a local section $\tau$ of $G_{N-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k-1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$ such that $\omega-d\tau$ is a local section of $G_{-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$. \item Let $\omega$ be a local section of $G_{-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$ such that $d\omega=0$. If we have a local section $\tau$ of $G_{N}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k-1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$ such that $\omega=d\tau$, then we can find a local section $\sigma$ of $G_{N-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k-2}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$ such that $\tau-d\sigma$ is a local section of $G_{-1}\bigl( \nbigm^{\alpha}_0(-D^{(1)}) \otimes\Omega^{k-1}_{X^{(1)}/\cnum_{\tau}}(\log D^{(1)}) \bigr)$. We have $\omega=d(\tau-d\sigma)$. \end{itemize} Then, we obtain the claim of Proposition . \qed \section{The case of mixed twistor $\nbigd$-modules} \subsection{Preliminary} \paragraph{$\nbigr$-modules and $\nbigrtilde$-modules} Here, let us recall the concept of $\nbigr$-modules . Let $Y$ be any complex manifold. We set $\nbigy:=\cnum_{\lambda}\times Y$. Let $p:\nbigy\lrarr Y$ denote the projection. Let $\nbigd_{\nbigy}$ denote the sheaf of holomorphic differential operators on $\nbigy$. Let $\nbigr_Y$ denote the sheaf of subalgebras of $\nbigd_{\nbigy}$ generated by $\lambda p^{\ast}\Theta _Y$ over $\nbigo_{\nbigy}$, where $\Theta_Y$ denote the tangent sheaf of $Y$. Let $\nbigrtilde_Y$ denote the sheaf of subalgebras of $\nbigd_{\nbigy}$ generated by $\lambda^2\del_{\lambda}$ over $\nbigr_Y$. A left $\nbigr_Y$-module is equivalent to an $\nbigo_{\nbigy}$-module $\nbigm$ with a relative flat meromorphic connection $\nabla^{\rel}: \nbigm\lrarr \nbigm\otimes \lambda^{-1}\Omega_{\nbigy/\cnum_{\lambda}}^1$. A left $\nbigrtilde_Y$-module is equivalent to an $\nbigo_{\nbigy}$-module $\nbigm$ with a flat meromorphic connection $\nabla: \nbigm\lrarr \nbigm\otimes \lambda^{-1} \Omega^1_{\nbigy}(\log \lambda)$. \paragraph{Push-forward by projection} We recall the functoriality of $\nbigr$-modules with respect to the push-forward by a projection . Suppose that $Y=Z\times W$ for complex manifolds $Z$ and $W$, and that $Z$ is projective. Let $\pi:Y\lrarr W$ denote the projection. For any $\nbigr_Y$-module $\nbign$, we have the $\nbigr_W$-modules $\pi^j_{\dagger}\nbign$ $(-\dim Z\leq j\leq \dim Z)$ as follows. Let $q_Z:\nbigy\lrarr Z$ denote the projection. We set $\Omegatilde^j_Z:=\lambda^{-j}q_Z^{\ast}\Omega^j_Z$. Then, we have a naturally defined complex $\Omegatilde^{\bullet}_Z\otimes\nbign$ on $\nbigy$ induced by the exterior derivative of $\Omega_Z^{\bullet}$ and the relative flat meromorphic connection $\nabla^{\rel}$ of $\nbign$. We have \[ \pi_{\dagger}^j\nbign \simeq R^{j+\dim Z}(\id_{\cnum_{\lambda}}\times\pi)_{\ast} \Bigl( \Omegatilde^{\bullet}_Z\otimes\nbign \Bigr). \] If $\nbign$ is an $\nbigrtilde_Y$-module, then $\pi_{\dagger}^j\nbign$ are naturally $\nbigrtilde_W$-modules. \paragraph{$V$-filtrations} Suppose that $Y=Y_0\times\cnum_t$. For simplicity, we assume that $Y_0$ is relatively compact and open in a larger complex manifold $Y_0'$. Let $V_0\nbigr_Y$ be the sheaf of subalgebras in $\nbigr_Y$ generated by $\lambda p^{\ast}\Theta_Y(\log t)$ over $\nbigo_{\nbigy}$. Let $\nbign$ be an $\nbigr_Y$-module underlying a mixed twistor $\nbigd$-module on $Y$, which can be extended to a mixed twistor $\nbigd$-module on $Y_0'\times\cnum_t$. Then, by the definition of mixed twistor $\nbigd$-module, $\nbign(\ast t)$ is strictly specializable along $t$ as an $\nbigr_Y(\ast t)$-module. Namely, for any $\lambda_0\in\cnum_{\lambda}$, we have a neighbourhood $B(\lambda_0)\subset\cnum_{\lambda}$ of $\lambda_0$ and a unique filtration $V^{(\lambda_0)}\nbign(\ast t)_{|B(\lambda_0)\times Y}$ by coherent $V_0\nbigr_{Y}$-submodules of $\nbign(\ast t)_{|B(\lambda_0)\times Y}$ satisfying the conditions as in \cite[Definition 22.4.1]{Mochizuki-wild}. Note that, for each $a\in\real$, we have the finite subset $\nbigk(a,\lambda_0)\subset \real\times\cnum$ such that $\prod_{u\in\nbigk(a,\lambda_0)} (-\deldel_{t}t+\eigenmap(\lambda,u))$ is nilpotent on $\Gr^{\Vzero}_{a}\bigl(\nbign(\ast t)_{|B(\lambda_0)\times Y}\bigr)$, where we set $\eigenmap(\lambda,(b,\beta)) =\beta-\lambda b-\betabar\lambda^2$ for $(b,\beta)\in\real\times\cnum$. Suppose moreover that $\nbign$ is enhanced to an $\nbigrtilde_Y$-module. As in \cite[Proposition 7.3.1]{sabbah2}, we have $\nbigk(a,\lambda_0)=\{(a,0)\}$, and \[ \lambda^2\del_{\lambda} \Vzero_a \bigl(\nbign(\ast t)_{|B(\lambda_0)\times Y}\bigr) \subset \Vzero_a \bigl(\nbign(\ast t)_{|B(\lambda_0)\times Y}\bigr) \] for any $a\in\real$. We obtain a global filtration $V_{a}\bigl(\nbign(\ast t)\bigr)$ $(a\in\real)$ of $V_0\nbigr_{Y}$-coherent submodules of $\nbign(\ast t)$ by gluing $\Vzero$ $(\lambda_0\in\cnum)$, which is uniquely characterized by the following conditions. \begin{itemize} \item We have $\bigcup_{a\in\real}V_a\bigl( \nbign(\ast t)\bigr) =\nbign(\ast t)$. \item For any $a\in\real$, we have $\epsilon>0$ such that $V_{a}\bigl(\nbign(\ast t)\bigr) =V_{a+\epsilon}\bigl( \nbign(\ast t)\bigr)$. \item We have $t V_a\bigl(\nbign(\ast t)\bigr) =V_{a-1}\bigl(\nbign(\ast t)\bigr)$ and $\deldel_{t}V_a\bigl(\nbign(\ast t)\bigr) \subset V_{a+1}\bigl(\nbign(\ast t)\bigr)$ for any $a\in\real$. \item The induced endomorphisms $\deldel_{t}t+\lambda a$ are nilpotent on \[ \Gr^V_a(\nbign(\ast t)) :=V_{a}\bigl(\nbign(\ast t)\bigr) \big/ V_{<a}\bigl( \nbign(\ast t) \bigr) \] for any $a\in\real$. Here, we set $V_{<a}\bigl(\nbign(\ast t)\bigr) =\bigcup_{b<a} V_b\bigl(\nbign(\ast t)\bigr)$. \item $\Gr^V_a(\nbign(\ast t))$ are strict, i.e., flat over $\nbigo_{\cnum_{\lambda}}$. \end{itemize} We set $\psitilde_a(\nbign):= \Gr^V_a(\nbign(\ast t))$. \subsection{Mixed twistor $\nbigd$-modules} \subsubsection{Wild harmonic bundles} We use the notation in \S. Let $E$ be a product line bundle on $X^{(1)}\setminus P_{\reduced}^{(1)}$ with a frame $e$. The natural holomorphic structure is denoted by $\delbar_E$. The holomorphic line bundle $(E,\delbar_E)$ is equipped with the Higgs field $\theta=d(\tau f)$, and the metric $h$ given by $h(e,e)=1$. Then, $(E,\delbar_E,\theta,h)$ is a wild harmonic bundle. It is homogeneous with respect to the $S^1$-action on $X^{(1)}$ given by $t(\tau,Q)=(t\tau,Q)$ and $t^{\ast}e=e$, in the sense of . We set $\nbigx^{(1)}:=\cnum_{\lambda}\times X^{(1)}$ and $\nbigp^{(1)}_{\reduced}:= \cnum_{\lambda}\times P^{(1)}_{\reduced}$. Let $p:\nbigx^{(1)}\setminus\nbigp^{(1)}_{\reduced} \lrarr X^{(1)}\setminus P^{(1)}_{\reduced}$ denote the projection. Let us recall that a family of $\lambda$-flat bundles is associated to a harmonic bundle in this situation. The $C^{\infty}$-bundle $p^{-1}E$ is equipped with the holomorphic structure $\delbar_1$ given by $\delbar_1p^{-1}(e)=p^{-1}(e)\cdot \lambda d\bigl( \overline{\tau f}\bigr)$. It is also equipped with the family of flat $\lambda$-connections $\DD$ given by $\DD p^{-1}(e) =p^{-1}(e)\cdot \bigl( d(\tau f) +\lambda d\bigl(\overline{\tau f}\bigr) \bigr)$. We set $\upsilon:=p^{-1}(e)\,\exp(-\lambda \overline{\tau f})$. It is a holomorphic frame of $(p^{-1}E,\delbar_1)$. We have $\DD\upsilon=\upsilon\,d(\tau f)$. Let $\nbige$ denote the sheaf of holomorphic sections of $(p^{-1}E,\delbar_1)$. Multiplying $\lambda^{-1}$ to the $(1,0)$-part of $\DD$, we obtain the family of flat connections \[ \DD^f: \nbige \lrarr \nbige\otimes \Bigl( \lambda^{-1} \Omega^1_{(\nbigx^{(1)}\setminus\nbigp^{(1)}_{\reduced}) /\cnum_{\lambda}} \Bigr). \] In terms of the frame $\upsilon$, it is given by $\DD^f\upsilon= \upsilon \cdot \lambda^{-1}d(\tau f)$. The bundle $p^{-1}E$ is naturally $S^1$-equivariant with respect to the action given by $t(\lambda,\tau,Q)=(t\lambda,t\tau,Q)$, for which we have $t^{\ast}\upsilon=\upsilon$. We have $t^{\ast}\DD^f=\DD^f$. Hence, $\DD^f$ is extended to a meromorphic flat connection \[ \nabla: \nbige\lrarr \nbige\otimes \Bigl( \lambda^{-1} \Omega^1_{\nbigx^{(1)}\setminus\nbigp^{(1)}_{\reduced}} \bigl( \log(\{0\}\times(X^{(1)}\setminus P^{(1)}_{\reduced})) \bigr) \Bigr). \] Namely, the derivative in the $\lambda$-direction is induced. (See , for example.) In terms of the frame $\upsilon$, it is given by $\nabla \upsilon =\upsilon d(\lambda^{-1}\tau f)$. The bundle $\nbige$ with $\DD^f$ gives an $\nbigr_{X^{(1)}\setminus P^{(1)}_{\reduced}}$-module, which is also denoted by $\nbige$. Because $\DD^f$ is extended to the meromorphic connection $\nabla$ as mentioned above, $\nbige$ is naturally enhanced to an $\nbigrtilde_{X^{(1)}\setminus P^{(1)}_{\reduced}}$-module. We set $\vecS:=\{|\lambda|=1\}$. Let $\sigma:\vecS\times (X^{(1)}\setminus P^{(1)}_{\reduced}) \lrarr\vecS\times(X^{(1)}\setminus P^{(1)}_{\reduced})$ be given by $\sigma(\lambda,\tau,Q)=(-\lambda,\tau,Q)$. We have the sesqui-linear pairing induced by the metric $h$: \[ C_h: \nbige_{|\vecS\times (X^{(1)}\setminus D^{(1)})} \times \sigma^{\ast} \nbige_{|\vecS\times (X^{(1)}\setminus D^{(1)})} \lrarr \distribution_{\vecS\times (X^{(1)}\setminus D^{(1)})/\vecS}. \] (See for the concept of sesqui-linear pairings.) In this case, it is given by $C_h(\upsilon,\sigma^{\ast}\upsilon) =\exp\bigl( -\lambda (\overline{\tau f}) +\lambdabar(\tau f) \bigr) =\exp\bigl( 2\sqrt{-1} \Image(\lambdabar \tau f) \bigr)$. Thus, we obtain an $\nbigr$-triple $(\nbige,\nbige,C_h)$ on $X^{(1)}\setminus P^{(1)}_{\reduced}$. It is naturally enhanced to an $\nbigrtilde$-triple. \subsubsection{The associated pure twistor $\nbigd$-modules} According to , $(\nbige,\nbige,C_h)$ is uniquely extended to a pure twistor $\nbigd$-module $\gbigt=(\gbigm,\gbigm,\gbigc)$ on $X^{(1)}$ of weight $0$ with the polarization $(\id,\id)$ such that (i) the strict support of $\gbigt$ is $X^{(1)}$, (ii) $\gbigt_{|X^{(1)}\setminus P^{(1)}_{\reduced}}$ is identified with $(\nbige,\nbige,C_h)$. Let us describe the $\nbigr_{X^{(1)}}$-module $\gbigm$. The holomorphic bundle $\nbige$ is naturally extended to an $\nbigo_{\nbigx^{(1)}}(\ast \nbigp^{(1)}_{\reduced})$-module $\nbigq\nbige:= \nbigo_{\nbigx^{(1)}}(\ast \nbigp^{(1)}_{\reduced})\upsilon$ with a meromorphic connection \[ \nabla: \nbigq\nbige \lrarr \nbigq\nbige\otimes \Bigl( \lambda^{-1}\Omega_{\nbigx^{(1)}} \bigl(\log(\{0\}\times X^{(1)})\bigr) \Bigr) \] given by $\nabla\upsilon=\upsilon\cdot d(\lambda^{-1}\tau f)$. It naturally induces an $\nbigrtilde_{X^{(1)}}$-module denoted by $\nbigq\nbige$. \begin{lemma} We have a natural isomorphism of $\nbigr_{X^{(1)}}$-modules $\gbigm\simeq \nbigq\nbige$. \end{lemma} \pf We take a projective birational morphism $F:X^{(2)}\lrarr X^{(1)}$ such that (i) $P^{(2)}_{\reduced}:= F^{-1}(P^{(1)}_{\reduced})$ is simply normal crossing, (ii) $X^{(2)}\setminus P^{(2)}\simeq X^{(1)}\setminus P^{(1)}$, (iii) the zeroes and the poles of $F^{\ast}(\tau f)$ are separated. Let $(E',\delbar_{E'},\theta',h')$ be the pull back of the harmonic bundle $(E,\delbar_E,\theta,h)$ by $F$. Then, it is a good wild harmonic bundle on $(X^{(2)},D^{(2)})$. We have the associated pure twistor $\nbigd$-module $\gbigt'=(\gbigm',\gbigm',\gbigc')$ on $X^{(2)}$. By the construction of $\gbigm'$ in , we have a natural isomorphism $\gbigm'(\ast \nbigp^{(2)}) \simeq F^{\ast}\nbigq\nbige$. Then, we have a natural isomorphism $F_{\dagger}\bigl( \gbigm'\bigr)(\ast \nbigp_{\reduced}^{(1)}) \simeq \nbigq\nbige$. Note that $\gbigm$ is a direct summand of $F_{\dagger}(\gbigm')$ such that $\gbigm_{|\nbigx^{(1)}\setminus \nbigp_{\reduced}^{(1)}} =F_{\dagger}(\gbigm')_{| \nbigx^{(1)}\setminus\nbigp_{\reduced}^{(1)}}$, and that $\gbigm$ is strictly $S$-decomposable. Because the strict support of $\gbigm$ is $X^{(1)}$, a local section of $\gbigm$ is $0$ if its support is contained in $\nbigp^{(1)}$. Hence, we obtain $\gbigm\subset \gbigm(\ast \nbigp^{(1)}_{\reduced}) \simeq F_{\dagger}(\gbigm')(\ast\nbigp^{(1)}_{\reduced}) \simeq \nbigq\nbige$. Moreover, we have $\nbigo_{\nbigx^{(1)}}(-N\nbigp_{\reduced}^{(1)}) \upsilon \subset \gbigm$ for some positive integer $N$. By using $\deldel_{\tau}\upsilon=f\upsilon$, we obtain that $\upsilon$ is a section of $\gbigm$, and hence $\gbigm\supset \nbigq\nbige$. \qed \vspace{.1in} In particular, $\gbigm$ is naturally an $\nbigrtilde_{X^{(1)}}$-module. \subsubsection{The associated mixed twistor $\nbigd$-modules} Recall that $H$ is a hypersurface of $X$ such that $D=H\cup P_{\reduced}$ and $\codim (H\cap P_{\reduced})\geq 2$. We set $H^{(1)}:=\cnum_{\tau}\times H$. We set $\nbigd^{(1)}:=\cnum_{\lambda}\times D^{(1)}$ and $\nbigh^{(1)}:=\cnum_{\lambda}\times H^{(1)}$. We have the localization of $\gbigt$ along $H^{(1)}$ in the category of mixed twistor $\nbigd$-modules on $X^{(1)}$. (See .) It is denoted by $\gbigt[\ast H^{(1)}]$. It consists of an $\nbigr_{X^{(1)}}$-triple $\bigl(\gbigm[!H^{(1)}],\gbigm[\ast H^{(1)}], \gbigc[\ast H^{(1)}]\bigr)$ with a weight filtration $\nbigw$. We have a natural morphism of $\nbigr_{X^{(1)}}$-triples $\gbigt\lrarr\gbigt[\ast H^{(1)}]$. Let us describe the $\nbigr_X(\ast \tau)$-module $\gbigm[\ast H^{(1)}](\ast \tau)$. We have the $\nbigr_{X^{(1)}}(\ast\tau)$-module $\nbigq\nbige(\ast\nbigh^{(1)})(\ast \tau)$ and the $\nbigo_{\nbigx^{(1)}}(\ast\tau)$-submodule $\nbigq\nbige(\nbigh^{(1)})(\ast\tau)$. Let $\nbigmtilde$ denote the $\nbigr_{X^{(1)}}(\ast \tau)$-submodule of $\nbigq\nbige(\ast\nbigh^{(1)})(\ast\tau)$ generated by $\nbigq\nbige(\nbigh^{(1)})(\ast\tau)$ over $\nbigr_{X^{(1)}}$: \[ \nbigmtilde:= \nbigr_{X^{(1)}}\cdot \Bigl( \nbigq\nbige(\nbigh^{(1)})(\ast\tau) \Bigr) \subset \nbigq\nbige(\ast\nbigh^{(1)})(\ast\tau). \] It is naturally an $\nbigrtilde_{X^{(1)}}(\ast\tau)$-submodule of $\nbigq\nbige(\ast\nbigh^{(1)})(\ast\tau)$. \begin{lemma} We have a natural isomorphism of $\nbigr_{X^{(1)}}(\ast\tau)$-modules: \begin{equation} \gbigm[\ast H^{(1)}](\ast\tau) \simeq \nbigmtilde. \end{equation} \end{lemma} \pf We have a natural injection $\gbigm\lrarr\gbigm[\ast H^{(1)}]$ whose restriction to $\nbigx^{(1)}\setminus\nbigh^{(1)}$ is an isomorphism. It induces $\gbigm(\ast\nbigh^{(1)})(\ast \tau) \simeq \bigl( \gbigm[\ast H^{(1)}] \bigr)(\ast\nbigh^{(1)})(\ast\tau)$. We obtain a morphism $\gbigm[\ast H^{(1)}](\ast\tau)\lrarr \gbigm(\ast\nbigh^{(1)})(\ast\tau) =\nbigq\nbige(\ast\nbigh^{(1)})(\ast\tau)$. Take a locally defined holomorphic function $g$ on $U\subset X^{(1)}$ such that $g^{-1}(0)=H^{(1)}$. Let $\iota_g:U\lrarr U\times\cnum_t$ denote the graph of $g$. We have $\iota_{g\dagger}(\gbigm[\ast H^{(1)}]) \simeq \bigl( \iota_{g\dagger}\gbigm \bigr)[\ast t] \subset \iota_{g\dagger}(\gbigm(\ast\nbigh^{(1)}))$. (See \S3.1.2, \S3.3 and \S5.4.1 of .) It implies that a local section of $\gbigm[\ast H^{(1)}]$ is $0$ if its support is contained in $\nbigh^{(1)}$. Hence, we may regard $\gbigm[\ast H^{(1)}](\ast\tau)$ as an $\nbigr_{X^{(1)}}$-submodule of $\nbigq\nbige(\ast\nbigh^{(1)})(\ast\tau)$. By the construction of $\nbigr$-module $\gbigm[\ast H^{(1)}]_{| \nbigx^{(1)}\setminus\{\tau=0\}}$ in \S5.3.3 of , outside $\{\tau=0\}$, $\gbigm[\ast H^{(1)}]$ is generated by $\nbigq\nbige(\nbigh^{(1)})$ over $\nbigr_{X^{(1)}}$. Hence, we have $\gbigm[\ast H^{(1)}](\ast\tau)_{| \nbigx^{(1)}\setminus\{\tau=0\}} = \nbigmtilde_{| \nbigx^{(1)}\setminus\{\tau=0\}}$. Because the actions of $\tau$ on $\gbigm[\ast H^{(1)}](\ast\tau)$ and $\nbigmtilde$ are invertible, we obtain $\gbigm[\ast H^{(1)}](\ast\tau) =\nbigmtilde$ in $\nbigq\nbige(\ast\nbigh^{(1)})(\ast\tau)$. \qed \vspace{.1in} Because the restriction of () to $\nbigx^{(1)}\setminus \nbigd^{(1)}$ is an isomorphism of $\nbigrtilde_{X^{(1)}\setminus D^{(1)}}$-modules, we can easily deduce that () is an isomorphism of $\nbigrtilde_{X^{(1)}}$-modules. We have the $V$-filtration of $\nbigmtilde$. We put $U_{\alpha}\nbigmtilde:=V_{\alpha-1}\nbigmtilde$ for any $\alpha\in\real$, which is a unique filtration satisfying the condition in \S. The filtration $U_{\alpha}$ $(\alpha\in\real)$ is also called the $V$-filtration in this paper. We shall explicitly describe the filtration in \S. \subsubsection{Push-forward} Let $p_1:X^{(1)}\lrarr\cnum_{\tau}$ denote the projection. As the cohomology of the push-forward, we obtain $\nbigr_{\cnum_{\tau}}$-triples for $j\in\seisuu$: \[ p_{1\dagger}^j\gbigt[\ast H^{(1)}] =\Bigl( p_{1\dagger}^{-j}\gbigm[!H^{(1)}], p_{1\dagger}^j\gbigm[\ast H^{(1)}], p_{1\dagger}^j\gbigc[\ast H^{(1)}] \Bigr). \] They are mixed twistor $\nbigd$-modules with the induced filtrations $\nbigw$. \begin{lemma} For any $i$, $p_{1\dagger}^i\bigl( \gbigt[\ast H^{(1)}] \bigr)(\ast \tau)$ are admissible variations of mixed twistor structure on $(\cnum_{\tau},0)$ in the sense of {\rm\cite[\S9]{Mochizuki-MTM}}. In particular, $p^{i}_{1\dagger}\gbigm[\star H^{(1)}](\ast \tau)$ are locally free $\nbigo_{\cnum^2_{\lambda,\tau}}(\ast\tau)$-modules. \end{lemma} \pf We set $\gbigm^1[\star H^{(1)}]:= \gbigm[\star H^{(1)}]\big/ (\lambda-1) \gbigm[\star H^{(1)}]$. We have \[ p_{1\dagger}^i\bigl( \gbigm[\star H^{(1)}] \bigr) \big/ (\lambda-1) p_{1\dagger}^i\bigl( \gbigm[\star H^{(1)}] \bigr) \simeq p_{1\dagger}^i\bigl( \gbigm^1[\star H^{(1)}] \bigr). \] Because $\gbigm^1[\ast H^{(1)}] \simeq \nbigm$, $p_{1\dagger}^i\bigl( \gbigm^1[\ast H^{(1)}] \bigr)_{|\cnum_{\tau}^{\ast}}$ is a locally free $\nbigo_{\cnum_{\tau}^{\ast}}$-module according to Corollary . (We can check the claim directly more easily than Corollary .) Hence, $p_{1\dagger}^i\bigl( \gbigt[\ast H^{(1)}] \bigr)_{|\cnum_{\tau}^{\ast}}$ is a smooth $\nbigr_{\cnum^{\ast}_{\tau}}$-triple. Because $\Gr^{\nbigw}_wp_{1\dagger}^i(\gbigt[\ast H^{(1)}])$ are polarizable pure twistor $\nbigd$-modules, $\Gr^{\nbigw}_wp_{1\dagger}^i\bigl(\gbigt[\ast H^{(1)}]\bigr)(\ast\tau)$ are obtained as the canonical prolongation of good wild polarizable variation of pure twistor structure of weight $w$. Hence, the tuple of the identity $\id:\cnum_{\tau}\lrarr\cnum_{\tau}$, the open subset $\cnum_{\tau}^{\ast}\subset\cnum_{\tau}$, and $p_{1\dagger}^i\bigl(\gbigt[\ast H^{(1)}]\bigr)(\ast\tau)$ with $\nbigw$ gives a cell of the mixed twistor $\nbigd$-module $p_{1\dagger}^i\bigl(\gbigt[\ast H^{(1)}]\bigr)$ in the sense of \S11.1 of . Then, according to Proposition 11.1 of , $p_{1\dagger}^i\bigl(\gbigt[\ast H^{(1)}]\bigr)(\ast \tau)$ is an admissible variation of mixed twistor structure. \qed \vspace{.1in} The $\nbigr_{\cnum_{\tau}}(\ast \tau)$-modules $p^i_{1\dagger}\gbigm[\ast H^{(1)}](\ast \tau)$ are strictly specializable along $\tau$. As in the case of $\nbigmtilde$, we have a unique filtration $V_{a}p^i_{1\dagger}\bigl( \gbigm[\ast H^{(1)}](\ast\tau) \bigr)$ $(a\in\real)$ by $\lefttop{\tau}V_0\nbigr_{\cnum_{\tau}}$-coherent submodules of $p^i_{1\dagger}\bigl(\gbigm[\ast H^{(1)}](\ast\tau)\bigr)$ satisfying the conditions in \S. We set $U_{\alpha}p^i_{1\dagger}\bigl( \gbigm[\ast H^{(1)}](\ast\tau) \bigr):= V_{\alpha-1}p^i_{1\dagger}\bigl( \gbigm[\ast H^{(1)}](\ast\tau) \bigr)$. \begin{corollary} $U_{\alpha}\bigl( p^i_{1\dagger}\gbigm[\ast H^{(1)}](\ast \tau) \bigr)$ are locally free $\nbigo_{\cnum^2_{\lambda,\tau}}$-modules for any $\alpha\in\real$. \end{corollary} \pf Because $p_{1\dagger}^i\bigl(\gbigt[\ast H^{(1)}]\bigr)(\ast\tau)$ are admissible variations of mixed twistor structure, the $\nbigr_{\cnum_{\tau}}(\ast \tau)$-modules $p^i_{1\dagger}\gbigm[\ast H^{(1)}](\ast \tau)$ are good-KMS in the sense of \cite[\S5.11]{Mochizuki-MTM}. By using the $\cnum^{\ast}$-equivariance, we can easily observe that $p^i_{1\dagger}\gbigm[\ast H^{(1)}](\ast \tau)$ are regular-KMS in the sense of \cite[\S5.11]{Mochizuki-MTM}. Hence, $U_{\alpha}\bigl( p^i_{1\dagger}\gbigm[\ast H^{(1)}](\ast \tau) \bigr)$ are locally free $\nbigo_{\cnum^2_{\lambda,\tau}}$-modules. \qed \subsubsection{Hodge structure on the nearby cycle sheaf} We consider the $\cnum^{\ast}$-action on $\cnum_{\lambda}\times\cnum_{\tau}$ given by $t(\lambda,\tau)=(t\lambda,t\tau)$. It induces a $\cnum^{\ast}$-action on $\nbigx^{(1)}$. The $\nbigr_{X^{(1)}}$-modules $\gbigm[\star H^{(1)}]$ $(\star=\ast,!)$ are $\cnum^{\ast}$-equivariant, and the sesqui-linear pairing of $\gbigm[!H^{(1)}]$ and $\gbigm[\ast H^{(1)}]$ is equivariant with respect to the $S^1$-action. We have the induced natural $\cnum^{\ast}$-actions on $p^i_{1\dagger}\gbigm[\star H^{(1)}]$. The induced $S^1$-action is compatible with the sesqui-linear pairing of $p^{-i}_{1\dagger}\gbigm[!H^{(1)}]$ and $p^{i}_{1\dagger}\gbigm[\ast H^{(1)}]$. For $i\in\seisuu$, we have the nearby $\nbigr_{\{0\}}$-triple \[ \psitilde_ap_{1\dagger}^i\gbigt[\ast H^{(1)}] =\bigl( \psitilde_ap_{1\dagger}^{-i}\gbigm[!H^{(1)}], \psitilde_ap_{1\dagger}^i\gbigm[\ast H^{(1)}], \psitilde_ap_{1\dagger}^i\gbigc[\ast H^{(1)}] \bigr). \] Here, $\psitilde_ap_{1\dagger}^i\gbigc[\ast H^{(1)}]$ denotes the induced sesqui-linear pairing. With the relative monodromy filtration $W$, it is a mixed twistor structure. We have the induced $\cnum^{\ast}$-actions on $\psitilde_ap_{1\dagger}^{-i}\gbigm[!H^{(1)}]$ and $\psitilde_ap_{1\dagger}^i\gbigm[\ast H^{(1)}]$, which are compatible with the filtration $W$. The sesqui-linear pairing $\psitilde_ap_{1\dagger}^i\gbigc[\ast H^{(1)}]$ is compatible with the $S^1$-action. Hence, the mixed twistor structure comes from a complex mixed Hodge structure. Namely, let $\gbigh^i_a$ denote the vector space obtained as the fiber of $\psitilde_ap_{1\dagger}^i\gbigm[\ast H^{(1)}]$ over $1\in\cnum_{\lambda}$. Then, we have the decreasing filtrations $F$ and $G$ on $\gbigh^i_a$ such that the $\nbigr$-modules $\psitilde_ap_{1\dagger}^i\gbigm[\ast H^{(1)}]$ and $\psitilde_ap_{1\dagger}^{-i}\gbigm[!H^{(1)}]$ with the $\cnum^{\ast}$-action are isomorphic to the analytification of the Rees modules of the filtered vector spaces $(\gbigh^i_a,F)$ and $(\gbigh^i_a,F)^{\lor}$. Moreover, we have the increasing filtration $W$ on $\gbigh^i_a$ induced by the filtration $W$ of $\psitilde_ap_{1\dagger}^i\gbigm[\ast H^{(1)}]$ so that $(\gbigh^i_a,F,G,W)$ is a complex mixed Hodge structure. For any $k\in\seisuu$, let $N^k_a$ denote the morphisms $\psitilde_ap_{1\dagger}^k\gbigm[\star H] \lrarr \lambda^{-1} \psitilde_ap_{1\dagger}^k\gbigm[\star H]$ $(\star=\ast,!)$ induced as the nilpotent part of $\tau\del_{\tau}$. Let $\gbign^i_a$ denote the endomorphism of $\gbigh^i_a$ induced by $N^i_a$. Recall that $(-N^{-i}_a,-N^i_a)$ gives a morphism of mixed twistor structures $\psitilde_ap_{1\dagger}^i(\gbigt[\ast H]) \lrarr \psitilde_ap_{1\dagger}^i(\gbigt[\ast H]) \otimes \newTate(-1)$. Hence, $\gbign^i_a$ induces a morphism of mixed complex Hodge structure $(\gbigh^i_a,F,G,W) \longrightarrow (\gbigh^i_a,F,G,W)(-1)$, where $(-1)$ indicates the Tate twist. In particular, we obtain the following. \begin{lemma} $\gbign^i_a(F^j) =\Image(\gbign^i_a)\cap F^{j-1}$ for any $i\in\seisuu$, $j\in\seisuu$ and $a\in\real$. \hfill\qed \end{lemma} \subsubsection{Functoriality of $V$-filtrations} Let $\ptilde_1:\nbigx^{(1)}\lrarr \cnum_{\lambda}\times\cnum_{\tau}$ denote the projection. Let $\Omega^1_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}$ denote the kernel of $\Omega^1_{\nbigx^{(1)}} \lrarr \ptilde^{\ast}_{1}\Omega^1_{\cnum^2_{\lambda,\tau}}$. We set $\Omegatilde^1_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} := \lambda^{-1} \Omega^1_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}$ and $\Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} := \bigwedge^k \Omegatilde^1_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}$. Let $\pi:X^{(1)}\lrarr X$ denote the projection. Put $\pitilde:=\id_{\cnum_{\lambda}}\times\pi$. We set $\pi^{\ast}\nbigr_X:= \nbigo_{\nbigx^{(1)}}\otimes_{\pitilde^{-1}\nbigo_{\nbigx}} \pitilde^{-1}\nbigr_X$. For any $\pi^{\ast}\nbigr_{X}$-module $\nbigb$, we have a naturally defined complex $\nbigb\otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}$. By the functoriality of mixed twistor $\nbigd$-modules with respect to the push-forward via projective morphisms, $p^j_{1\dagger} \Gr^U_{\alpha}\gbigm[\ast H^{(1)}]$ are strict for any $\alpha\in\real$. Hence, we obtain the following isomorphism (see the proof of Theorem 3.18 of ): \begin{equation} U_{\alpha}\bigl( p_{1\dagger}^j\gbigm[\ast H^{(1)}](\ast \tau) \bigr) \simeq \hyperr^j\ptilde_{1\ast} \Bigl( U_{\alpha}\bigl( \gbigm[\ast H^{(1)}](\ast \tau) \bigr) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}[\dim X] \Bigr). \end{equation} It is isomorphic to $\hyperr^j\ptilde_{1\ast} \bigl( U_{\alpha}\nbigmtilde \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}[\dim X] \bigr)$. \subsection{Main theorem} \subsubsection{Quasi-isomorphism} Let $q_X:\nbigx^{(1)}\lrarr X$ denote the projection. We set $\Omegatilde^{k}_{f,\lambda,\tau}(\alpha):= \lambda^{-k}q_X^{\ast}\Omega_f^{k}(\alpha)$. We have the following morphism of sheaves: \[ d+\lambda^{-1}\tau df: \Omegatilde^k_{f,\lambda,\tau}(\alpha) \lrarr \Omegatilde^{k+1}_{f,\lambda,\tau}(\alpha). \] We have $\bigl(d+\lambda^{-1}\tau df\bigr)^2=0$. We shall prove the following theorem later in \S. \begin{theorem} We have a natural quasi-isomorphism \[ \bigl( \Omegatilde_{f,\lambda,\tau}^{\bullet}(\alpha), d+\lambda^{-1}\tau df \bigr) \simeq U_{\alpha}\bigl( \gbigm[\ast H^{(1)}](\ast \tau) \bigr) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}. \] It is equivariant with respect to the $\cnum^{\ast}$-action. \end{theorem} \subsubsection{Consequences} We shall explain some consequences of Theorem . \paragraph{Proof of Theorem } We set $\nbigktilde^j_f(\alpha):= \hyperr^j\ptilde_{1\ast}\Omegatilde^{\bullet}_{f,\lambda,\tau}(\alpha)$. By Theorem and the isomorphism (), we have \[ \nbigktilde^j_f(\alpha) \simeq U_{\alpha}\bigl( p_{1\dagger}^{j+\dim X}\gbigm[\ast H^{(1)}](\ast \tau) \bigr) \] for any $j$. Because $U_{\alpha}\bigl( p_{1\dagger}^j\gbigm[\ast H^{(1)}](\ast \tau) \bigr)$ are locally free $\nbigo_{\cnum^2_{\lambda,\tau}}$-modules for any $j$, we obtain that the fiber of $\nbigktilde^j_f(\alpha)$ at $(\lambda,\tau)$ is quasi-isomorphic to the $j$-th hypercohomology group of $(\Omega^{\bullet}_f(\alpha),\lambda d+\tau df)$ by using a general result on any perfect complexes (Lemma below). In particular, the dimension of the hypercohomology groups are independent of $(\lambda,\tau)$, i.e., Theorem holds. \begin{lemma} Let $Y$ be any complex manifold. Let $C^{\bullet}$ be a bounded complex of locally free $\nbigo_Y$-modules. Let $\nbigf$ be any $\nbigo_Y$-module. If the cohomology sheaves $\nbigh^j(C^{\bullet})$ $(j\in\seisuu)$ are locally free $\nbigo_Y$-modules, we naturally have $\nbigh^j(C^{\bullet}) \otimes\nbigf \simeq \nbigh^j(C^{\bullet}\otimes\nbigf)$ for any $j$. \qed \end{lemma} \paragraph{Proof of Theorem } The $\nbigr_{\cnum_{\tau}}$-module $p^j_{1\dagger}\gbigm[\ast H^{(1)}]$ is naturally an $\nbigrtilde_{\cnum_{\tau}}$-module, i.e., it is equipped with the action of $\lambda^2\del_{\lambda}$. The action of $\lambda^2\del_{\lambda}$ preserves $\nbigktilde^{j+\dim X}_f(\alpha)= U_{\alpha}\bigl( p^j_{1\dagger}\gbigm[\ast H^{(1)}] (\ast\tau) \bigr)$. It means that $\nbigktilde^i_f(\alpha)$ is equipped with a meromorphic connection $\nabla$ such that $\lambda\nabla_{\tau\del_{\tau}} \nbigktilde^i_f(\alpha) \subset \nbigktilde^i_f(\alpha)$ and $\nabla_{\lambda^2\del_{\lambda}} \nbigktilde^i_f(\alpha) \subset \nbigktilde^i_f(\alpha)$. The holomorphic bundle $\nbigktilde^i_f(\alpha)$ is $\cnum^{\ast}$-equivariant with respect to the action $t(\lambda,\tau)=(t\lambda,t\tau)$. The meromorphic connection $\nabla$ is equivariant by the construction, and the $\cnum^{\ast}$-action on the bundle is equal to the parallel transport by the connection. The induced vector bundle on $\proj^1$ is isomorphic to $\nbigk^i_f(\alpha)$. It is equipped with the induced meromorphic connection $\nabla$ such that $\nabla\nbigk^i_f(\alpha) \subset \nbigk^i_f(\alpha) \otimes \Omega^1_{\proj^1}(\{0\}+2\{\infty\})$. Here, $0$ corresponds to $[\lambda:\tau]=[1:0]$ and $\infty$ corresponds to $[\lambda:\tau]=[0:1]$. By the construction, it is equal to the connection in Corollary around $0$. Thus, we obtain Theorem . \paragraph{Proof of Theorem } For $\alpha-1<b\leq\alpha$, let $U_{b}\bigl( \nbigktilde^i_f(\alpha)_{|\tau=0} \bigr)$ denote the image of \[ U_{b} p^{i-\dim X}_{1\dagger}\bigl( \gbigm[\ast H^{(1)}] \bigr)_{|\tau=0} \lrarr U_{\alpha} p^{i-\dim X}_{1\dagger}\bigl( \gbigm[\ast H^{(1)}] \bigr)_{|\tau=0} =\nbigktilde^i_f(\alpha)_{|\tau=0}. \] We have $\psitilde_{b-1} p^{i-\dim X}_{1\dagger}\bigl( \gbigm[\ast H^{(1)}] \bigr) \simeq \Gr^{U}_{b}\bigl( \nbigktilde^i_f(\alpha)_{|\tau=0} \bigr)$. We also have the filtration $U$ on $\nbigk^i_f(\alpha)_{|0}$ as in \S. By the constructions, we have the following natural identifications: \[ \gbigh^{i-\dim X}_{b-1}\simeq \Gr^U_{b}\bigl( \nbigktilde^i_f(\alpha)_{|\tau=0} \bigr)_{|\lambda=1} \simeq \Gr^U_b \nbigk^i_f(\alpha)_{|0}. \] (See \S for $\gbigh^k_a$.) By Theorem , we have the filtration $F^{(1)}$ on $\nbigktilde^i_f(\alpha)$ induced by the truncations as in \S (see ()). In particular, it induces a filtration of $\nbigktilde^i_f(\alpha)_{|\cnum_{\lambda}\times\{0\}}$. We also have a filtration $F^{(2)}$ on $\nbigktilde^i_f(\alpha)_{|\cnum_{\lambda}\times\{0\}}$ corresponding to the $\cnum^{\ast}$-action, according to the Rees construction. It is easy to observe that $F^{(1)}=F^{(2)}$. Hence, the Hodge filtration $F$ on $\gbigh^{i-\dim X}_{b-1}$ is equal to the filtration induced by $F^{(1)}$. Then, the claim of Theorem follows from Lemma . \subsection{A description of the $V$-filtration} \subsubsection{The $\nbigr_{X^{(1)}}(\ast \tau)$-module $\nbigmtilde$} We consider the $\nbigr_{X^{(1)}}(\ast \tau)$-module $\nbigmtilde$ given in \S. It has the global section $\upsilon$. For a local coordinate system of $X$ around a point of $D$ as in \S, we have $\deldel_i \upsilon=-k_i(\tau fx_i^{-1})\upsilon$ and $\deldel_{\tau}\upsilon=f\upsilon$. We have \[ \tau\deldel_{\tau}(x^{-\vecdelta+\vecm}\upsilon) =x^{-\vecdelta+\vecm}\tau f\upsilon, \quad \deldel_{i}x_ix^{-\vecdelta+\vecm}\upsilon =x^{-\vecdelta+\vecm} (\lambda m_i-k_i\tau f)\upsilon. \] Hence, for any $i$ with $k_i\neq 0$, we have \begin{equation} (\tau\deldel_{\tau}+k_i^{-1}\deldel_ix_i) (x^{-\vecdelta+\vecm}\upsilon) =x^{-\vecdelta+\vecm} (m_i/k_i)\lambda \upsilon. \end{equation} \subsubsection{The $V$-filtration of $\nbigmtilde$ along $\tau=0$} Let $\pi:X^{(1)}\lrarr X$ denote the projection. Let $\pitilde:\nbigx^{(1)}\lrarr\nbigx$ denote the induced morphism. Set $\pi^{\ast}\nbigr_X:= \nbigo_{\nbigx^{(1)}} \otimes_{\pitilde^{-1}\nbigo_{\nbigx}} \pitilde^{-1}\nbigr_X $. Let $\lefttop{\tau}V_0\nbigr_{X^{(1)}}$ denote the sheaf of subalgebras in $\nbigr_{X^{(1)}}$ generated by $\pi^{\ast}\nbigr_{X}$ and $\tau\deldel_{\tau}$. We shall define $\lefttop{\tau}V_0\nbigr_{X^{(1)}}$-coherent submodules $U_{\alpha}\nbigmtilde$ of $\nbigmtilde$ for any $\alpha\in\real$. For $0\leq\alpha\leq 1$, we set \[ U'_{\alpha}\nbigmtilde:= \pi^{\ast}\nbigr_X \Bigl( \nbigo_{\nbigx^{(1)}}(\nbigd^{(1)}+[\alpha \nbigp^{(1)}])\upsilon \Bigr) \subset \nbigmtilde. \] For any real number $\alpha$, we take the integer $n$ such that $0\leq \alpha+n<1$, and we set $U_{\alpha}\nbigmtilde:= \tau^{n}U'_{\alpha+n}\nbigmtilde$. \begin{remark} We do not need $U'_1(\nbigmtilde)$ for the construction of $U_{\alpha}\nbigmtilde$. We use it in the proof of Theorem {\rm}. \end{remark} We shall prove the following theorem in \S--. \begin{theorem} $U_{\bullet}\nbigmtilde$ is the $V$-filtration of $\nbigmtilde$. \end{theorem} \begin{remark} Theorem {\rm} is not precisely the twistor version of Proposition {\rm} because $\nbigmtilde_{|\{\lambda=1\}}=\nbigm(\ast \tau)$. \end{remark} \subsubsection{Easy property of $U_{\bullet}\nbigmtilde$} We may prove the following lemma by the argument in the proof of Lemma . \begin{lemma} Outside $\{\tau=0\}$, we have $U_{\alpha}\nbigmtilde=\nbigmtilde$. \qed \end{lemma} \begin{lemma} We have a natural action of $\lefttop{\tau}V_0\nbigr_{X^{(1)}}$ on $U_{\alpha}\nbigmtilde$. \end{lemma} \pf We have only to check it around any point of $\nbigp_{\reduced}^{(1)}$ by using a coordinate system as in \S, and the relation (). The argument is the same as that in Lemma . \qed \begin{lemma} For $\alpha-1<\beta\leq \alpha$, we have $\tau U_{\alpha}\nbigmtilde \subset U_{\beta}\nbigmtilde$. As a result, we have $U_{\alpha}\nbigmtilde \subset U_{\alpha'}\nbigmtilde$ for any $\alpha\leq\alpha'$. \end{lemma} \pf We have only to prove the claim around any point of $\nbigp_{\reduced}^{(1)}$ by using a coordinate system in \S. Set $\vecp:=[\alpha\veck]$. It is enough to consider the case $0\leq \alpha<1$. If $\alpha-1<\beta<0$, we have $\tau U_{\alpha}\nbigmtilde \subset U_{\beta}\nbigmtilde$ by the construction. By using the relation $\tau x^{-\vecdelta-\vecp}\upsilon =\tau\deldel_{\tau}\bigl( x^{\veck-\vecdelta-\vecp}\upsilon \bigr)$, we can prove the claim in the case $0\leq \beta<\alpha$. \qed \vspace{.1in} The following lemma is proved as in the case of the $\nbigd$-modules (Lemma ). \begin{lemma} The $\lefttop{\tau}V_0\nbigr_{X^{(1)}}$-module $U_{\alpha}\nbigmtilde$ is coherent. \qed \end{lemma} We set $X_0:=\{0\}\times X\subset X^{(1)}$. We set $U_{<\alpha}\nbigmtilde:= \bigcup_{\beta<\alpha} U_{\beta}\nbigmtilde$. We obtain an $\nbigr_{X_0}$-module $U_{\alpha}\nbigmtilde\big/ U_{<\alpha}\nbigmtilde$. It is equipped with an endomorphism $[\tau\deldel_{\tau}]$ induced by $\tau\deldel_{\tau}$. The following lemma is a generalization of Lemma . \begin{lemma} For $0<\alpha<1$, the action of $[\tau\deldel_{\tau}]+\alpha\lambda$ on $U_{\alpha}\nbigmtilde\big/U_{<\alpha}\nbigmtilde$ is nilpotent. Indeed, the action of $\bigl([\tau\deldel_{\tau}] +\alpha\lambda\bigr)^{\dim X}$ is $0$. \end{lemma} \pf We have only to check the claim around any point of $\nbigp^{(1)}_{\reduced}$ by using a coordinate system as in \S. Set $\vecp:=[\alpha\veck]$. By using $\bigl( \tau\deldel_{\tau}+\lambda\alpha \bigr)x^{-\vecdelta-\vecp}\upsilon =-k_i^{-1}\deldel_{i}(x_ix^{-\vecdelta-\vecp}\upsilon)$, we obtain \[ (\tau\deldel_{\tau}+\lambda\alpha)^N (x^{-\vecdelta-\vecp}\upsilon) \in U_{<\alpha}\nbigmtilde \] for any $N\geq \ell_1$ as in the proof of Lemma . We have $\bigl( \tau\deldel_{\tau}+\lambda\alpha \bigr)(gx^{-\vecdelta-\vecp}\upsilon) =g\cdot(\tau\deldel_{\tau}+\lambda\alpha) (x^{-\vecdelta-\vecp}\upsilon) +(\tau\deldel_{\tau}g)x^{-\vecdelta-\vecp}\upsilon$. We have $(\tau\deldel_{\tau}g)x^{-\vecdelta-\vecp}\upsilon \in\tau U_{\alpha}\nbigmtilde \subset U_{<\alpha}\nbigmtilde$. Then, the claim of the lemma follows. \qed \vspace{.1in} The following lemma can be also checked as in the case of the $\nbigd$-modules (Lemma ). \begin{lemma} If $N\geq \dim X+1$, we have $(\tau\deldel_{\tau})^NU_0\nbigmtilde \subset \tau U_{<1}\nbigmtilde$. \end{lemma} \pf We have only to check the claim around any point of $\nbigp^{(1)}_{\reduced}$ by using a coordinate system as in \S. Let $\vecdelta_1$ be as in the proof of Lemma . We have $(\tau\deldel_{\tau})^{\ell_1}(x^{-\vecdelta}\upsilon) =\prod_{i=1}^{\ell_1} \bigl(-k_i^{-1}\deldel_i\bigr)\cdot x^{-(\vecdelta-\vecdelta_1)}\upsilon$. Hence, we have \begin{multline} (\tau\deldel_{\tau})^{\ell_1+1} (x^{-\vecdelta}\upsilon) =\prod_{i=1}^{\ell_1}(-k_i^{-1}\deldel_i) (\tau f x^{-(\vecdelta-\vecdelta_1)}\upsilon) \\ =\tau \prod_{i=1}^{\ell_1}(-k_i^{-1}\deldel_i) (x^{-\vecdelta-(\veck-\vecdelta_1)}\upsilon) \in \tau U_{<1}\nbigmtilde. \end{multline} For any section $s$ of $U_0\nbigmtilde$ and any holomorphic function $g$, we have $\tau\deldel_{\tau}(gs) =\tau(\deldel_{\tau}g)\,s +g\,\tau\deldel_{\tau}s$, and $\tau(\deldel_{\tau}g)s\in \tau U_{<1}(\nbigmtilde)$. Then, we obtain the claim of the lemma. \qed \vspace{.1in} We obtain the following from Lemma and Lemma . \begin{lemma} $[\tau\deldel_{\tau}]+\alpha\lambda$ is nilpotent on $U_{\alpha}\nbigmtilde/U_{<\alpha}\nbigmtilde$ for any $\alpha\in\real$. \qed \end{lemma} \begin{lemma} We have $\bigcup_{\alpha\in\real} U_{\alpha}\nbigmtilde =\nbigmtilde$. \end{lemma} \pf We have only to prove the claim locally around any point of $\nbigd^{(1)}$. By the construction, $\nbigmtilde':=\bigcup_{\alpha\in\real} U_{\alpha}\nbigmtilde$ is an $\nbigr_{X^{(1)}}(\ast \tau)$-module. By the construction, we have $U_{\alpha}\nbigmtilde \subset \nbigmtilde$ for any $\alpha$, and hence $\nbigmtilde'\subset\nbigmtilde$. We have $\deldel_{\tau}^N x^{-\vecdelta-\vecp} \upsilon =x^{-\vecdelta-\vecp-N\veck}\upsilon$. Hence, we have $\nbigq\nbige(\nbigh^{(1)})(\ast\tau) \subset \nbigmtilde'$. We obtain $\nbigmtilde\subset\nbigmtilde'$. \qed \vspace{.1in} To prove Theorem , it remains to prove that $U_{\alpha}\nbigmtilde\big/U_{<\alpha}\nbigmtilde$ are strict for any $\alpha$, i.e., we need to establish that the multiplication of $\lambda-\lambda_1$ on $U_{\alpha}\nbigmtilde\big/U_{<\alpha}\nbigmtilde$ is injective for any $\lambda_1\in\cnum$. We shall prove it in \S after some preparations. \subsubsection{Preliminary} Let $(U,x_1,\ldots,x_n)$ and $Y$ be as in \S. We set $\nbigy:=\cnum_{\lambda}\times Y$ and $\nbigy^{\lambda_0}:=\{\lambda_0\}\times Y$. In the following, $\nbigy^{(\lambda_0)}$ denotes a neighbourhood of $\nbigy^{\lambda_0}$ in $\nbigy$, and $\nbign^{(\lambda_0)}$ denote the product of $\nbigy^{(\lambda_0)}$ and $\bigl\{(x_1,\ldots,x_{\ell})\,\big|\,|x_i|<\epsilon\bigr\}$ for some $\epsilon>0$. For any section $s$ of $\nbigo_{\nbigx^{(1)}}(\ast \nbigd^{(1)})$ on $\nbign^{(\lambda_0)}$, we have the Laurent expansion \[ s=\sum_{\vecm\in\seisuu^{\ell}} \sum_{j\in\seisuu_{\geq 0}} h_{\vecm,j}x^{\vecm}\tau^j, \] where $h_{\vecm,j}$ are holomorphic functions on $\nbigy^{(\lambda_0)}$. As before, we have the unique expansion \[ s=\sum_{\vecm\in\seisuu^{\ell}} \sum_{j\in\seisuu_{\geq 0}} h^{(1)}_{\vecm,j}x^{\vecm}(\tau f)^j, \] where $h^{(1)}_{\vecm,j}$ are holomorphic functions on $\nbigy^{(\lambda_0)}$. Indeed, we have $h^{(1)}_{\vecm,j} =h_{\vecm-j\veck,j}$. \begin{definition} A section $s$ of $\nbigo_{\nbigx^{(1)}}(\ast \nbigd^{(1)})$ on $\nbign^{(\lambda_0)}$ is called $(\vecm,j)$-primitive if it is expressed as $s=gx^{\vecm}(\tau f)^j$ for a holomorphic function $g$ on $\nbign^{(\lambda_0)}$ with $g_{|\nbigy^{(\lambda_0)}}\neq 0$. \qed \end{definition} \begin{definition} A primitive expression of a section $s$ of $\nbigo_{\nbigx^{(1)}}(\ast\nbigd^{(1)})$ on $\nbign^{(\lambda_0)}$ is an expression \[ s=\sum_{(\vecm,j)\in\nbigs} g_{\vecm,j}x^{\vecm}(\tau f)^{j}, \] where $\nbigs$ is a finite subset in $\seisuu^{\ell}\times\seisuu_{\geq 0}$ and $g_{\vecm,j}$ are holomorphic functions on $\nbign^{(\lambda_0)}$ with $g_{\vecm,j|\nbigy^{(\lambda_0)}}\neq 0$. \qed \end{definition} We use the partial orders on $\seisuu^{\ell}$ and $\seisuu^{\ell}\times\seisuu_{\geq 0}$ as in \S. We reword Lemma and Corollary in this context. \begin{lemma} Any section $s$ of $\nbigo_{\nbigx^{(1)}}(\ast\nbigd^{(1)})$ on $\nbign^{(\lambda_0)}$ has a primitive expression $s=\sum_{(\vecm,j)\in\nbigs} g_{\vecm,j}x^{\vecm}(\tau f)^{j}$. Moreover the following claims hold. \begin{itemize} \item The set $\min\pi(\nbigs)$ is well defined for $s$. For any $\vecm\in\min\pi(\nbigs)$, $g_{\vecm,j|\nbigy^{(\lambda_0)}}$ is well defined for $s$. \item The set $\min(\nbigs)$ is well defined for $s$. For any $(\vecm,j)\in\min(\nbigs)$, $g_{\vecm,j|\nbigy^{(\lambda_0)}}$ is well defined for $s$. \qed \end{itemize} \end{lemma} We reword Lemma in this context. \begin{lemma} Let $\nbigt\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ be a finite subset. Suppose we are given holomorphic functions $g_{\vecm,j}$ $((\vecm,j)\in\nbigt)$ on $\nbign^{(\lambda_0)}$ such that $\sum_{(\vecm,j)\in\nbigt} g_{\vecm,j}x^{\vecm}(\tau f)^j=0$. \begin{itemize} \item For any $\vecm\in\min\pi(\nbigt)$ and any $j\in\seisuu_{\geq 0}$, we have $g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$. \item For any $(\vecm,j)\in\min(\nbigt)$, we have $g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$. \qed \end{itemize} \end{lemma} \subsubsection{Primitive expressions} Let $\Ptilde$ be any effective divisor of $X$ such that $\Ptilde_{\reduced}\subset P_{\reduced}$. We set $\Ptilde^{(1)}:=\cnum_{\tau}\times \Ptilde$ and $\nbigptilde^{(1)}:=\cnum_{\lambda}\times \Ptilde^{(1)}$. For any non-negative integer $N$, we set \[ G_N\nbigmtilde_{\Ptilde}:= \sum_{j=0}^N \nbigo_{\nbigx^{(1)}} \bigl(\nbigd^{(1)}+\nbigptilde^{(1)}\bigr) (\tau f)^j\upsilon \subset \nbigmtilde. \] Let $\nbigmtilde_{\Ptilde}$ be the sheafification of $\bigcup_{N}G_N\nbigmtilde_{\Ptilde}$. We set $U_{\Ptilde}\nbigmtilde:= \pi^{\ast}\nbigr_X\cdot \nbigmtilde_{\Ptilde}$ in $\nbigmtilde$. \vspace{.1in} Let $V_0\nbigr_X\subset\nbigr_X$ be generated by $\lambda\Theta_{\nbigx/\cnum}(\log \nbigd)$ over $\nbigo_{\nbigx}$. Set $\pi^{\ast}V_0\nbigr_X:= \nbigo_{\nbigx^{(1)}} \otimes_{\pitilde^{-1}\nbigo_{\nbigx}} \pitilde^{-1}V_0\nbigr_X$. We naturally regard $\pi^{\ast}V_0\nbigr_X\subset \nbigr_{X^{(1)}}$. We can check the following lemma as in the case of the $\nbigd$-modules (Lemma ). \begin{lemma} We have $\nbigmtilde_{\Ptilde} =\pi^{\ast}V_0\nbigr_X \cdot\Bigl( \nbigo_{\nbigx^{(1)}} \bigl(\nbigd^{(1)}+\nbigptilde^{(1)}\bigr) \upsilon \Bigr)$, and hence \[ U_{\Ptilde}\nbigmtilde= \pi^{\ast}\nbigr_X\cdot \Bigl( \nbigo_{\nbigx^{(1)}} \bigl(\nbigd^{(1)}+\nbigptilde^{(1)}\bigr) \upsilon \Bigr). \] \qed \end{lemma} By the lemma, we have $U_{[\alpha P]}\nbigmtilde =U'_{\alpha}\nbigmtilde$ for any $0\leq\alpha\leq 1$. \vspace{.1in} We use the notation in \S. Take any $\lambda_0\in\cnum$. We set $D_i=\{z_i=0\}$ on the neighbourhood $U$. Let $\vecp\in\seisuu_{\geq 0}^{\ell}$ be determined by $\sum_{i=1}^{\ell} p_iD_i=\Ptilde$ on $U$. Note we have $p_i=0$ for $i>\ell_1$. Let $s$ be a non-zero section of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$. We have an expression $s=\sum_{\vecn\in\seisuu_{\geq 0}^{\ell}} \deldel^{\vecn}s_{\vecn}$ as an essentially finite sum, where $s_{\vecn}$ are sections of $\nbigmtilde_{\Ptilde}$ on $\nbign^{(\lambda_0)}$. \begin{definition} A section $s$ of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ is called $(\vecm,j)$-primitive if it is expressed as \[ s=\deldel^{\vecm_-}\bigl( g x^{-\vecdelta-\vecp+\vecm_+}(\tau f)^j\upsilon\bigr) \] for a holomorphic function $g$ on $\nbign^{(\lambda_0)}$ with $g_{|\nbigy^{(\lambda_0)}}\neq 0$. \qed \end{definition} \begin{definition} A primitive expression of a section $s$ of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ is a decomposition \[ s=\sum_{(\vecm,j)\in\nbigs} s_{(\vecm,j)}, \] where $\nbigs$ is a finite subset of $\seisuu^{\ell}\times\seisuu_{\geq 0}$, and $s_{(\vecm,j)}$ are $(\vecm,j)$-primitive sections of $U_{\Ptilde}\nbigmtilde$. \qed \end{definition} \begin{lemma} Any section $s$ of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ has a primitive expression \[ s=\sum_{(\vecm,j)\in\nbigs} \deldel^{\vecm_-}\bigl( g_{\vecm,j}x^{-\vecdelta-\vecp+\vecm_+} (\tau f)^j\upsilon \bigr). \] \end{lemma} \pf We give an algorithm to obtain a primitive expression. Let $s$ be any section of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$. We have the following expression as an essentially finite sum: \begin{equation} s=\sum_{\vecn\in\seisuu^{\ell}_{\geq 0}} \sum_{\vecq\in\seisuu_{\geq 0}^{\ell}} \sum_{j\in\seisuu_{\geq 0}} \deldel^{\vecn} \Bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^j\upsilon \Bigr). \end{equation} For $a\in\seisuu_{\geq 0}$, we say that an expression () has the property $(R_a)$ if the following holds for any $(\vecn,\vecq,j)$ with $|\vecn|\geq a$: \begin{itemize} \item We have $g_{\vecn,\vecq,j}=0$ unless $\{i\,|\,n_i\neq 0,q_i\neq 0\}=\emptyset$. \item If $g_{\vecn,\vecq,j}\neq 0$, then $\deldel^{\vecn}\bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^j\upsilon \bigr)$ are $(\vecq-\vecn,j)$-primitive. \end{itemize} Take any expression $s= \sum_{\vecn,\vecq,j} \deldel^{\vecn} \bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^j\upsilon \bigr)$. If $a$ is sufficiently large, the expression has the property $(R_a)$. In general, if $q_i>0$, we have \begin{multline} \deldel_ix_i\bigl( g x^{-\vecdelta-\vecp+\vecq}x_i^{-1} (\tau f)^j\upsilon \bigr) = \Bigl( x_i\deldel_ig -\lambda(p_i-q_i+1+jk_i) g \Bigr) x^{-\vecdelta-\vecp+\vecq}x_i^{-1} (\tau f)^j\upsilon \\ -k_ig x^{-\vecdelta-\vecp+\vecq}x_i^{-1} (\tau f)^{j+1}\upsilon. \end{multline} Let $s= \sum_{\vecn,\vecq,j} \deldel^{\vecn} \bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^j\upsilon \bigr)$ be an expression with the property $(R_a)$ such that $a\geq 1$. Applying () and Lemma to each $\deldel^{\vecn} \bigl( g_{\vecn,\vecq,j} x^{-\vecdelta-\vecp+\vecq} (\tau f)^j\upsilon\bigr)$ with $|\vecn|=a-1$, we can obtain an expression with the property $(R_{a-1})$. We can arrive at an expression with the property $(R_0)$, which is a primitive expression of $s$. \qed \vspace{.1in} Suppose that we are given a finite set $\nbigs\subset\seisuu^{\ell}\times\seisuu_{\geq 0}$ and sections of $g_{\vecm,j}$ of $\nbigo_{\nbigx^{(1)}}$ on $\nbign^{(\lambda_0)}$ for $(\vecm,j)\in\nbigs$, such that $0=\sum_{(\vecm,j)\in\nbigs} \deldel^{\vecm_-} \Bigl( g_{\vecm,j}x^{-\vecdelta-\vecp+\vecm_+} (\tau f)^j\upsilon \Bigr)$, where $\vecm=\vecm_+-\vecm_-$ is the decomposition as in \S. \begin{lemma} For any $\vecm\in\min\pi(\nbigs)$ and $j\in\seisuu_{\geq 0}$, we have $g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$. For any $(\vecm,j)\in\min\nbigs$, we have $g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$. \end{lemma} \pf In general, we have the following for any section $g$ of $\nbigo_{\nbigx^{(1)}}$ on $\nbign^{(\lambda_0)}$: \begin{multline} \deldel_i\bigl( gx^{-\vecdelta-\vecp+\vecn}(\tau f)^j\upsilon\bigr) = \Bigl( x_i\deldel_ig -(1+p_i-n_i+jk_i)\lambda g \Bigr) x^{-\vecdelta-\vecp+\vecn}x_i^{-1}(\tau f)^j\upsilon \\ -k_igx^{-\vecdelta-\vecp+\vecn}x_i^{-1} (\tau f)^{j+1}\upsilon. \end{multline} Hence, we have the following expression: \[ \deldel^{\vecm_-} \bigl(g_{\vecm,j}x^{-\vecdelta-\vecp}x^{\vecm_+} (\tau f)^j\upsilon \bigr) =\sum_{0\leq k\leq |\vecm_-|} h_{\vecm,j,k}x^{-\vecdelta-\vecp+\vecm} (\tau f)^{j+k}\upsilon, \] where $h_{\vecm,j,k}$ are sections of $\nbigo_{\nbigx^{(1)}}$ on $\nbign^{(\lambda_0)}$ such that \[ h_{\vecm,j,k|\nbigy^{(\lambda_0)}}= C_{\vecm,j,k}\cdot g_{\vecm,j|\nbigy^{(\lambda_0)}} \] for some $C_{\vecm,j,k}\in\rnum$. By $\{i\,|\,m_{+,i}\neq 0,m_{-,i}\neq 0\}=\emptyset$, we have $C_{\vecm,j,0}\neq 0$. We have the following in $\nbigo_{\nbigx^{(1)}}(\ast \nbigd^{(1)})$: \[ 0=\sum_{(\vecm,j)\in\nbigs} \sum_{0\leq k\leq |\vecm_-|} h_{\vecm,j,k}x^{-\vecdelta-\vecp+\vecm} (\tau f)^{j+k}. \] For $\vecm\in\min\pi(\nbigs)$ and $p\geq 0$, we have $\sum_{j+k=p} C_{\vecm,j,k}g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$. We obtain $g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$ by an ascending induction on $j$. For $(\vecm,j)\in\min\nbigs$, we have $C_{\vecm,j,0}g_{\vecm,j|\nbigy^{(\lambda_0)}}=0$. Thus, the proof of Lemma is finished. \qed \begin{corollary} Let $s$ be a section of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ with a primitive expression \[ s=\sum_{(\vecm,j)\in\nbigs} \deldel^{\vecm_-}\bigl( g_{\vecn,j}x^{-\vecdelta-\vecp+\vecm_+}(\tau f)^j\upsilon \bigr). \] \begin{itemize} \item The set $\min\pi(\nbigs)$ is well defined for $s$. For any $\vecm\in\min\pi(\nbigs)$ and any $j\in\seisuu_{\geq 0}$, $g_{\vecm,j|\nbigy^{(\lambda_0)}}$ is well defined for $s$. \item The set $\min\nbigs$ is well defined for $s$. For any $(\vecm,j)\in\min(\nbigs)$, $g_{\vecm,j|\nbigy^{(\lambda_0)}}$ is well defined for $s$. \qed \end{itemize} \end{corollary} \subsubsection{Variant of primitive expression} Let $P_1$ be an irreducible component of $P_{\reduced}$. We have the inclusion $U_{\Ptilde}\nbigmtilde \subset U_{\Ptilde+P_1}\nbigmtilde$. We assume that $P_1=\{x_1=0\}$ on $(U,x_1,\ldots,x_n)$. \begin{lemma} Any section $s$ of $U_{\Ptilde+P_1}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ has an expression as an essentially finite sum \[ s=s'+\sum_{\vecn\in\seisuu^{\ell}} \deldel^{\vecn_-}\bigl( g_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecn_+}x_1^{-1} \upsilon \bigr), \] such that (i) $s'$ is a section of $U_{\Ptilde}\nbigmtilde$, (ii) each $\deldel^{\vecn_-}\bigl( g_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecn_+}x_1^{-1} \upsilon \bigr)$ is $(\vecn,0)$-primitive as a section of $U_{\Ptilde+P_1}\nbigmtilde$. \end{lemma} \pf We prepare two procedures which are used to obtain an expression with the desired property. \paragraph{(A)} Suppose that a section $s$ of $U_{\Ptilde+P_1}\nbigm$ on $\nbign^{(\lambda_0)}$ has an expression as an essentially finite sum \begin{equation} s=\sum_{\vecn\in\seisuu^{\ell}_{\geq 0}} \sum_{j\geq 0} \deldel^{\vecn}\bigl( g_{\vecn,j}(\lambda,\tau,x) x^{-\vecdelta-\vecp}x_1^{-1} (\tau f)^j\upsilon \bigr). \end{equation} Then, by using the relations (), we obtain an expression of $s$ as an essentially finite sum: \begin{equation} s=s'+\sum_{\vecn\in\seisuu^{\ell}_{\geq 0}} \deldel^{\vecn}\bigl( g^{(1)}_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp}x_1^{-1}\upsilon \bigr). \end{equation} Here, (i) $s'$ is a section of $U_{\Ptilde}\nbigm$, (ii) $\max\bigl\{ |\vecn|\,\big|\, g^{(1)}_{\vecn}\neq 0 \bigr\}\leq \max\bigl\{ |\vecn|\,\big|\, \exists j\,\, g_{\vecn,j}\neq 0 \bigr\}$. \paragraph{(B)} Suppose that we are given a section $\deldel^{\vecn}\bigl( g(\lambda,\tau,x)x^{-\vecdelta-\vecp}x_1^{-1}\upsilon \bigr)$. By applying Lemma and (), we obtain an expression of $\deldel^{\vecn}\bigl( g(\lambda,\tau,x)x^{-\vecdelta-\vecp}x_1^{-1}\upsilon \bigr)$ as an essentially finite sum: \begin{multline} \deldel^{\vecn}\bigl( g(\lambda,\tau,x)x^{-\vecdelta-\vecp}x_1^{-1} \upsilon \bigr) =s' +\sum_{\vecq} \deldel^{\vecn}(g^{(2)}_{\vecq}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecq}x_1^{-1}\upsilon) \\ +\sum_{\vecm\in\seisuu^{\ell}_{\geq 0}} \sum_{j\geq 0} \deldel^{\vecm}\bigl( g^{(3)}_{\vecm,j}x^{-\vecdelta-\vecp}x_1^{-1} (\tau f)^j\upsilon \bigr). \end{multline} Here, (i) $s'$ is a section of $U_{\Ptilde}\nbigm$, (ii) if $g^{(2)}_{\vecq}\neq 0$, $\{i\,|\,n_i\neq 0,q_i\neq 0\}$ is empty and $\deldel^{\vecn}\bigl(g^{(2)}_{\vecq}(\lambda,\tau,x') x^{-\vecdelta-\vecp+\vecq}x_1^{-1}\upsilon\bigr)$ is $(\vecq-\vecn,0)$-primitive, (iii) if $g^{(3)}_{\vecm,j}\neq 0$, we have $|\vecm|<|\vecn|$. \vspace{.1in} Suppose that we have an expression () of $s$ such that $|\vecn|\leq a$ for any $(\vecn,j)$ such that $g_{\vecn,j}\neq 0$. By applying {\bf(A)}, we obtain an expression (). We have $|\vecn|\leq a$ if $g^{(1)}_{\vecn}\neq 0$. By applying {\bf(B)} to $\deldel^{\vecn}\bigl( g^{(1)}_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp}x_1^{-1}\upsilon \bigr)$, we obtain an expression of $s$ as follows: \begin{multline} s= s''+ \sum_{|\vecn|=a}\sum_{\vecq} \deldel^{\vecn}\bigl( g^{(4)}_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecq}x_1^{-1}\upsilon \bigr) \\ +\sum_{|\vecn|<a}\sum_j \deldel^{\vecn}\bigl( g^{(5)}_{\vecn,j} x^{-\vecdelta-\vecp}x_1^{-1} (\tau f)^j\upsilon \bigr). \end{multline} Here, (i) $s''$ is a section of $U_{\Ptilde}\nbigm$, (ii) if $g^{(4)}_{\vecn}\neq 0$, $\{i\,|\,n_i\neq 0,q_i\neq 0\}$ is empty and $\deldel^{\vecn}\bigl( g^{(4)}_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecq}x_1^{-1}\upsilon \bigr)$ is $(\vecq-\vecn,0)$-primitive. Then, we can obtain an expression of $s$ with the desired property by an inductive argument. \qed \begin{lemma} Let $\nbigt_1$ be a finite subset in $\seisuu^{\ell}$ such that $m_1\leq 0$ for any $\vecm\in\nbigt_1$. Suppose that we are given a section $s'$ of $U_{\Ptilde}\nbigm$ and holomorphic functions $g_{\vecm}(\lambda,\tau,x)$ $(\vecm\in\nbigt_1)$ on $\nbign^{(\lambda_0)}$ such that the following holds: \begin{equation} s'= \sum_{\vecm\in\nbigt_1} \deldel^{\vecm_-}\bigl( g_{\vecm}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecm_+} x_1^{-1}\upsilon \bigr). \end{equation} Then, $g_{\vecm|\nbigy^{(\lambda_0)}}=0$ for any $\vecm\in\min\nbigt_1$. \end{lemma} \pf Take a primitive expression of $s'$ in $U_{\Ptilde}\nbigmtilde$: \[ s'=\sum_{\vecn\in\nbigt_2}\sum_{j=0}^{j_0(\vecn)} \deldel^{\vecn_-} \bigl(h_{\vecn,j}x^{-\vecdelta-\vecp+\vecn_+} (\tau f)^j\upsilon\bigr). \] Here, $\nbigt_2$ be a finite subset of $\seisuu^{\ell}$. We set $\nbigt_{2+}:=\bigl\{ \vecn\in\nbigt_2\,\big|\, n_1\geq 0 \bigr\}$ and $\nbigt_{2-}:=\bigl\{ \vecn\in\nbigt_2\,\big|\, n_1< 0 \bigr\}$. For $\vecn$, we set $\vecn_-':=\vecn_--(1,0,\ldots,0)$ and $\vecn_+':=\vecn_++(1,0,\ldots,0)$. Then, by using (), we obtain an expression as follows: \begin{multline} s'= \sum_{\vecn\in\nbigt_{2+}} \sum_{j=0}^{j_0(\vecn)} \deldel^{\vecn_-}\bigl( h_{\vecn,j}x^{-\vecdelta-\vecp+\vecn'_+}x_i^{-1} (\tau f)^j \upsilon \bigr) \\ +\sum_{\vecn\in\nbigt_{2-}} \sum_{j=0}^{j_0(\vecn)+1} \deldel^{\vecn'_-} \bigl( h^{(1)}_{\vecn,j}x^{-\vecdelta-\vecp+\vecn_+}x_1^{-1} (\tau f)^j\upsilon \bigr). \end{multline} Here, we can observe that $\deldel^{\vecn_-'} \bigl( h^{(1)}_{\vecn,j_0(\vecn)+1} x^{-\vecdelta-\vecp+\vecn_+}x_1^{-1} (\tau f)^{j_0(\vecn)+1}\upsilon \bigr)$ is primitive. We set $\nbigt'_2:= \{\vecn+(1,0,\ldots,0)\,|\,\vecn\in\nbigt_2\}$ and $\nbigt'_{2\pm}:= \{\vecn+(1,0,\ldots,0)\,|\,\vecn\in\nbigt_{2\pm}\}$. Take any $\vecm\in \min\bigl(\nbigt_1\cup\nbigt_2'\bigr)$. If $\vecm\in\nbigt'_{2-}$, i.e., $\vecm=\vecn+(1,0,\ldots,0)$ for $\vecn\in\nbigt_{2-}$, we obtain $h^{(1)}_{\vecn,j|\nbigy^{(\lambda_0)}}=0$ for any $j\in\seisuu_{> 0}$ by using (), () and Lemma . But, it contradicts with $h^{(1)}_{\vecn,j_0(\vecn)+1|\nbigy^{(\lambda_0)}}\neq 0$. Hence, we obtain $\vecm\not\in\nbigt_{2-}'$. Then, we obtain $\min(\nbigt_1) \subset \min\bigl(\nbigt_1\cup\nbigt_2'\bigr)$. Then, we obtain the claim of the lemma by using (), () and Lemma again. \qed \begin{corollary} Let $s$ be a section of $U_{\Ptilde+P_1}\nbigmtilde$ with an expression as in Lemma {\rm }: \begin{equation} s=s'+\sum_{\vecn\in\nbigt} \deldel^{\vecn_-}\bigl( g_{\vecn}(\lambda,\tau,x) x^{-\vecdelta-\vecp+\vecn_+}x_1^{-1} \upsilon \bigr). \end{equation} Then, the set $\min(\nbigt)$ and the functions $g_{\vecn|\nbigy^{(\lambda_0)}}$ $(\vecn\in\min(\nbigt))$ are well defined for $s$. \qed \end{corollary} We obtain the following corollary. \begin{corollary} $U_{\Ptilde+P_1}\nbigmtilde\big/ U_{\Ptilde}\nbigmtilde$ is strict. \end{corollary} \pf We have only to consider the issue locally around any point of $\nbigp^{(1)}_{\reduced}$ by using the notation above. Let $s$ be a section of $U_{\Ptilde+P_1}\nbigmtilde$ on $\nbign^{(\lambda_0)}$. Take an expression () of $s$ as in Lemma {\rm }. By Corollary , $s$ is a section of $U_{\Ptilde}\nbigmtilde$ if and only if $\nbigt=\emptyset$. Then, for any $\lambda_1\in\cnum$, the conditions for $s$ and $(\lambda-\lambda_1)s$ are equivalent. \qed \vspace{.1in} Note that we have $\tau U_{\Ptilde+P} \subset U_{\Ptilde}$. \begin{lemma} Let $s$ be a section of $U_{\Ptilde}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ with a primitive expression $s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}$. It is a section of $\tau U_{\Ptilde+P}\nbigmtilde$ if and only if $j\neq 0$ for any $(\vecm,j)\in\nbigs$. \end{lemma} \pf Note that the condition is independent of the choice of a primitive expression, by Corollary . Then, the claim is clear. \qed \begin{corollary} $U_{\Ptilde}\nbigmtilde\big/ \tau U_{\Ptilde+P}\nbigmtilde$ is strict. \qed \end{corollary} \subsubsection{The end of the proof of Theorem } Let us finish the proof of Theorem . It remains to prove that $U_{\alpha}\nbigmtilde/U_{<\alpha}\nbigmtilde$ is strict for any $\alpha\in\real$. We have only to prove the claim locally around any point of $\nbigd^{(1)}$ by using a coordinate system as in \S. It is enough to consider the case $0\leq\alpha<1$. Take any $\lambda_0\in\cnum_{\lambda}$. Note that $U_{[\alpha P]}\nbigmtilde=U'_{\alpha}\nbigmtilde$ for $0\leq \alpha\leq 1$. We have an effective divisor $P(<\alpha)$ for $0<\alpha\leq 1$ such that (i) $U_{P(<\alpha)}\nbigmtilde=U'_{<\alpha}\nbigmtilde$, (ii) $[\alpha P]-P(<\alpha)$ is effective. Hence, by using Corollary successively, we obtain that $U'_{\alpha}\nbigmtilde\big/U'_{<\alpha}\nbigmtilde$ is strict for $0<\alpha<1$. We have $U_{<0}\nbigmtilde=\tau U_{<1}\nbigmtilde \subset \tau U'_1\nbigmtilde \subset U_0\nbigmtilde$. We obtain that $(\tau U'_1\nbigmtilde)/U_{<0}\nbigmtilde$ is strict as above. By using Corollary , $U_0\nbigmtilde\big/(\tau U'_1\nbigmtilde)$ is strict. Thus, we are done. \qed \subsection{Proof of Theorem } \subsubsection{Quasi-isomorphisms} We generalize the results in \S in the context of mixed twistor $\nbigd$-modules. Let $\Omega^1_{\nbigx^{(1)}/\cnum^{2}_{\lambda,\tau}} (\log \nbigd^{(1)})$ denote the kernel of $\Omega^1_{\nbigx^{(1)}}(\log \nbigd^{(1)}) \lrarr \ptilde_1^{\ast} \Omega^1_{\cnum^2_{\lambda,\tau}}$. We set $\Omegatilde^1_{\nbigx^{(1)}/\cnum^{2}_{\lambda,\tau}} (\log \nbigd^{(1)}):= \lambda^{-1} \Omega^1_{\nbigx^{(1)}/\cnum^{2}_{\lambda,\tau}} (\log \nbigd^{(1)})$ and $\Omegatilde^k_{\nbigx^{(1)}/\cnum^{2}_{\lambda,\tau}} (\log \nbigd^{(1)}):= \bigwedge^k\Omegatilde^1_{\nbigx^{(1)}/\cnum^{2}_{\lambda,\tau}} (\log \nbigd^{(1)})$. Let $0\leq\alpha<1$. We set $\nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}):= \nbigmtilde_{[\alpha P]} \otimes\nbigo(-\nbigd^{(1)})$. We have an inclusion of complexes: \begin{equation} \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \lrarr U_{\alpha}\nbigmtilde \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} \end{equation} The following proposition is an analogue of Proposition , which we shall prove in \S. \begin{proposition} The morphism {\rm()} is a quasi-isomorphism. \end{proposition} The correspondence $1\longmapsto \upsilon$ induces a natural morphism of complexes: \begin{equation} \Omegatilde^{\bullet}_{f,\lambda,\tau}(\alpha) \lrarr \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}). \end{equation} We obtain the following proposition as in the case of Proposition , which we shall prove in \S. \begin{proposition} The morphism {\rm()} is a quasi-isomorphism. \end{proposition} We immediately obtain Theorem from Proposition and Proposition . \subsubsection{Proof of Proposition } We essentially repeat the argument in \S. We have only to check the claim around any point $(\lambda_0,\tau_0,Q)$ of $\nbigd^{(1)}$. We use a coordinate system as in \S. We set $\nbigd_i^{(1)}:=\{x_i=0\}$ on the coordinate neighbourhood. Set $\vecp:=[\alpha\veck]$. Let us consider the case $\tau_0\neq 0$. For $\ell_1\leq p\leq \ell$, we set $S(p):=\{\overbrace{(0,\ldots,0)}^{\ell_1}\} \times\seisuu^{p-\ell_1}_{\geq 0} \times\{\overbrace{(0,\ldots,0)}^{\ell-p}\} \subset\seisuu^{\ell_1}\times\seisuu^{p-\ell_1} \times\seisuu^{\ell-p} =\seisuu^{\ell}$. We consider \[ \nbigmtilde^{\leq p}:= \sum_{\vecn\in S(p)} \deldel^{\vecn} \nbigmtilde_{[\alpha P]}. \] For any $0\leq \alpha<1$, it is equal to the following in $\nbigo_{\nbigx^{(1)}}(\ast\nbigd^{(1)})$ around $(\lambda_0,\tau_0,Q)$: \[ \sum_{\vecn\in S(p)} \deldel^{\vecn} \bigl( \nbigo_{\nbigx^{(1)}}(\nbigd^{(1)}) (\ast\nbigp_{\reduced}^{(1)}) \upsilon \bigr) =\sum_{\vecn\in S(p)} \lambda^{|\vecn|} \nbigo_{\nbigx^{(1)}} \Bigl(\nbigd^{(1)} +\sum n_i\nbigd^{(1)}_i\Bigr) (\ast\nbigp_{\reduced}^{(1)}) \upsilon. \] We have $\nbigmtilde^{\leq \ell_1}=\nbigmtilde_{[\alpha P]}$. \begin{lemma} For $\ell_1+1\leq p\leq \ell$, the complexes $x_p\nbigmtilde^{\leq p-1} \stackrel{\deldel_p}{\lrarr} \nbigmtilde^{\leq p-1}$ and $\nbigmtilde^{\leq p} \stackrel{\deldel_p}{\lrarr} \nbigmtilde^{\leq p}$ are quasi-isomorphic with respect to the inclusion. \end{lemma} \pf We set $\nbigd^{(1)}_{<p}:= \bigcup_{i=1}^{p-1} \{x_i=0\}$ and $\nbigd^{(1)}_{\leq p}:= \bigcup_{i=1}^{p} \{x_i=0\}$ on the neighbourhood. For $j\in\seisuu$, we consider the following: \[ \lefttop{p}F_j\nbigl^{\leq p}:= \nbigo_{\nbigx^{(1)}}\bigl( \nbigd^{(1)}+j\nbigd_p^{(1)} \bigr)(\ast\nbigd_{< p}^{(1)}). \] We set $\nbigl^{\leq p}:= \bigcup_{j}\lefttop{p}F_j\nbigl^{\leq p} =\nbigo_{\nbigx^{(1)}}\bigl( \nbigd^{(1)} \bigr)(\ast\nbigd_{\leq p}^{(1)})$. We have $\lefttop{p}F_j\nbigl^{\leq p} \lrarr \lefttop{p}F_{j+1}\nbigl^{\leq p}$ induced by $\deldel_p$. If $j\geq 0$, the induced morphisms \[ \lefttop{p}F_j\nbigl^{\leq p} \big/\lefttop{p}F_{j-1}\nbigl^{\leq p} \lrarr \lefttop{p}F_{j+1}\nbigl^{\leq p} \big/\lefttop{p}F_{j}\nbigl^{\leq p} \] are injective. Indeed, the induced morphism $x_p\deldel_{p}$ on $\lefttop{p}F_j\nbigl^{\leq p} \big/\lefttop{p}F_{j-1}\nbigl^{\leq p}$ is the multiplication of $(j+1)\lambda$. By using the case $j=0$, we obtain that the kernel of $\lefttop{p}F_0 \lrarr \lefttop{p}F_1\big/\lefttop{p}F_0$ is $x_p\lefttop{p}F_0$. For $j\in\seisuu_{\geq 0}$, we consider the following: \[ \lefttop{p}F_j\nbigmtilde^{\leq p} :=\sum_{\substack{\vecn\in S(p)\\ n_p\leq j}} \deldel^{\vecn} \nbigmtilde_{[\alpha P]} =\sum_{\substack{\vecn\in S(p)\\ n_p\leq j}} \lambda^{|\vecn|} \nbigo_{\nbigx^{(1)}} \Bigl(\nbigd^{(1)} +\sum n_i\nbigd^{(1)}_i\Bigr)(\ast \nbigp_{\reduced}^{(1)}) \upsilon \] We have $\lefttop{p}F_0\nbigm^{\leq p} =\nbigm^{\leq p-1}$. We have $\lefttop{p}F_j\nbigmtilde^{\leq p} \subset \lefttop{p}F_j\nbigl^{\leq p}$ and $x_p\lefttop{p}F_0\nbigmtilde^{\leq p} \subset \lefttop{p}F_{-1}\nbigl^{\leq p}$. Let us consider the following commutative diagram: \[ \begin{CD} \lefttop{p}F_0\nbigmtilde^{\leq p}\big/ x_p\lefttop{p}F_0\nbigmtilde^{\leq p} @>{a_1}>> \lefttop{p}F_{0}\nbigltilde^{\leq p}\big/ \lefttop{p}F_{-1}\nbigltilde^{\leq p} \\ @V{b_1}VV @V{b_2}VV \\ \lefttop{p}F_j\nbigmtilde^{\leq p}\big/ \lefttop{p}F_{j-1}\nbigmtilde^{\leq p} @>{a_2}>> \lefttop{p}F_{j}\nbigltilde^{\leq p}\big/ \lefttop{p}F_{j-1}\nbigltilde^{\leq p}. \end{CD} \] Here, $a_i$ are induced by the inclusions, and $b_i$ are induced by $\deldel_p^j$. By using an explicit description, we can check that $a_1$ is a monomorphism. As mentioned, $b_2$ is a monomorphism. By the construction, $b_1$ is epimorphism. Hence, we obtain that $b_1$ is an isomorphism, and that $a_2$ is a monomorphism. We can deduce the following claims on the morphisms induced by $\deldel_p$: \begin{itemize} \item If $j\geq 1$, the induced morphisms $\lefttop{p}F_j\big/\lefttop{p}F_{j-1} \stackrel{\deldel_p}{\lrarr} \lefttop{p}F_{j+1}\big/\lefttop{p}F_{j}$ are isomorphisms. \item The kernel of $\lefttop{p}F_0\nbigmtilde^{\leq p} =\nbigmtilde^{\leq p-1} \lrarr \lefttop{p}F_1\big/\lefttop{p}F_0$ is $x_p\nbigmtilde^{\leq p-1}$. \end{itemize} Then, the claim of Lemma follows. \qed \vspace{.1in} For $\ell_1\leq p\leq \ell$, we set $\nbigh^{(1)}_{>p}:= \bigcup_{i=p+1}^{\ell}\{x_i=0\}$. By using Lemma , the following inclusions of the complexes of sheaves are quasi-isomorphisms (see the argument after Lemma ): \[ \nbigmtilde^{\leq p-1}(-\nbigh^{(1)}_{>p-1}) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigh^{(1)}_{>p-1}) \lrarr \nbigmtilde^{\leq p}(-\nbigh^{(1)}_{>p}) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigh^{(1)}_{>p}). \] We have $\nbigmtilde^{\leq \ell}(-\nbigh^{(1)}_{>\ell}) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigh^{(1)}_{>\ell}) =\nbigmtilde\otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}$, and \begin{multline} \nbigm^{\leq \ell_1}(-\nbigh^{(1)}_{>\ell_1}) \otimes \Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigh^{(1)}_{>\ell_1}) =\nbigmtilde_{[\alpha P]}(-\nbigh^{(1)}) \otimes\Omegatilde^{\bullet} _{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigh^{(1)}) \\ =\nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{\bullet}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}). \end{multline} Hence, we are done in the case $\tau_0\neq 0$. \vspace{.1in} Let us consider the case $\tau_0=0$. For $1\leq p\leq \ell$, we regard $\seisuu^{p}=\seisuu^{p}\times \{\overbrace{(0,\ldots,0)}^{\ell-p}\} \subset \seisuu^{\ell}$. We set \[ U_{\alpha}\nbigmtilde^{\leq p}:= \sum_{\vecn\in\seisuu^p_{\geq 0}} \deldel^{\vecn}\nbigmtilde_{[\alpha P]}, \quad\quad \lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p} =\sum_{\substack{ \vecn\in\seisuu^p_{\geq 0}\\ n_p\leq j }} \deldel^{\vecn}\nbigmtilde_{[\alpha P]}. \] We have $U_{\alpha}\nbigmtilde^{\leq \ell} =U_{\alpha}\nbigmtilde$ and $\lefttop{p}F_0U_{\alpha}\nbigmtilde^{\leq p} =U_{\alpha}\nbigmtilde^{\leq p-1}$. We consider the following maps: \[ \deldel_p: \lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p} \lrarr \lefttop{p}F_{j+1}U_{\alpha}\nbigmtilde^{\leq p}. \] The following lemma is easy to see. \begin{lemma} Let $s$ be a section of $U_{\alpha}\nbigmtilde$ on $\nbign^{(\lambda_0)}$ with a primitive expression $s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}$. Then, $s$ is a section of $\lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p}$ if and only if we have $m_i\geq 0$ $(i>p)$ and $m_p\geq -j$ for any $\vecm\in\min\pi(\nbigs)$. \qed \end{lemma} \begin{lemma} If $j\geq 1$, the following induced morphism of $\nbigo_{\nbigx^{(1)}}/x_p\nbigo_{\nbigx^{(1)}}$-modules is an isomorphism: \[ \frac{\lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p}} {\lefttop{p}F_{j-1}U_{\alpha}\nbigmtilde^{\leq p}} \stackrel{\deldel_p}{\lrarr} \frac{\lefttop{p}F_{j+1}U_{\alpha}\nbigmtilde^{\leq p}} {\lefttop{p}F_{j}U_{\alpha}\nbigmtilde^{\leq p}} \] \end{lemma} \pf It is surjective by construction. Let $s$ be a non-zero section of $\lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p}$ on $\nbign^{(\lambda_0)}$ with a primitive decomposition $s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}$ such that $\deldel_ps$ is also a section of $\lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p}$. We set $s':=\sum_{m_p=-j}s_{\vecm,j}$ and $s'':=\sum_{m_p>-j}s_{\vecm,j}$. Because $\deldel_ps''\in \lefttop{p}F_{j}U_{\alpha}\nbigmtilde^{\leq p}$, we obtain $\deldel_ps'\in\lefttop{p}F_jU_{\alpha}\nbigmtilde^{\leq p}$. If $s'$ is non-zero, $\deldel_ps'=\sum_{m_p=-j} \deldel_ps_{\vecm,j}$ is a primitive expression of $\deldel_ps'$. We obtain $\deldel_ps'\not\in \lefttop{p}F_{j}U_{\alpha}\nbigmtilde^{\leq p}$, and thus we arrive at a contradiction. Hence, $s'=0$, i.e., $s\in \lefttop{p}F_{j-1}U_{\alpha}\nbigmtilde^{\leq p}$. \qed \begin{lemma} The kernel of the following induced morphism is $x_pU_{\alpha}\nbigmtilde^{\leq p-1}$: \[ U_{\alpha}\nbigmtilde^{\leq p-1} \stackrel{\deldel_p}{\lrarr} \frac{\lefttop{p}F_1U_{\alpha}\nbigmtilde^{\leq p}} {U_{\alpha}\nbigmtilde^{\leq p-1}} \] \end{lemma} \pf Let $s$ be a section of $U_{\alpha}\nbigmtilde^{\leq p-1}$, and we take a primitive expression $s=\sum_{(\vecm,j)\in\nbigs}s_{\vecm,j}$ such that $\delbar_ps\in U_{\alpha}\nbigmtilde^{p-1}$. We set $s':=\sum_{m_p=0}s_{\vecm,j}$ and $s'':=\sum_{m_p>0}s_{\vecm,j}$. Because $s''\in x_pU_{\alpha}\nbigmtilde^{\leq p-1}$, we have $\deldel_ps''\in U_{\alpha}\nbigmtilde^{\leq p-1}$ as in the proof of Lemma . Hence, we have $\deldel_ps'\in U_{\alpha}\nbigmtilde^{\leq p-1}$. If $s'\neq 0$, $\deldel_ps'=\sum_{m_p=0}\deldel_ps_{\vecm,j}$ is a primitive expression of $s'$. We obtain $\deldel_ps'\not\in U_{\alpha}\nbigmtilde^{\leq p-1}$, and we arrive at a contradiction. Hence, we have $s'=0$, i.e., $s\in x_p U_{\alpha}\nbigm^{\leq p-1}$. \qed \vspace{.1in} By Lemma and Lemma , the inclusions of the complexes \[ \bigl( x_pU_{\alpha}\nbigmtilde^{\leq p-1} \stackrel{\deldel_p}{\lrarr} U_{\alpha}\nbigmtilde^{\leq p-1} \bigr) \lrarr \bigl( U_{\alpha}\nbigmtilde^{\leq p} \stackrel{\deldel_p}{\lrarr} U_{\alpha}\nbigmtilde^{\leq p} \bigr) \] are quasi-isomorphisms. For $0\leq p \leq \ell$, we set $\nbigd^{(1)}_{>p}:=\bigcup_{i=p+1}^{\ell}\{x_i=0\}$ on the neighbourhood. We obtain that the following inclusions of the complexes of sheaves are quasi-isomorphisms: \begin{equation} U_{\alpha}\nbigmtilde^{\leq p-1}(-\nbigd^{(1)}_{>p-1}) \otimes \Omegatilde_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}^{\bullet} (\log \nbigd^{(1)}_{>p-1}) \lrarr U_{\alpha}\nbigmtilde^{\leq p}(-\nbigd^{(1)}_{>p}) \otimes \Omegatilde_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}^{\bullet} (\log \nbigd^{(1)}_{>p}) \end{equation} We have $U_{\alpha}\nbigmtilde^{\leq \ell}(-\nbigd^{(1)}_{>\ell}) \otimes \Omegatilde_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}^{\bullet} (\log \nbigd^{(1)}_{>\ell}) =U_{\alpha}\nbigmtilde\otimes \Omegatilde_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}^{\bullet}$. We also have \[ U_{\alpha}\nbigmtilde^{\leq 0}(-\nbigd^{(1)}_{>0}) \otimes \Omegatilde_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}^{\bullet} (\log \nbigd^{(1)}_{>0}) =\nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}^{\bullet} (\log \nbigd^{(1)}). \] Hence, Proposition is proved. \qed \subsubsection{Proof of Proposition } We essentially repeat the argument in \S. We have only to check the claim around any point of $\nbigp_{\reduced}^{(1)}$. We use the coordinate system as in \S. We set $\vecp:=[\alpha\veck]$. For any non-negative integer $N$, we set \[ G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \bigr) :=\sum_{j=0}^N \nbigo_{\nbigx^{(1)}}x^{-\vecp} (\tau f)^j\upsilon. \] We define $G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$ as \[ G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \bigr) \otimes \Omegatilde^k_{\nbigx^{(1)}/ \cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}). \] We set $G_{-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log \nbigd^{(1)}) \bigr) :=\Omegatilde_{f,\lambda,\tau}^{k}(\alpha)\upsilon =\Omegatilde_{f,\lambda,\tau}^kx^{-\vecp}\upsilon$, where $\Omegatilde_{f,\lambda,\tau}^k:= \Omegatilde^k_{f,\lambda,\tau}(0)$. Let $N\geq 0$. Take a section \[ \omega =\sum_{j=0}^N \omega_j x^{-\vecp}(\tau f)^j\cdot \upsilon \in G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)})\otimes \Omegatilde^k _{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log\nbigd^{(1)}) \bigr), \] where $\omega_j\in \Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log \nbigd^{(1)})$. If $d\omega \in G_{N}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k+1} _{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log\nbigd^{(1)}) \bigr)$, then we have \[ \lambda^{-1}\tau df\wedge \omega_N\,(\tau f)^Nx^{-\vecp}\upsilon \in G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde^{k+1}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log \nbigd^{(1)}) \bigr). \] As in Lemma , we obtain $df\wedge\omega_N \in \Omegatilde^{k+1} _{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log \nbigd^{(1)})$, i.e., $\omega_N$ is a section of $\Omegatilde^k_{f,\lambda,\tau}$. \begin{lemma} We have an expression \[ \omega_N= (df/f)\wedge \kappa_1 +f^{-1}\kappa_2, \] where $\kappa_1$ and $\kappa_2$ are sections of $\Omegatilde^{k-1}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log\nbigd^{(1)})$ and $\Omegatilde^{k}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)})$, respectively. \qed \end{lemma} If $N\geq 1$, we have $f^{-1}\kappa_2(\tau f)^N =\tau \kappa_2(\tau f)^{N-1}$. Hence, we have \[ \omega-d\bigl(\kappa_1(\tau f)^{N-1}x^{-\vecp}\upsilon\bigr) \in G_{N-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr). \] We also have $f^{-1}\kappa_1(\tau f)^Nx^{-\vecp}\upsilon \in G_{N-1}$. Let $\omega$ be a local section of $G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$ such that $d\omega$ is a local section of $G_{-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$. Then, by applying the previous argument successively, we can find a local section $\tau$ of $G_{N-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k-1}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$ such that $\omega-d\tau$ is a local section of $G_{-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log\nbigd^{(1)}) \bigr)$. \begin{itemize} \item If a local section $\omega$ of $G_N\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^k_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$ satisfies $d\omega=0$, we can find a local section $\tau$ of $G_{N-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k-1}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$ such that $\omega-d\tau$ is a local section of $G_{-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log\nbigd^{(1)}) \bigr)$. \item Let $\omega$ be a local section of $G_{-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log \nbigd^{(1)}) \bigr)$ such that $d\omega=0$. Suppose that we have a local section $\tau$ of $G_{N}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde^{k-1}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log\nbigd^{(1)}) \bigr)$ such that $\omega=d\tau$. Then, we can find a local section $\sigma$ of $G_{N-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes\Omegatilde^{k-2}_{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}} (\log\nbigd^{(1)}) \bigr)$ such that $\tau-d\sigma$ is a local section of $G_{-1}\bigl( \nbigmtilde_{[\alpha P]}(-\nbigd^{(1)}) \otimes \Omegatilde^{k-1} _{\nbigx^{(1)}/\cnum^2_{\lambda,\tau}}(\log\nbigd^{(1)}) \bigr)$. We have $\omega=d(\tau-d\sigma)$. \end{itemize} Then, we obtain the claim of Proposition . The proof of Theorem is also completed. \qed \begin{thebibliography}{99} \bibitem{Esnault-Sabbah-Yu} H. Esnault, C. Sabbah, J.-D. Yu, (with an appendix by M. Saito), {\em $E_1$-degeneration of the irregular Hodge filtration}, J. reine angew. Math. (2015), doi:10.1515/crelle-2014-0118, \bibitem{kashiwara_text} M. Kashiwara, {\em $D$-modules and microlocal calculus,} Translations of Mathematical Monographs, {\bf 217}. American Mathematical Society, Providence, (2003). \bibitem{Katzarkov-Kontsevich-Pantev-2017} L. Katzarkov, M. Kontsevich, T. Pantev, {\em Bogomolov-Tian-Todorov theorems for Landau-Ginzburg models}, J. Differential Geom. 105 (2017), no. 1, 55--117. \bibitem{mochi2} T. Mochizuki, {\em Asymptotic behaviour of tame harmonic bundles and an application to pure twistor $D$-modules I, II}, Mem. AMS. {\bf 185}, (2007). \bibitem{Mochizuki-wild} T. Mochizuki, {\em Wild harmonic bundles and wild pure twistor $D$-modules}, Ast\'{e}risque {\bf 340}, Soci\'{e}t\'{e} Math\'{e}matique de France, Paris, 2011. \bibitem{Mochizuki-Toda-lattice} T. Mochizuki, {\em Harmonic bundles and Toda lattices with opposite sign II}, Comm. Math. Phys. {\bf 328}, 1159--1198, DOI:10.1007/s00220-014-1994-0. \bibitem{Mochizuki-MTM} T. Mochizuki, {\em Mixed twistor $D$-modules}, Lecture Notes in Mathematics, {\bf 2125}. Springer, 2015. \bibitem{sabbah2} C. Sabbah, {\em Polarizable twistor $D$-modules} Ast\'{e}risque, {\bf 300}, (2005) \bibitem{sabbah5} C. Sabbah, {\em Wild twistor $D$-modules}, in {\em Algebraic analysis and around}, Adv. Stud. Pure Math., {\bf 54}, Math. Soc. Japan, Tokyo, (2009), 293--353. \bibitem{Sabbah-Yu} C. Sabbah, J.-D. Yu. {\em On the irregular Hodge filtration of exponentially twisted mixed Hodge modules}, Forum Math. Sigma {\bf 3} (2015), 71 pp. \bibitem{saito1} M. Saito, {\em Modules de Hodge polarisables}, Publ. RIMS., {\bf 24}, (1988), 849--995. \bibitem{saito2} M. Saito, {\em Mixed Hodge modules}, Publ. RIMS., {\bf 26}, (1990), 221--333. \bibitem{Yu} J.-D. Yu, {\em Irregular Hodge filtration on twisted de Rham cohomology}, Manuscripta Math. 144(1–2) (2014), 99--133. \end{thebibliography} \title{A twistor approach to the Kontsevich complexes } \author{Takuro Mochizuki} \date{} \maketitle \begin{abstract} We study the $V$-filtration of the mixed twistor $\nbigd$-modules associated to algebraic meromorphic functions. We prove that their relative de Rham complexes are quasi-isomorphic to the family of Kontsevich complexes. It reveals a generalized Hodge theoretic meaning of Kontsevich complexes. On the basis of the quasi-isomorphism, we revisit the results on the Kontsevich complexes due to H. Esnault, M. Kontsevich, C. Sabbah, M. Saito and J.-D. Yu from a viewpoint of mixed twistor $\nbigd$-modules. \vspace{.1in} \noindent Keywords: Mixed twistor $\nbigd$-module, Kontsevich complex, $V$-filtration. \vspace{.1in} \noindent 14F10, 32C38, 32S35. \end{abstract} \input{10} \input{computation_ref} \vspace{.1in} \noindent Address: Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-8502, Japan\\ Email: takuro@kurims.kyoto-u.ac.jp \newcounter{goodsection} \newtheorem{theorem}{Theorem}[section] \newtheorem{corollary}[theorem]{Corollary} \newtheorem{conj}[theorem]{Conjecture} \newtheorem{remark}[theorem]{Remark} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{definition}[theorem]{Definition} \newtheorem{problem}[theorem]{Problem} \newtheorem{condition}[theorem]{Condition} \newtheorem{assumption}[theorem]{Assumption} \newtheorem{principle}[theorem]{Principle} \newtheorem{notation}[theorem]{Notation} \newtheorem{convention}[theorem]{Convention} \newcommand{\nbiga}{\mathcal{A}} \newcommand{\nbigb}{\mathcal{B}} \newcommand{\nbigc}{\mathcal{C}} \newcommand{\nbigd}{\mathcal{D}} \newcommand{\nbige}{\mathcal{E}} \newcommand{\nbigf}{\mathcal{F}} \newcommand{\nbigg}{\mathcal{G}} \newcommand{\nbigh}{\mathcal{H}} \newcommand{\nbigi}{\mathcal{I}} \newcommand{\nbigj}{\mathcal{J}} \newcommand{\nbigk}{\mathcal{K}} \newcommand{\nbigl}{\mathcal{L}} \newcommand{\nbigm}{\mathcal{M}} \newcommand{\nbign}{\mathcal{N}} \newcommand{\nbigo}{\mathcal{O}} \newcommand{\nbigp}{\mathcal{P}} \newcommand{\nbigq}{\mathcal{Q}} \newcommand{\nbigr}{\mathcal{R}} \newcommand{\nbigs}{\mathcal{S}} \newcommand{\nbigt}{\mathcal{T}} \newcommand{\nbigu}{\mathcal{U}} \newcommand{\nbigv}{\mathcal{V}} \newcommand{\nbigw}{\mathcal{W}} \newcommand{\nbigx}{\mathcal{X}} \newcommand{\nbigy}{\mathcal{Y}} \newcommand{\nbigz}{\mathcal{Z}} \newcommand{\nbiggr}{\mathcal{GR}} \newcommand{\Epsilon}{\nbige} \newcommand{\pb}[1]{{#1}^{\ast}} \newcommand{\proj}{\mathbb{P}} \newcommand{\seisuu}{{\mathbb Z}} \newcommand{\rnum}{{\mathbb Q}} \newcommand{\nnum}{{\mathbb N}} \newcommand{\cnum}{{\mathbb C}} \newcommand{\real}{{\mathbb R}} \newcommand{\hyperh}{\mathbb{H}} \newcommand{\hyperr}{\mathbb{R}} \newcommand{\iti}{\mathbb{I}} \newcommand{\gm}{\mathbb{G}_m} \newcommand{\Tate}{\mathbb{T}} \newcommand{\newTate}{\pmb{T}} \newcommand{\DD}{\mathbb{D}} \newcommand{\EE}{\mathbb{E}} \newcommand{\FF}{\mathbb{F}} \newcommand{\GG}{\mathbb{G}} \newcommand{\gbiga}{\mathfrak A} \newcommand{\gbigb}{\mathfrak B} \newcommand{\gbigc}{\mathfrak C} \newcommand{\gbigd}{\mathfrak D} \newcommand{\gbige}{\mathfrak E} \newcommand{\gbigf}{\mathfrak F} \newcommand{\gbigg}{\mathfrak G} \newcommand{\gbigh}{\mathfrak H} \newcommand{\gbigi}{\mathfrak I} \newcommand{\gbigj}{\mathfrak J} \newcommand{\gbigk}{\mathfrak K} \newcommand{\gbigl}{\mathfrak L} \newcommand{\gbigm}{\mathfrak M} \newcommand{\gbign}{\mathfrak N} \newcommand{\gbigo}{\mathfrak O} \newcommand{\gbigp}{\mathfrak P} \newcommand{\gbigq}{\mathfrak Q} \newcommand{\gbigr}{\mathfrak R} \newcommand{\gbigs}{\mathfrak S} \newcommand{\gbigt}{\mathfrak T} \newcommand{\gbigu}{\mathfrak U} \newcommand{\gbigv}{\mathfrak V} \newcommand{\gbigw}{\mathfrak W} \newcommand{\gbigx}{\mathfrak X} \newcommand{\gbigy}{\mathfrak Y} \newcommand{\gbigz}{\mathfrak Z} \newcommand{\gminia}{\mathfrak a} \newcommand{\gminib}{\mathfrak b} \newcommand{\gminic}{\mathfrak c} \newcommand{\gminid}{\mathfrak d} \newcommand{\gminie}{\mathfrak e} \newcommand{\gminif}{\mathfrak f} \newcommand{\gminig}{\mathfrak g} \newcommand{\gminih}{\mathfrak h} \newcommand{\gminii}{\mathfrak i} \newcommand{\gminij}{\mathfrak j} \newcommand{\gminik}{\mathfrak k} \newcommand{\gminil}{\mathfrak l} \newcommand{\gminim}{\mathfrak m} \newcommand{\gminin}{\mathfrak n} \newcommand{\gminio}{\mathfrak o} \newcommand{\gminip}{\mathfrak p} \newcommand{\gminiq}{\mathfrak q} \newcommand{\gminir}{\mathfrak r} \newcommand{\gminis}{\mathfrak s} \newcommand{\gminit}{\mathfrak t} \newcommand{\gminiu}{\mathfrak u} \newcommand{\gminiv}{\mathfrak v} \newcommand{\gminiw}{\mathfrak w} \newcommand{\gminix}{\mathfrak x} \newcommand{\gminiy}{\mathfrak y} \newcommand{\gminiz}{\mathfrak z} \newcommand{\vexi}{{\boldsymbol \xi}} \newcommand{\veeta}{{\boldsymbol \eta}} \newcommand{\vecxi}{{\boldsymbol \xi}} \newcommand{\veceta}{{\boldsymbol \eta}} \newcommand{\vecrho}{{\boldsymbol \rho}} \newcommand{\vece}{{\boldsymbol e}} \newcommand{\vecr}{{\boldsymbol r}} \newcommand{\vecv}{{\boldsymbol v}} \newcommand{\vecu}{{\boldsymbol u}} \newcommand{\vecw}{{\boldsymbol w}} \newcommand{\vecgamma}{{\boldsymbol \gamma}} \newcommand{\vecl}{{\boldsymbol l}} \newcommand{\veczero}{{\boldsymbol 0}} \newcommand{\vecalpha}{{\boldsymbol \alpha}} \newcommand{\veca}{{\boldsymbol a}} \newcommand{\vecb}{{\boldsymbol b}} \newcommand{\vecbeta}{{\boldsymbol \beta}} \newcommand{\vecdelta}{{\boldsymbol \delta}} \newcommand{\vecs}{{\boldsymbol s}} \newcommand{\vect}{{\boldsymbol t}} \newcommand{\vecc}{{\boldsymbol c}} \newcommand{\vecd}{{\boldsymbol d}} \newcommand{\vech}{{\boldsymbol h}} \newcommand{\veck}{{\boldsymbol k}} \newcommand{\vecm}{{\boldsymbol m}} \newcommand{\vecM}{{\boldsymbol M}} \newcommand{\vecN}{{\boldsymbol N}} \newcommand{\vecI}{{\boldsymbol I}} \newcommand{\vecomega}{{\boldsymbol \omega}} \newcommand{\vecx}{{\boldsymbol x}} \newcommand{\vecf}{{\boldsymbol f}} \newcommand{\vecepsilon}{{\boldsymbol \epsilon}} \newcommand{\vecF}{{\boldsymbol F}} \newcommand{\vecE}{{\boldsymbol E}} \newcommand{\vecW}{{\boldsymbol W}} \newcommand{\vecn}{{\boldsymbol n}} \newcommand{\vecp}{{\boldsymbol p}} \newcommand{\veczeta}{{\boldsymbol \zeta}} \newcommand{\vecz}{{\boldsymbol z}} \newcommand{\vecC}{{\boldsymbol C}} \newcommand{\vecU}{{\boldsymbol U}} \newcommand{\vecV}{{\boldsymbol V}} \newcommand{\vecA}{{\boldsymbol A}} \newcommand{\vecB}{{\boldsymbol B}} \newcommand{\vecS}{{\boldsymbol S}} \newcommand{\dual}{\lor} \newcommand{\larr}{\leftarrow} \newcommand{\llarr}{\longleftarrow} \newcommand{\uarr}{\uparrow} \newcommand{\Larr}{\Longleftarrow} \newcommand{\Llarr}{\Longleftarrow} \newcommand{\Uarr}{\Uparrow} \newcommand{\rarr}{\rightarrow} \newcommand{\lrarr}{\longrightarrow} \newcommand{\darr}{\downarrow} \newcommand{\Rarr}{\Rightarrow} \newcommand{\Lrarr}{\Longrightarrow} \newcommand{\Darr}{\Downarrow} \newcommand{\barr}{\leftrightarrow} \newcommand{\lbarr}{\longleftrightarrow} \newcommand{\udarr}{\updownarrow} \newcommand{\Barr}{\Leftrightarrow} \newcommand{\Lbarr}{\Longleftrightarrow} \newcommand{\Udarr}{\Updownarrow} \newcommand{\lmapsto}{\longmapsto} \newcommand{\hlarr}{\hookleftarrow} \newcommand{\hrarr}{\hookrightarrow} \newcommand{\lhrarr}{ \mbox{ \begin{picture}(2,8)(0,-2) \put(2,2){\oval(4,4)[l]} \end{picture}\rule[.65mm]{6mm}{.1mm}\raisebox{-.1mm}{$\rightarrow$}}} \newcommand{\hdarr} {\mbox{$ \begin{array}{r} \raisebox{-.6ex} {\begin{picture}(6,2)(-4,0) \put(-2,0){\oval(4,4)[t]} \put(0,0){\line(0,-1){8}} \end{picture}}\hspace{.23mm}\\ \rule[-1.5mm]{.1mm}{3mm}\hspace{.9mm}\\ \raisebox{.6ex}{$\downarrow$} \end{array}$}} \newcommand{\pf}{{\bf Proof}\hspace{.1in}} \newcommand{\qed}{\hfill\mbox{\rule{1.2mm}{3mm}}} \def\Hom{\mathop{\rm Hom}\nolimits} \def\poly{\mathop{\rm poly}\nolimits} \def\End{\mathop{\rm End}\nolimits} \def\Ext{\mathop{\rm Ext}\nolimits} \def\Extbar{\underline{\mathop{\rm Ext}\nolimits}} \def\Cok{\mathop{\rm Cok}\nolimits} \def\cok{\mathop{\rm Cok}\nolimits} \def\Image{\mathop{\rm Im}\nolimits} \def\Realpart{\mathop{\rm Re}\nolimits} \def\realpart{\mathop{\rm Re}\nolimits} \def\Re{\mathop{\rm Re}\nolimits} \def\CH{\mathop{\rm CH}\nolimits} \def\Gr{\mathop{\rm Gr}\nolimits} \def\gr{\mathop{\rm Gr}\nolimits} \def\GL{\mathop{\rm GL}\nolimits} \def\SL{\mathop{\rm SL}\nolimits} \def\Tot{\mathop{\rm Tot}\nolimits} \def\Cone{\mathop{\rm Cone}\nolimits} \def\rank{\mathop{\rm rank}\nolimits} \def\Spec{\mathop{\rm Spec}\nolimits} \def\Nu{{\mathcal V}} \def\RHom{\mathop{\rm RHom}\nolimits} \def\Rhombar{\underline{\mathop{\rm RHom}\nolimits}} \def\Hombar{\underline{\mathop{\rm Hom}\nolimits}} \def\trdeg{\mathop{\rm tr.deg}\nolimits} \def\Ker{\mathop{\rm Ker}\nolimits} \def\ker{\mathop{\rm Ker}\nolimits} \def\modulo{\mathop{\rm modulo}\nolimits} \def\length{\mathop{\rm length}\nolimits} \def\Gr{\mathop{\rm Gr}\nolimits} \def\Sym{\mathop{\rm Sym}\nolimits} \def\sym{\mathop{\rm sym}\nolimits} \def\ad{\mathop{\rm ad}\nolimits} \def\Res{\mathop{\rm Res}\nolimits} \def\Rep{\mathop{\rm Rep}\nolimits} \def\ord{\mathop{\rm ord}\nolimits} \def\degpar{\mathop{\rm par\textrm{-}deg}\nolimits} \def\parch{\mathop{\rm par\textrm{-}ch}\nolimits} \def\ch{\mathop{ch}\nolimits} \def\parchern{\mathop{\rm par\textrm{-}c}\nolimits} \def\degnil{\mathop{\rm nil\textrm{-}deg}\nolimits} \def\Ric{\mathop{\rm Ric}\nolimits} \def\tr{\mathop{\rm tr}\nolimits} \def\Tr{\mathop{\rm Tr}\nolimits} \def\vol{\mathop{\rm dvol}\nolimits} \def\dvol{\mathop{\rm dvol}\nolimits} \def\Diff{\mathop{\rm Diff}\nolimits} \def\vecdeg{\mathop{\rm \bf deg}\nolimits} \def\vecord{\mathop{\rm \bf ord}\nolimits} \def\can{\mathop{\rm can}\nolimits} \def\var{\mathop{\rm var}\nolimits} \def\id{\mathop{\rm id}\nolimits} \def\Pat{\mathop{\rm Pat}\nolimits} \def\Prim{\mathop{\rm Prim}\nolimits} \def\Red{\mathop{\rm Red}\nolimits} \def\gcd{\mathop{\rm g.c.d.}\nolimits} \def\codim{\mathop{\rm codim}\nolimits} \def\gap{\mathop{\rm gap}\nolimits} \def\reduced{\mathop{\rm red}\nolimits} \def\filt{\mathop{\rm Filt}\nolimits} \def\equi{\mathop{\rm Equi}\nolimits} \def\Supp{\mathop{\rm Supp}\nolimits} \def\VPTgen{\mathop{\rm VPT_{gen}}\nolimits} \def\VPTgenwild{\mathop{\rm VPT_{gen}^{\wild}}\nolimits} \def\MPT{\mathop{\rm MPT}\nolimits} \def\RHD{\mathop{\rm RHD}\nolimits} \def\ch{\mathop{\rm ch}\nolimits} \def\Irrep{\mathop{\rm Irrep}\nolimits} \def\Irr{\mathop{\rm Irr}\nolimits} \def\Dom{\mathop{\rm Dom}\nolimits} \def\Zar{\mathop{\rm Zar}\nolimits} \def\wt{\mathop{\rm wt}\nolimits} \def\diag{\mathop{\rm diag}\nolimits} \def\Lef{\mathop{\rm Lef}\nolimits} \def\Obv{\mathop{\rm Obv}\nolimits} \newcommand{\Cpoly}[1]{\nbigc^{poly}_{#1}} \newcommand{\hh}[1]{{\rm H}^{#1}} \newcommand{\nbighh}[1]{\nbigh ^{#1}} \newcommand{\hhbar}[1]{\underbar{{\rm H}}^{#1}} \newcommand{\tentimes}{\otimes \cdots \otimes} \newcommand{\tenwedge}{\wedge \cdots \wedge} \newcommand{\del}{\partial} \newcommand{\delbar}{\overline{\del}} \newcommand{\bardel}{\overline{\del}} \newcommand{\delbartilde}{\widetilde{\delbar}} \newcommand{\deltilde}{\widetilde{\del}} \newcommand{\dtilde}{\widetilde{d}} \newcommand{\deltatilde}{\widetilde{\delta}} \newcommand{\deltabar}{(\Delta ^{l\,\circ})_{-}} \newcommand{\Deltabar}{\overline{\Delta}} \newcommand{\Deltabarast}{\Deltabar^{\ast}} \newcommand{\pardeg}{\degpar} \newcommand{\nildeg}{\degnil} \newcommand{\nhom}{{\mathcal Hom}} \newcommand{\ntor}{{\mathcal Tor}} \newcommand{\nend}{{\mathcal End}} \newcommand{\next}{{\mathcal Ext}} \newcommand{\hsdgla}[1]{\Omega (\gminig_{#1} ^{\cdot})^{\cdot}} \newcommand{\hsdglai}[2]{\Omega (\gminig_{#1} ^{\cdot})^{#2}} \newcommand{\vecP}{\bf P} \newcommand{\barh}{\hbar} \newcommand{\tpoly}{T_{poly}} \newcommand{\dpoly}{D_{poly}} \newcommand{\tpolybar}{\overline{T}_{poly}} \newcommand{\dpolybar}{\overline{D}_{poly}} \newcommand{\gaisekig}{\bigwedge^{\cdot}\gminig} \newcommand{\gaisekigg}{\bigwedge^{\cdot}\gminig^{\lor}} \newcommand{\gaisekih}{\bigwedge^{\cdot}\gminih} \newcommand{\gaisekihh}{\bigwedge^{\cdot}\gminih^{\bot}} \newcommand{\iremonot}{T_{poly}(U)\otimes\gaisekigg} \newcommand{\iremonod}{D_{poly}(U)\otimes\gaisekigg} \newcommand{\aaa}{\alpha} \newcommand{\ttt}{\theta} \newcommand{\TTT}{\Theta} \newcommand{\nbar}{\underline{n}} \newcommand{\dbar}{\underline{d}} \newcommand{\jbar}{\underline{j}} \newcommand{\mbar}{\underline{m}} \newcommand{\kbar}{\underline{k}} \newcommand{\jitibar}{\underline{j+1}} \newcommand{\jminusitibar}{\underline{j-1}} \newcommand{\mitibar}{\underline{m+1}} \newcommand{\mzerobar}{\underline{m_0}} \newcommand{\mzeroijbar}{\underline{m_0(i,j)}} \newcommand{\mitiitibar}{\underline{m_1+1}} \newcommand{\mbariti}{\underline{m_1}} \newcommand{\mminusitibar}{\underline{m-1}} \newcommand{\lbariti}{\underline{l_1}} \newcommand{\litibar}{\underline{l-1}} \newcommand{\ibar}{\underline{i}} \newcommand{\iitibar}{\underline{i+1}} \newcommand{\nminusitibar}{\underline{n-1}} \newcommand{\nminusnibar}{\underline{n-2}} \newcommand{\lminusitibar}{\underline{l-1}} \newcommand{\lbar}{\underline{l}} \newcommand{\pbar}{\underline{p}} \newcommand{\itibar}{\underline{1}} \newcommand{\nibar}{\underline{2}} \newcommand{\zerobar}{\underline{0}} \newcommand{\alphasitabar}{\underline{\alpha}} \newcommand{\shikaku}{\sharp} \newcommand{\sankaku}{\triangle} \newcommand{\twoprime}{\prime\prime} \newcommand{\harmonicbundle}{(E,\delbar_E,\theta,h)} \newcommand{\harmonicbundledual}{(E^{\lor},\delbar_{E^{\lor}},\theta^{\lor},h^{\lor})} \newcommand{\conjugate}{\clubsuit} \newcommand{\barz}{\overline{z}} \newcommand{\zbar}{\barz} \newcommand{\zetabar}{\overline{\zeta}} \newcommand{\baralpha}{\overline{\alpha}} \newcommand{\alphabar}{\baralpha} \newcommand{\barlambda}{\overline{\lambda}} \newcommand{\lambdabar}{\barlambda} \newcommand{\fbar}{\overline{f}} \newcommand{\varphibar}{\overline{\varphi}} \newcommand{\etabar}{\overline{\eta}} \newcommand{\nablabar}{\overline{\nabla}} \newcommand{\tbar}{\overline{t}} \newcommand{\xbar}{\overline{x}} \newcommand{\sbar}{\overline{s}} \newcommand{\Hbar}{\overline{H}} \newcommand{\Hbarepsilon}{\overline{H}_{\epsilon}} \newcommand{\Abar}{\overline{A}} \newcommand{\modelbundle}[2]{E({#1},{#2})} \newcommand{\modeldeform}{{\mathcal Mod}} \newcommand{\transmatrix}{B} \newcommand{\volume}{\Omega} \newcommand{\dotvecw}{\dot{\vecw}} \newcommand{\vecwdot}{\dotvecw} \newcommand{\dotF}{\dot{F}} \newcommand{\dotvece}{\dot{\vece}} \newcommand{\vecedot}{\dotvece} \newcommand{\dote}{\dot{e}} \newcommand{\dotw}{\dot{w}} \newcommand{\dotPhi}{\dot{\Phi}} \newcommand{\doth}{\dot{h}} \newcommand{\dottheta}{\dot{\theta}} \newcommand{\dotn}{\dot{n}} \newcommand{\dotdelbar}{\dot{\delbar}} \newcommand{\dotvecv}{\dot{\vecv}} \newcommand{\dotv}{\dot{v}} \newcommand{\soeji}{I} \newcommand{\Poin}{{\bf p}} \newcommand{\poin}{\Poin} \newcommand{\prolong}[1]{{}^{\diamond}{#1}} \newcommand{\prolongg}[2]{{}_{#1}{#2}} \newcommand{\leftupdown}[2]{{}^{#1}_{#2}} \newcommand{\resddlambda}{\Res(\DD^{\lambda})} \newcommand{\residuei}{\Res_{\nbigd_i}(\DD)} \newcommand{\residueidagger}{\Res_{\nbigd_i^{\dagger}}(\DD^{\dagger})} \newcommand{\residueiflat}{\Res_{\nbigd_i}(\DD^f)} \newcommand{\residueiflatdagger}{\Res_{\nbigd^{\dagger}_i}(\DD^{\dagger\,f})} \newcommand{\DDlambda}{\DD^{\lambda}} \newcommand{\DDlambdastar}{\DD^{\lambda\,\star}} \newcommand{\DDlambdahat}{\widehat{\DD}^{\lambda}} \newcommand{\DDhatlambda}{\DDlambdahat} \newcommand{\DDlambdaf}{\DD^{\lambda\,f}} \newcommand{\DDlambdaast}{\DD^{\lambda\,\ast}} \newcommand{\weightfilt}{W} \newcommand{\laplacian}{\Delta''} \newcommand{\Deltalambda}{\Delta^{\lambda}} \newcommand{\doublelangle}{\langle\langle} \newcommand{\doublerangle}{\rangle\rangle} \newcommand{\doublebiglangle}{\big\langle\big\langle} \newcommand{\doublebigrangle}{\big\rangle\big\rangle} \newcommand{\graded}{{\mathcal Gr}} \newcommand{\deformoverorigin}{\prolong{\nbige}_{|\cnum_{\lambda}\times O}} \newcommand{\deformoverorigindagger}{\prolong{\nbige^{\dagger}}_{|\cnum_{\mu}\times O}} \newcommand{\nbigelambda}{\nbige^{\lambda}} \newcommand{\nbigelambdazero}{\nbige^{\lambda_0}} \newcommand{\nbigelor}{\nbige^{\lor}} \newcommand{\tiisai}{\nbigu} \newcommand{\blowup}{\widetilde} \newcommand{\projection}{\gminiq} \newcommand{\directsummand}{\nbigu} \newcommand{\thetainfty}{\theta^{(\infty)}} \newcommand{\hinfty}{h^{(\infty)}} \newcommand{\superinfty}{^{(\infty)}} \newcommand{\partialkoszul}{\Pi(N_1,\ldots,N_n)} \newcommand{\partialkoszulinfty}{\Pi(N_1\superinfty,\ldots,N_n\superinfty)} \newcommand{\gradedrestrictedtoni}{\graded_{b|\nbigd_{\nibar}}^{(1)}} \newcommand{\nbigxlambda}{\nbigx^{\lambda}} \newcommand{\nbigxlambdazero}{\nbigx^{\lambda_0}} \newcommand{\nbigdlambda}{\nbigd^{\lambda}} \newcommand{\KMS}{{\mathcal{KMS}}} \newcommand{\EKMS}{{\mathcal{EKMS}}} \newcommand{\KMSoverline}{\overline{\mathcal{KMS}}} \newcommand{\KMSE}[1]{\KMS(\nbige^{#1})} \newcommand{\KMSEoverline}[1]{\overline{\KMS}(\nbige^{#1})} \newcommand{\KMSEprolongg}[2]{\KMS(\prolongg{#1}{\nbige^{#2}})} \newcommand{\Par}{{\mathcal Par}} \newcommand{\Sp}{{\mathcal Sp}} \newcommand{\ParE}[1]{{\mathcal Par}(\nbige^{#1})} \newcommand{\SpE}[1]{{\mathcal Sp}(\nbige^{#1})} \newcommand{\ParEprolongg}[2]{{\mathcal Par}(\prolongg{#1}{\nbige}^{#2})} \newcommand{\SpEprolongg}[2]{{\mathcal Sp}(\prolongg{#1}{\nbige}^{#2})} \newcommand{\kmsmap}{\gminik} \newcommand{\paramap}{\gminip} \newcommand{\eigenmap}{\gminie} \newcommand{\multiplicity}{\gminim} \newcommand{\twistmap}{\gminit} \newcommand{\lefttop}[1]{{}^{#1}\!} \newcommand{\leftbottom}[1]{{}_{#1}\!} \def\Filt{\mathop{\rm Filt}\nolimits} \def\Equi{\mathop{\rm Equi}\nolimits} \def\Bifilt{\mathop{\rm Bifilt}\nolimits} \def\irr{\mathop{\rm irr}\nolimits} \def\reg{\mathop{\rm reg}\nolimits} \def\Def{\mathop{\rm Def}\nolimits} \def\Harm{\mathop{\rm Harm}\nolimits} \def\nil{\mathop{\rm nil}\nolimits} \def\Nil{\mathop{\rm Nil}\nolimits} \newcommand{\FEzero}{(F^{(\lambda_0)},\EE^{(\lambda_0)})} \newcommand{\Fzero}{F^{(\lambda_0)}} \newcommand{\EEzero}{\EE^{(\lambda_0)}} \newcommand{\FFzero}{\FF^{(\lambda_0)}} \newcommand{\GGzero}{\GG^{(\lambda_0)}} \newcommand{\Vzero}{V^{(\lambda_0)}} \newcommand{\Vzeroprime}{V^{\prime\,(\lambda_0)}} \newcommand{\psizero}{\psi^{(\lambda_0)}} \newcommand{\tildepsizero}{\widetilde{\psi}^{(\lambda_0)}} \newcommand{\tildepsi}{\widetilde{\psi}} \newcommand{\psitilde}{\tildepsi} \newcommand{\nbigvzero}{\nbigv^{(\lambda_0)}} \newcommand{\nbigfzero}{\nbigf^{(\lambda_0)}} \newcommand{\nbigqzero}{\nbigq^{(\lambda_0)}} \newcommand{\Ulambdazero}{U^{(\lambda_0)}} \newcommand{\Uzero}{U^{(\lambda_0)}} \newcommand{\Psizero}{\Psi^{(\lambda_0)}} \newcommand{\deldel}{\eth} \newcommand{\deldelbar}{\overline{\deldel}} \newcommand{\lamda}{\lambda} \newcommand{\lambdazero}{\lambda_{0}} \newcommand{\lamdazero}{\lamda_0} \newcommand{\kataim}{\gbigm} \newcommand{\AAA}{{\boldsymbol A}} \newcommand{\distribution}{\gbigd\gminib} \newcommand{\naiveprolong}[1]{\lefttop{\square}{#1}} \newcommand{\naiveprolongg}[2]{{}^{\square}_{#1}{#2}} \newcommand{\rmoduleprolong}[1]{\mathfrak #1} \newcommand{\vecnbign}{{\mathcal {\boldsymbol {\tilde{N}}}}} \newcommand{\nbigvecn}{{\mathcal{\boldsymbol N}}} \newcommand{\supp}{\gminis} \newcommand{\kakkolambdazero}{^{(\lambda_0)}} \newcommand{\tildenbigm}{\widetilde{\nbigm}} \newcommand{\tildenbigt}{\widetilde{\nbigt}} \newcommand{\tildegbige}{\widetilde{\gbige}} \newcommand{\tildenbige}{\widetilde{\nbige}} \newcommand{\dualhenomap}{\clubsuit} \newcommand{\rtriplecat}{\nbigr\textrm{-Triples}} \newcommand{\closedopen}[2]{[#1,#2[} \newcommand{\openclosed}[2]{]#1,#2]} \newcommand{\openopen}[2]{]#1,#2[} \newcommand{\closedclosed}[2]{[#1,#2]} \newcommand{\PH}{{\mathcal P}{\mathcal H}} \newcommand{\harmonicbundleretu}{\bigl(E^{(i)},\delbar^{(i)},\theta^{(i)},h^{(i)}\bigr)} \newcommand{\harmonicbundlebubunretu}{\bigl(E^{(i')},\delbar^{(i')},\theta^{(i')},h^{(i')}\bigr)} \newcommand{\harmonicbundleinfty}{\bigl(E^{(\infty)},\delbar^{(\infty)},\theta^{(\infty)},h^{(\infty)}\bigr)} \newcommand{\regularfilteredHiggsbundleretu}{\bigl(E^{(i)},\delbar^{(i)},\vecF^{(i)},\theta^{(i)}\bigr)} \newcommand{\regularfilteredHiggsbundlebubunretu}{\bigl(E^{(i')},\delbar^{(i')},\vecF^{(i')},\theta^{(i')}\bigr)} \newcommand{\regularfilteredHiggsbundleinfty}{\bigl(E^{(\infty)},\delbar^{(\infty)},\vecF^{(\infty)},\theta^{(\infty)}\bigr)} \newcommand{\nbigeilambda}{\nbige^{(i)\,\lambda}} \newcommand{\nbigeinftylambda}{\nbige^{(\infty)\,\lambda}} \newcommand{\DDilambda}{\DD^{(i)\,\lambda}} \newcommand{\DDinftylambda}{\DD^{(\infty)\,\lambda}} \newcommand{\nbigeizero}{\nbige^{(i)\,0}} \newcommand{\nbigeinftyzero}{\nbige^{(\infty)\,0}} \newcommand{\unitary}{\gminiu} \newcommand{\gl}{\gminig\gminil} \newcommand{\minisl}{\gminis\gminil} \newcommand{\delbarhat}{\widehat{\delbar}} \newcommand{\delhat}{\widehat{\del}} \newcommand{\omegahat}{\widehat{\omega}} \newcommand{\omegatilde}{\widetilde{\omega}} \newcommand{\vecEhat}{\widehat{\vecE}} \newcommand{\thetahat}{\widehat{\theta}} \newcommand{\Ehat}{\widehat{E}} \newcommand{\hhat}{\widehat{h}} \newcommand{\vecvhat}{\widehat{\vecv}} \newcommand{\vhat}{\widehat{v}} \newcommand{\Ahat}{\widehat{A}} \newcommand{\Gtilde}{\widetilde{G}} \newcommand{\rhotilde}{\widetilde{\rho}} \newcommand{\gun}{G} \newcommand{\xtilde}{\widetilde{x}} \newcommand{\Etilde}{\widetilde{E}} \newcommand{\vecEtilde}{\widetilde{\vecE}} \newcommand{\thetatilde}{\widetilde{\theta}} \newcommand{\Vhat}{\widehat{V}} \newcommand{\nablahat}{\widehat{\nabla}} \newcommand{\Vtilde}{\widetilde{V}} \newcommand{\nablatilde}{\widetilde{\nabla}} \newcommand{\vecvtilde}{\widetilde{\vecv}} \newcommand{\vecwtilde}{\widetilde{\vecw}} \newcommand{\vecutilde}{\widetilde{\vecu}} \newcommand{\zerohat}{\widehat{0}} \newcommand{\vecvbar}{\overline{\vecv}} \newcommand{\fhat}{\widehat{f}} \newcommand{\vecVhat}{\widehat{\vecV}} \newcommand{\vecwbar}{\overline{\vecw}} \newcommand{\wbar}{\overline{w}} \newcommand{\gbar}{\overline{g}} \newcommand{\Chat}{\widehat{C}} \newcommand{\kappatilde}{\widetilde{\kappa}} \newcommand{\mutilde}{\widetilde{\mu}} \newcommand{\nbigetilde}{\widetilde{\nbige}} \newcommand{\DDlambdatilde}{\widetilde{\DD}^{\lambda}} \newcommand{\htilde}{\widetilde{h}} \newcommand{\nbigelambdatilde}{\widetilde{\nbige}^{\lambda}} \newcommand{\ptilde}{\widetilde{p}} \newcommand{\pitilde}{\widetilde{\pi}} \newcommand{\Phitilde}{\widetilde{\Phi}} \newcommand{\vtilde}{\widetilde{v}} \newcommand{\phat}{\widehat{p}} \newcommand{\nbigbhat}{\widehat{\nbigb}} \newcommand{\ftilde}{\widetilde{f}} \newcommand{\utilde}{\widetilde{u}} \newcommand{\Ftilde}{\widetilde{F}} \newcommand{\Phibar}{\overline{\Phi}} \newcommand{\gminiabar}{\overline{\gminia}} \newcommand{\atilde}{\widetilde{a}} \newcommand{\stilde}{\widetilde{s}} \newcommand{\mubar}{\overline{\mu}} \newcommand{\Psitilde}{\widetilde{\Psi}} \newcommand{\nbiglhat}{\widehat{\nbigl}} \newcommand{\sigmatilde}{\widetilde{\sigma}} \newcommand{\nbigltilde}{\widetilde{\nbigl}} \newcommand{\nbigdhat}{\widehat{\nbigd}} \newcommand{\DDtilde}{\widetilde{\DD}} \newcommand{\nbigxhat}{\widehat{\nbigx}} \newcommand{\nbigkhat}{\widehat{\nbigk}} \newcommand{\Rtilde}{\widetilde{R}} \newcommand{\Dhat}{\widehat{D}} \newcommand{\vecgammabar}{\overline{\vecgamma}} \newcommand{\Itilde}{\widetilde{I}} \newcommand{\nbigbtilde}{\widetilde{\nbigb}} \newcommand{\vecwhat}{\widehat{\vecw}} \newcommand{\pihat}{\widehat{\pi}} \newcommand{\Omegabar}{\overline{\Omega}} \newcommand{\Fhat}{\widehat{F}} \newcommand{\nbigfhat}{\widehat{\nbigf}} \newcommand{\Ghat}{\widehat{G}} \newcommand{\puebar}{\overline{p}} \newcommand{\Rhat}{\widehat{R}} \newcommand{\vecuhat}{\widehat{\vecu}} \newcommand{\Ohat}{\widehat{O}} \newcommand{\Sbar}{\overline{S}} \newcommand{\nbigehat}{\widehat{\nbige}} \newcommand{\Phihat}{\widehat{\Phi}} \newcommand{\nbigyhat}{\widehat{\nbigy}} \newcommand{\nbigvtilde}{\widetilde{\nbigv}} \newcommand{\nbigvhat}{\widehat{\nbigv}} \newcommand{\ellsitabar}{\underline{\ell}} \newcommand{\Utilde}{\widetilde{U}} \newcommand{\Dtilde}{\widetilde{D}} \newcommand{\Xtilde}{\widetilde{X}} \newcommand{\vecy}{\boldsymbol y} \newcommand{\Zhat}{\widehat{Z}} \newcommand{\What}{\widehat{W}} \newcommand{\DDhat}{\widehat{\DD}} \newcommand{\nbigrtilde}{\widetilde{\nbigr}} \newcommand{\nbigghat}{\widehat{\nbigg}} \newcommand{\Ltilde}{\widetilde{L}} \newcommand{\nbigmtilde}{\widetilde{\nbigm}} \def\Der{\mathop{\rm Der}\nolimits} \def\Stab{\mathop{\rm Stab}\nolimits} \def\ord{\mathop{\rm ord}\nolimits} \def\Herm{\mathop{\rm Herm}\nolimits} \def\full{\mathop{\rm full}\nolimits} \def\MT{\mathop{\rm MT}\nolimits} \def\MTW{\mathop{\rm MTW}\nolimits} \def\MTN{\mathop{\rm MTN}\nolimits} \def\wild{\mathop{\rm wild}\nolimits} \def\Gal{\mathop{\rm Gal}\nolimits} \def\strict{\mathop{\rm strict}\nolimits} \def\pol{\mathop{\rm pol}\nolimits} \def\Can{\mathop{\rm Can}\nolimits} \def\Var{\mathop{\rm Var}\nolimits} \def\moderate{\mathop{\rm mod}\nolimits} \def\op{\mathop{\rm op}\nolimits} \def\Mod{\mathop{\rm Mod}\nolimits} \def\ramification{\mathop{\rm rami}\nolimits} \def\pureimaginary{\mathop{\rm pi}\nolimits} \def\Hol{\mathop{\rm Hol}\nolimits} \def\Sec{\mathop{\rm Sec}\nolimits} \def\Sep{\mathop{\rm Sep}\nolimits} \def\sing{\mathop{\rm sing}\nolimits} \def\Gys{\mathop{\rm Gys}\nolimits} \def\exchange{\mathop{\rm exchange}\nolimits} \def\HS{\mathop{\rm HS}\nolimits} \def\TNIL{\mathop{\rm TNIL}\nolimits} \def\Glue{\mathop{\rm Glue}\nolimits} \def\Seg{\mathop{\rm Seg}\nolimits} \def\El{\mathop{\rm El}\nolimits} \def\Sing{\mathop{\rm Sing}\nolimits} \def\Ray{\mathop{\rm Ray}\nolimits} \def\Arc{\mathop{\rm Arc}\nolimits} \def\Lift{\mathop{\rm Lift}\nolimits} \def\slope{\mathop{\rm slope}\nolimits} \def\dec{\mathop{\rm dec}\nolimits} \def\local{\mathop{\rm local}\nolimits} \def\DR{\mathop{\rm DR}\nolimits} \def\ob{\mathop{\rm ob}\nolimits} \def\Mor{\mathop{\rm Mor}\nolimits} \def\Ch{\mathop{\rm Ch}\nolimits} \def\good{\mathop{\rm good}\nolimits} \def\Map{\mathop{\rm Map}\nolimits} \def\hol{\mathop{\rm hol}\nolimits} \def\Loc{\mathop{\rm Loc}\nolimits} \def\Tor{\mathop{\rm Tor}\nolimits} \def\Cat{\mathop{\rm Cat}\nolimits} \def\rapid{\mathop{\rm rap}\nolimits} \def\rel{\mathop{\rm rel}\nolimits} \newcommand{\Ubar}{\overline{U}} \newcommand{\Dbar}{\overline{D}} \newcommand{\Pbar}{\overline{P}} \newcommand{\Ibar}{\overline{I}} \newcommand{\Wtilde}{\widetilde{W}} \newcommand{\Ptilde}{\widetilde{P}} \newcommand{\psihat}{\widehat{\psi}} \newcommand{\iotahat}{\widehat{\iota}} \newcommand{\vecR}{{\boldsymbol R}} \newcommand{\vbar}{\overline{v}} \newcommand{\ubar}{\overline{u}} \newcommand{\Tbar}{\overline{T}} \newcommand{\gminibbar}{\overline{\gminib}} \newcommand{\vecj}{{\boldsymbol j}} \newcommand{\vecD}{{\boldsymbol D}} \newcommand{\nbigehatlambda}{\widehat{\nbige}^{\lambda}} \newcommand{\nbigelambdahat}{\nbigehatlambda} \newcommand{\nbigutilde}{\widetilde{\nbigu}} \newcommand{\nbigstilde}{\widetilde{\nbigs}} \newcommand{\tauhat}{\widehat{\tau}} \newcommand{\That}{\widehat{T}} \newcommand{\Phat}{\widehat{P}} \newcommand{\nbigmhat}{\widehat{\nbigm}} \newcommand{\nbigzhat}{\widehat{\nbigz}} \newcommand{\Nhat}{\widehat{N}} \newcommand{\gbigehat}{\widehat{\gbige}} \newcommand{\gtilde}{\widetilde{g}} \newcommand{\Ztilde}{\widetilde{Z}} \newcommand{\nbigttilde}{\widetilde{\nbigt}} \newcommand{\gbigstilde}{\widetilde{\gbigs}} \newcommand{\varphitilde}{\widetilde{\varphi}} \newcommand{\Ctilde}{\widetilde{C}} \newcommand{\nbigntilde}{\widetilde{\nbign}} \newcommand{\gbigetilde}{\widetilde{\gbige}} \newcommand{\gbigctilde}{\widetilde{\gbigc}} \newcommand{\Tzero}{T^{(\lambda_0)}} \newcommand{\Ttildezero}{\widetilde{T}^{(\lambda_0)}} \newcommand{\Stilde}{\widetilde{S}} \newcommand{\tautilde}{\widetilde{\tau}} \newcommand{\gbigelambda}{\gbige^{\lambda}} \newcommand{\ybar}{\overline{y}} \newcommand{\epsilonbar}{\overline{\epsilon}} \newcommand{\Atilde}{\widetilde{A}} \newcommand{\Otilde}{\widetilde{O}} \newcommand{\gbigrtilde}{\widetilde{\gbigr}} \newcommand{\gbigmtilde}{\widetilde{\gbigm}} \newcommand{\gbigltilde}{\widetilde{\gbigl}} \newcommand{\Deltatilde}{\widetilde{\Delta}} \newcommand{\nbigftilde}{\widetilde{\nbigf}} \newcommand{\nbigtzero}{\nbigt^{(\lambda_0)}} \newcommand{\nbigpzero}{\nbigp^{(\lambda_0)}} \newcommand{\Mtilde}{\widetilde{M}} \newcommand{\Ytilde}{\widetilde{Y}} \newcommand{\Ybar}{\overline{Y}} \newcommand{\wtilde}{\widetilde{w}} \newcommand{\ztilde}{\widetilde{z}} \newcommand{\nbiglbar}{\overline{\nbigl}} \newcommand{\nbigkbar}{\overline{\nbigk}} \newcommand{\nbigrmod}{\nbigr-\Mod} \newcommand{\itilde}{\widetilde{i}} \newcommand{\nbigxtilde}{\widetilde{\nbigx}} \newcommand{\gbigttilde}{\widetilde{\gbigt}} \newcommand{\nbigvlambda}{\nbigv^{\lambda}} \newcommand{\nbigztilde}{\widetilde{\nbigz}} \newcommand{\nbigdtilde}{\widetilde{\nbigd}} \newcommand{\Qtilde}{\widetilde{Q}} \newcommand{\Ttilde}{\widetilde{T}} \newcommand{\gbigvtilde}{\widetilde{\gbigv}} \newcommand{\gbigntilde}{\widetilde{\gbign}} \newcommand{\what}{\widehat{w}} \newcommand{\uhat}{\widehat{u}} \newcommand{\nbigsbar}{\overline{\nbigs}} \newcommand{\betabar}{\overline{\beta}} \newcommand{\vecatilde}{\widetilde{\veca}} \newcommand{\gminicbar}{\overline{\gminic}} \newcommand{\forget}{for} \newcommand{\nbigitilde}{\widetilde{\nbigi}} \newcommand{\Secbar}{\overline{\Sec}} \newcommand{\Multisector}{\nbigm\nbigs} \newcommand{\SD}{\nbigs\nbigd} \newcommand{\PSD}{\nbigp\nbigs\nbigd} \newcommand{\Irrbar}{\overline{\Irr}} \newcommand{\psitabar}{\underline{p}} \newcommand{\nbigktilde}{\widetilde{\nbigk}} \newcommand{\Ssitabar}{\underline{S}} \newcommand{\vecctilde}{\widetilde{\vecc}} \newcommand{\vecnbigi}{{\boldsymbol \nbigi}} \newcommand{\vectheta}{{\boldsymbol \theta}} \newcommand{\DDprimelambda}{\DD^{\prime\lambda}} \newcommand{\Lambdatilde}{\widetilde{\Lambda}} \newcommand{\nbigmlambda}{\nbigm^{\lambda}} \newcommand{\nrhom}{R{\mathcal Hom}} \newcommand{\DDD}{\boldsymbol D} \newcommand{\jtilde}{\widetilde{j}} \newcommand{\vecH}{{\boldsymbol H}} \newcommand{\nbiggzero}{\nbigg^{(\lambda_0)}} \newcommand{\Dtildehat}{\widehat{\Dtilde}} \newcommand{\ctilde}{\widetilde{c}} \newcommand{\taubar}{\overline{\tau}} \newcommand{\Kbar}{\overline{K}} \newcommand{\Ktilde}{\widetilde{K}} \newcommand{\veckappa}{{\boldsymbol \kappa}} \newcommand{\TTtilde}{{T}\widetilde{T}} \newcommand{\Hhat}{\widehat{H}} \newcommand{\Htilde}{\widetilde{H}} \newcommand{\Omegatilde}{\widetilde{\Omega}} \newcommand{\Ntilde}{\widetilde{N}} \newcommand{\nbigxzero}{\nbigx^{(\lambda_0)}} \newcommand{\nbigdzero}{\nbigd^{(\lambda_0)}} \newcommand{\nutilde}{\widetilde{\nu}} \newcommand{\vecmtilde}{\widetilde{\vecm}} \newcommand{\gminiatilde}{\widetilde{\gminia}} \newcommand{\gminibtilde}{\widetilde{\gminib}} \newcommand{\cnumtilde}{\widetilde{\cnum}} \newcommand{\phibar}{\overline{\phi}} \newcommand{\HSbar}{\overline{\HS}} \newcommand{\phitilde}{\widetilde{\phi}} \newcommand{\nbightilde}{\widetilde{\nbigh}} \newcommand{\Vbar}{\overline{V}} \newcommand{\Wbar}{\overline{W}} \newcommand{\Nbar}{\overline{N}} \newcommand{\DDbar}{\overline{\DD}} \newcommand{\Ebar}{\overline{E}} \newcommand{\Gbar}{\overline{G}} \newcommand{\Fourier}{\mathfrak{Four}} \newcommand{\Xbar}{\overline{X}} \newcommand{\nbigwhat}{\widehat{\nbigw}} \newcommand{\rhohat}{\widehat{\rho}} \newcommand{\varphihat}{\widehat{\varphi}} \newcommand{\gammatilde}{\widetilde{\gamma}} \newcommand{\Gammatilde}{\widetilde{\Gamma}} \newcommand{\Shat}{\widehat{S}} \newcommand{\Xitilde}{\widetilde{\Xi}} \newcommand{\gminiahat}{\widehat{\gminia}} \newcommand{\etahat}{\widehat{\eta}} \newcommand{\Cbar}{\overline{C}} \newcommand{\vecUtilde}{\widetilde{\vecU}} \newcommand{\vecUpsilon}{{\boldsymbol \Upsilon}} \newcommand{\vecUbar}{\overline{\vecU}} \newcommand{\Xibar}{\overline{\Xi}} \newcommand{\gminichat}{\widehat{\gminic}} \newcommand{\piinverseDhat}{\widehat{\pi^{-1}(D)}} \newcommand{\Nilhat}{\widehat{\Nil}} \newcommand{\qtilde}{\widetilde{q}} \newcommand{\veci}{\boldsymbol i} \newcommand{\vecell}{\boldsymbol \ell} \newcommand{\vecq}{\boldsymbol q} \newcommand{\nbigatilde}{\widetilde{\nbiga}} \newcommand{\starbar}{\overline{\star}} \newcommand{\mnuleq}{\prec} \newcommand{\mnugeq}{\succ} \newcommand{\nbigctilde}{\widetilde{\nbigc}} \newcommand{\iotatilde}{\widetilde{\iota}} \newcommand{\nbigqtilde}{\widetilde{\nbigq}} \newcommand{\ktilde}{\widetilde{k}} \newcommand{\nbigptilde}{\widetilde{\nbigp}} |
1501.04146 | Title: Twistor property of GKZ-hypergeometric systems
Abstract: We study the mixed twistor D-modules associated to meromorphic functions. In
particular, we describe their push-forward and specialization under some
situations. We apply the results to study the twistor property of a type of
GKZ-hypergeometric systems, and to study their specializations. As a result, we
obtain some isomorphisms of mixed TEP-structures in the local mirror symmetry.
Body: \subsection{Mixed twistor $\nbigd$-modules} In , C. Simpson introduced the concept of mixed twistor structure as a generalization of mixed Hodge structure. According to his principle, called Simpson's meta theorem, most objects and theorems concerning with mixed Hodge structure should have their counterparts in the context of mixed twistor structure. The concept of mixed Hodge module due to M. Saito is one of the ultimate in the Hodge theory. Roughly speaking, mixed Hodge modules are regular holonomic $\nbigd$-modules equipped with mixed Hodge structure. One of the main results in the theory is the functoriality of mixed Hodge modules. Namely, we have the standard operations on the derived category of algebraic mixed Hodge modules such as push-forward, pull back, duality, inner homomorphism, tensor product, nearby cycle functor, vanishing cycle functors, which are compatible with the standard operations for algebraic regular holonomic $\nbigd$-modules. According to Simpson's meta theorem, we should have a twistor version of mixed Hodge modules, that is the concept of mixed twistor $\nbigd$-module. The concept of pure twistor $\nbigd$-module was introduced by C. Sabbah , and studied by himself and the author . The mixed case was studied in . As in the Hodge case, we have the standard functors on the derived category of algebraic mixed twistor $\nbigd$-modules, that is one of the most useful points in the theory of mixed twistor $\nbigd$-modules. Another interesting point is that we have the mixed twistor $\nbigd$-modules associated to meromorphic functions. Let $f$ be a meromorphic function on a complex manifold $X$ whose poles are contained in a hypersurface $H$. Let $L_{\ast}(f,H)$ be the holonomic $\nbigd$-modules given by $\nbigo_X(\ast H)$ with the flat connection $d+df$. We also have the holonomic $\nbigd$-module $L_!(f,H):=\DDD(L_{\ast}(-f,H))$, where $\DDD$ denotes the duality of holonomic $\nbigd$-modules. We have the natural mixed twistor $\nbigd$-modules over $L_{\star}(f,H)$ $(\star=\ast,!)$. Applying the standard functors to such mixed twistor $\nbigd$-modules, we can obtain many mixed twistor $\nbigd$-modules. In other words, many important holonomic $\nbigd$-modules are equipped with a natural mixed twistor structure. It would be interesting to have applications of the mixed twistor structure. In this paper, we shall explain that a type of better behaved GKZ-hypergeometric systems are naturally equipped with the mixed twistor structure. We shall also explain an application in the study of toric local mirror symmetry. \subsubsection{Ingredients for mixed twistor $\nbigd$-modules} Originally, ingredients for twistor $\nbigd$-modules are given as $\nbigr$-triples by Sabbah . (See \S for a review.) But, in this paper, we consider only integrable mixed twistor $\nbigd$-modules with real structure. So, in this introduction, we explain it in a slightly different way, although we use the formalism of $\nbigrtilde$-triples after \S. Let $X$ be any complex manifold. Let $\nbigx:=\cnum\times X$ and $\nbigx^0:=\{0\}\times X$. Let $\lambda$ be the standard coordinate of $\cnum$. We have the sheaf of differential operators $\nbigd_{\nbigx}$ on $\nbigx$. Let $\nbigrtilde_X\subset \nbigd_{\nbigx}$ denote the sheaf of subalgebras generated by $\lambda\Theta_{\nbigx}(\log\nbigx^0)$. An $\nbigrtilde_X$-module $\nbigm$ is called strict if it is flat over $\nbigo_{\cnum}$. Any coherent $\nbigrtilde_X$-module $\nbigm$ has the characteristic variety $\Ch(\nbigm)$ in $T^{\ast}\nbigx(\log\nbigx^0)$ as in the case of $\nbigd$-modules. We say that $\nbigm$ is holonomic if $\Ch(\nbigm)$ is contained in $\cnum\times \Lambda$ for a Lagrangian subvariety $\Lambda\subset T^{\ast}X$. A real structure of strict holonomic $\nbigrtilde_X$-module $\nbigm$ is a $\real$-perverse sheaf $P_{\real}$ on $\cnum^{\ast}\times X$ with an isomorphism $P_{\real}\otimes\cnum \simeq \DR_{\cnum^{\ast}\times X} (\nbigm_{|\cnum^{\ast}\times X})$. Then, an integrable mixed twistor $\nbigd$-module with real structure on $X$ is a strict holonomic $\nbigrtilde_X$-module $\nbigm$ with a real structure and a weight filtration $(\nbigm,P_{\real},W)$ satisfying some conditions. Here, $W$ is an increasing filtration of strict holonomic $\nbigrtilde_X$-modules with real structure. In this introduction, ``mixed twistor $\nbigd$-module'' means ``integrable mixed twistor $\nbigd$-module with real structure''. \subsection{Better behaved GKZ-hypergeometric systems} \subsubsection{$\nbigd$-modules} L. Borisov and P. Horja introduced the concept of better behaved GKZ-hypergeometric system as a generalization of GKZ-hypergeometric systems . (See also .) Let us recall a special type of better behaved GKZ-hypergeometric systems. Let $\nbiga=\{\veca_1,\ldots,\veca_m\}\subset \seisuu^n$ be a finite subset generating $\seisuu^n$. We have the cone $K_{\real}(\nbiga):=\bigl\{ \sum_{j=1}^m r_j\veca_j\,\big|\, r_j\geq 0 \bigr\} \subset\real^n$ generated by $\nbiga$. We set $K(\nbiga):=K_{\real}(\nbiga)\cap\seisuu^n$. The semigroup $K(\nbiga)$ is the saturation of $\seisuu_{\geq 0}\nbiga =\bigl\{ \sum_{j=1}^m n_{j}\veca_j\,\big|\, n_{j}\in\seisuu_{\geq 0} \bigr\}$. We denote $\veca_j=(a_{j1},\ldots,a_{jn})$. Let $\Gamma\subset K(\nbiga)$ be any subset such that $\Gamma+\veca\subset \Gamma$ for any $\veca\in\nbiga$. Let $\vecbeta\in\cnum^n$. The following system of differential equations $\GKZ(\nbiga,\Gamma,\vecbeta)$ for tuples of holomorphic functions $(\Phi_{\vecc}\,|\,\vecc\in \Gamma)$ on any open subset in $\cnum^m$ is called the better behaved GKZ-hypergeometric system associated to $(\nbiga,\Gamma,\vecbeta)$: \[ \del_{x_j}\Phi_{\vecc}=\Phi_{\vecc+\veca_j} \quad(\forall \vecc\in \Gamma,\,\forall j=1,\ldots,m) \] \[ \Bigl( \sum_{j}a_{ji}x_j\del_{x_j}+c_i-\beta_i \Bigr) \Phi_{\vecc}=0 \quad (\forall \vecc\in \Gamma,\,\forall i=1,\ldots,n) \] Here, $(x_1,\ldots,x_m)$ denotes the standard coordinate system of $\cnum^m$. \begin{rem} In {\rm}, any tuple in a finitely generated abelian group is considered instead of a finite subset in $\seisuu^n$. In that sense, the above system is a special case of better behaved GKZ-hypergeometric systems. Later, moreover, we shall mainly consider the case $\vecbeta=0$. But, as in {\rm }, we omit the existence of an element $\vecalpha\in(\seisuu^n)^{\lor}$ such that $\vecalpha(\veca_i)=1$ for any $i$. \hfill\qed \end{rem} \paragraph{Expression as the twisted Gauss-Manin systems} We can describe the corresponding $\nbigd$-modules using the twisted Gauss-Manin system. Let $T^n:=(\cnum^{\ast})^n$. We consider the morphism $\psi^{\aff}_{\nbiga}:T^n\lrarr \cnum^m$ given by $\psi^{\aff}_{\nbiga}(t_1,\ldots,t_n) =(t^{\veca_1},\ldots,t^{\veca_m})$, where $t^{\veca_j}=\prod_{i=1}^nt_i^{a_{ji}}$. Let $X_{\nbiga}^{\aff}$ denote the closure of the image of $\psi^{\aff}_{\nbiga}$. Let $\check{X}^{\aff}_{\nbiga}\lrarr X^{\aff}_{\nbiga}$ be the normalization. Let $\check{D}^{\aff}_{\nbiga}$ denote the complement of $T^n$ in $\check{X}^{\aff}_{\nbiga}$. Note that $\check{X}^{\aff}_{\nbiga} =\Spec \cnum[K(\nbiga)]$ and $X^{\aff}_{\nbiga} =\Spec\cnum[\seisuu_{\geq 0}\nbiga]$. Let $\Omega^k_{\check{X}^{\aff}_{\nbiga}}(\log \check{D}^{\aff}_{\nbiga})$ denote the sheaf of meromorphic differential $k$-forms which are logarithmic along $\check{D}^{\aff}_{\nbiga}$, studied in . Let $\Omega^k_{(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga})}$ denote the sheaf of holomorphic $k$-forms on the normal variety $\check{X}^{\aff}_{\nbiga}$ which are $0$ along $\check{D}^{\aff}_{\nbiga}$, studied in . Let $q:\check{X}^{\aff}_{\nbiga}\times\cnum^m\lrarr \check{X}^{\aff}_{\nbiga}$ be the projection. We set \[ \Omega^k_{\check{X}^{\aff}_{\nbiga}\times\cnum^m/\cnum^m} (\log \check{D}^{\aff}_{\nbiga}\times\cnum^m):= q^{\ast}\Omega^k_{\check{X}^{\aff}_{\nbiga}}(\log \check{D}^{\aff}_{\nbiga}), \quad\quad \Omega^k_{ (\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga}) \times\cnum^m /\cnum^m}:= q^{\ast}\Omega^k_{(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga})}. \] The family of Laurent polynomials $\sum_{j=1}^m x_jt^{\veca_j}$ induces a meromorphic function $F_{\nbiga}$ on $(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga})\times\cnum^m$. We also have the logarithmic closed one form $\kappa(\vecbeta)=\sum_{i=1}^n \beta_i dt_i/t_i$. We obtain the relative algebraic de Rham complexes: \[ \nbigc^{\bullet}(\nbiga,\vecbeta)_{\bullet}:= \Bigl( \Omega^{\bullet}_{(\check{X}^{\aff}_{\nbiga}\times \cnum^m)/\cnum^m} (\log\check{D}^{\aff}_{\nbiga}\times\cnum^m), d+dF_{\nbiga}-\kappa(\vecbeta) \Bigr) \] \[ \nbigc^{\bullet}(\nbiga,\vecbeta)_{\circ}:= \Bigl( \Omega^{\bullet}_{(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga}) \times \cnum^m/\cnum^m}, d+dF_{\nbiga}-\kappa(\vecbeta) \Bigr) \] Let $\check{\pi}_{\nbiga}: \check{X}^{\aff}_{\nbiga}\times\cnum^m \lrarr \cnum^m$ denote the projection. Each $\check{\pi}_{\nbiga\ast}\nbigc^{k}(\nbiga,\vecbeta)_{\star}$ is naturally a $\nbigd_{\cnum^m}$-module by $\del_{x_j}\cdot(g)=\del_{x_j}g+(\del_{x_j}F_{\nbiga}) g$. The differential $d+dF_{\nbiga}-\kappa(\vecbeta)$ of the complexes are compatible with the actions of $\nbigd_{\cnum^m}$. So, we obtain the following $\nbigd_{\cnum^m}$-modules \[ M_{\nbiga,\vecbeta,\star}:= \hyperr^n\check{\pi}_{\nbiga\ast} \nbigc^{\bullet}(\nbiga,\vecbeta)_{\star} \quad (\star=\bullet,\circ). \] Let $K(\nbiga)^{\circ}$ be the intersection of $K(\nbiga)$ and the interior part of $K_{\real}(\nbiga)$. The $\nbigd_{\cnum^m}$-modules $M_{\nbiga,\vecbeta,\bullet}$ and $M_{\nbiga,\vecbeta,\circ}$ correspond to the systems $\GKZ(\nbiga,K(\nbiga),\vecbeta)$ and $\GKZ(\nbiga,K(\nbiga)^{\circ},\vecbeta)$ in the sense of the following lemma. (See Lemma and Proposition for more details.) \begin{prop} Let $U$ be any open subset in $\cnum^m$. We have a natural bijective correspondence between solutions of $\GKZ(\nbiga,K(\nbiga),\vecbeta)$ (resp. $\GKZ(\nbiga,K(\nbiga)^{\circ},\vecbeta)$) and $\nbigd_{U}$-homomorphisms $M_{\nbiga,\vecbeta,\bullet}\lrarr \nbigo_U$ (resp. $M_{\nbiga,\vecbeta,\circ}\lrarr \nbigo_U$). \end{prop} In this paper, we are mainly interested in the case $\vecbeta=0$. \paragraph{Expressions as the push-forward of $\nbigd$-modules} We take a toric resolution $\varphi_{\Sigma_1}:X_{\Sigma_1} \lrarr \check{X}^{\aff}_{\nbiga}$, i.e., $X_{\Sigma_1}$ be a smooth toric variety, and $\varphi_{\Sigma_1}$ is a projective morphism which is $T^n$-equivariant. Let $D_{\Sigma_1}:=X_{\Sigma_1}\setminus T^n$. We have the meromorphic algebraic function $F_{\nbiga,\Sigma_1}$ on $(X_{\Sigma_1},D_{\Sigma_1})\times\cnum^m$ given by $F_{\nbiga}$. We have the algebraic holonomic $\nbigd_{X_{\Sigma_1}\times\cnum^m}$-modules $L_{\star}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times \cnum^m)$ ($\star=\ast,!$). Let $\pi_{\Sigma_1}: X_{\Sigma_1}\times\cnum^m\lrarr \cnum^m$ denote the projection. By the results in \S and \S, we have the following proposition, which gives expressions of $\GKZ(\nbiga,K(\nbiga),0)$ and $\GKZ(\nbiga,K(\nbiga)^{\circ},0)$ in terms of $\nbigd$-modules. \begin{prop} We have natural isomorphisms \[ \pi^0_{\Sigma_1\ast} L_{\ast}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \simeq \pi^0_{\Sigma_1!} L_{\ast}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \simeq M_{\nbiga,0,\bullet}, \] \[ \pi^0_{\Sigma_1\ast} L_{!}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \simeq \pi^0_{\Sigma_1!} L_{!}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \simeq M_{\nbiga,0,\circ}. \] Here, $\pi^0_{\Sigma_1,\star}$ $(\star=\ast,!)$ denote the $0$-th cohomology of the push-forward functors of algebraic $\nbigd$-modules with respect to $\pi_{\Sigma_1}$. \end{prop} \paragraph{Special case} For any $\vecp=(p_1,\ldots,p_m)\in \seisuu^m$, we put $\supp_+(\vecp):=\{j\,|\,p_j\geq 0\}$ and $\supp_-(\vecp):=\{j\,|\,p_j\leq 0\}$. We set \[ \square_{\vecp} =\prod_{j\in\supp_+(\vecp)}\del_{x_j}^{p_j} -\prod_{j\in\supp_-(\vecp)}\del_{x_j}^{-p_j}. \] We have the morphism $\seisuu^m\lrarr \seisuu^n$ given by $\vecp=(p_1,\ldots,p_m)\longmapsto \sum_{j=1}^m p_j\veca_j$. Let $L_{\nbiga}$ denote the kernel. For $\vecc_0\in K(\nbiga)$, we have the following ordinary GKZ-hypergeometric system $\GKZ^{\ord}(\nbiga,\vecbeta-\vecc_0)$ for holomorphic functions $\Phi_{\vecc_0}$ on any open subset of $\cnum^m$: \[ \square_{\vecp}\Phi_{\vecc_0}=0 \quad (\forall\vecp\in L_{\nbiga}) \] \[ \Bigl(c_{0i}-\beta_i+\sum_{j=1}^ma_{ji}x_j\del_{x_j}\Bigr) \Phi_{\vecc_0}=0 \quad (i=1,\ldots,n) \] For any $\vecgamma\in\cnum^n$, let $I(\nbiga,\vecgamma)$ denote the left ideal of $\nbigd_{\cnum^m}$ generated by $\square_{\vecp}$ $(\vecp\in L_{\nbiga})$ and $-\gamma_i+\sum_{j=1}^ma_{ji}x_j\del_{x_j}$ $(i=1,\ldots,n)$. Then, for $\Gamma=\vecc_0+\seisuu_{\geq 0}\nbiga$, we have a natural isomorphism $M^{\GKZ}(\nbiga,\Gamma,\vecbeta) \simeq \nbigd_{\cnum^m}/I(\nbiga,\vecbeta-\vecc_0)$ which is the $\nbigd$-module corresponding to the system $\GKZ^{\ord}(\nbiga,\vecbeta-\vecc_0)$. Suppose that $K(\nbiga)=\seisuu_{\geq 0}\nbiga$. Then, we have a natural isomorphism \begin{equation} M_{\nbiga,0,\bullet} \simeq \nbigd_{\cnum^m}/I(\nbiga,0). \end{equation} Suppose moreover $K(\nbiga)^{\circ}=\seisuu_{\geq 0}\nbiga+\vecc_0$ for an element $\vecc_0\in K(\nbiga)$. Then, we have a natural isomorphism \begin{equation} M_{\nbiga,0,\circ} \simeq \nbigd_{\cnum^m}/I(\nbiga,-\vecc_0). \end{equation} For an expression $\vecc_0=\sum_{i=1}^m b_i\veca_i$, we have the following commutative diagram: \[ \begin{CD} M_{\nbiga,0,\circ} @>{a}>> M_{\nbiga,0,\bullet} \\ @V{\simeq}VV @V{\simeq}VV \\ \nbigd_{\cnum^m}\big/I(\nbiga,-\vecc_0) @>{b}>> \nbigd_{\cnum^m}\big/I(\nbiga,0). \end{CD} \] Here, the vertical morphisms are () and (), $a$ is the natural morphism, and $b$ is the induced by the multiplication of $\prod \del_{x_i}^{b_i}$. \subsubsection{Mixed twistor $\nbigd$-modules} By the geometric expression in Proposition , we naturally obtain mixed twistor $\nbigd$-modules $\gbigt_{\nbiga,0,\star}$ over $M_{\nbiga,0,\star}$ $(\star=\bullet,\circ)$. Namely, we have the algebraic mixed twistor $\nbigd$-modules $\nbigt_{\star}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m)$ $(\star=\ast,!)$ on $X_{\Sigma_1}\times \cnum^m$ whose underlying $\nbigd$-modules are $L_{\star}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m)$. We obtain the algebraic mixed twistor $\nbigd$-modules \[ \gbigt_{\nbiga,0,\bullet} =\pi^0_{\Sigma_1\ast} \nbigt_{\ast}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \simeq \pi^0_{\Sigma_1!} \nbigt_{\ast}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \] \[ \gbigt_{\nbiga,0,\circ} =\pi^0_{\Sigma_1\ast} \nbigt_{!}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \simeq \pi^0_{\Sigma_1!} \nbigt_{!}(F_{\nbiga,\Sigma_1},D_{\Sigma_1}\times\cnum^m) \] Let $\nbigm_{\nbiga,0,\star}$ denote the underlying $\nbigrtilde_{\cnum^m}$-modules of $\gbigt_{\nbiga,0,\star}$ $(\star=\ast,!)$. Let us describe them as systems of differential equations. Let $\Gamma\subset K(\nbiga)$ be any subset satisfying $\Gamma+\veca\subset \Gamma$ for any $\veca\in\nbiga$. We consider the following system of differential equations $\GKZ_{\nbigrtilde}(\nbiga,\Gamma,\vecbeta)$ for a tuple $\Phi_{\Gamma}=(\Phi_{\vecc}\,|\,\vecc\in \Gamma)$ of holomorphic functions on any open subset of $\cnum\times\cnum^m =\{(\lambda,x_1,\ldots,x_m)\}$: \[ \lambda\del_{x_j}\Phi_{\vecc} =\Phi_{\vecc+\veca_j}, \quad (\forall \vecc\in \Gamma,\,\,j=1,\ldots,m) \] \[ \Bigl( \lambda^2\del_{\lambda} +n\lambda +\sum_{j=1}^m\lambda x_j\del_{x_j} \Bigr) \Phi_{\vecc} =0, \quad (\forall \vecc\in \Gamma) \] \[ \Bigl( \lambda (c_i-\beta_i)+\sum_{j=1}^m a_{ji}\lambda x_j\del_{x_j} \Bigr) \Phi_{\vecc}=0, \quad (\forall\vecc\in \Gamma,\,\,\,i=1,\ldots,n) \] We have the $\nbigrtilde$-modules $\nbigm^{\GKZ}(\nbiga,\Gamma,\vecbeta)$ corresponding to the system $\GKZ_{\nbigrtilde}(\nbiga,\Gamma,\vecbeta)$. (See \S, in particular Lemma .) According to Proposition and Remark , we obtain the following proposition. \begin{prop} $\nbigm^{\GKZ}(\nbiga,K(\nbiga),0)$ (resp. $\nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0)$) is naturally isomorphic to $\nbigm_{\nbiga,0,\bullet}$ (resp. $\nbigm_{\nbiga,0,\circ}$). \end{prop} \paragraph{Special case} For any $\vecp\in \seisuu^m$, we set \[ \overline{\square}_{\vecp}:= \prod_{j\in \supp_+(\vecp)} (\lambda\del_j)^{p_j} -\prod_{i\in\supp_-(\vecp)} (\lambda\del_j)^{-p_j}. \] For any element $\vecc_0\in K(\nbiga)$, we consider the following system of differential equations $\GKZ_{\nbigrtilde}^{\ord}(\nbiga,\vecbeta-\vecc_0)$ for $\Phi_{\vecc_0}$: \[ \overline{\square}_{\vecp}\Phi_{\vecc_0}=0 \quad (\forall\vecp\in L_{\nbiga}) \] \[ \Bigl( \lambda^2\del_{\lambda} +n\lambda+\sum_{j=1}^mx_j\lambda\del_{x_j} \Bigr)\Phi_{\vecc_0}=0 \] \[ \Bigl( \lambda(c_{0i}-\beta_i) +\sum_{j=1}^ma_{ji}x_j\lambda\del_{x_j}\Bigr) \Phi_{\vecc_0}=0 \quad (i=1,\ldots,n) \] For any $\vecgamma\in\seisuu^m$, let $\nbigi(\nbiga,\vecgamma)$ denote the left ideal of $\nbigrtilde_{\cnum^m}$ generated by $\overline{\square}_{\vecp}$ $(\vecp\in L_{\nbiga})$, $\lambda^2\del_{\lambda}+n\lambda +\sum_{j=1}^m\lambda x_j\del_{x_j}$, and $-\lambda\gamma_i+\sum_{j=1}^ma_{ji}\lambda x_j\del_{x_j}$ $(i=1,\ldots,n)$. Then, we have a natural isomorphism \[ \nbigm^{\GKZ}(\nbiga,\Gamma,\vecbeta) \simeq \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,\vecbeta-\vecc_0) \] which is the $\nbigrtilde_{\cnum^m}$-module corresponding to $\GKZ^{\ord}(\nbiga,\vecbeta-\vecc_0)$. Suppose that $K(\nbiga)=\seisuu_{\geq 0}\nbiga$. Then, we have \begin{equation} \nbigm^{\GKZ}(\nbiga,K(\nbiga),0) \simeq \nbigrtilde_{\cnum^m}\big/ \nbigi(\nbiga,0). \end{equation} Suppose moreover that $K(\nbiga)^{\circ}=K(\nbiga)+\vecc_0$ for an element $\vecc_0\in K(\nbiga)$. Then, we have \begin{equation} \nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) \simeq \nbigrtilde_{\cnum^m}\big/ \nbigi(\nbiga,-\vecc_0). \end{equation} For an expression $\vecc_0=\sum_{i=1}^m b_i\veca_i$, we have the following commutative diagram: \begin{equation} \begin{CD} \nbigm_{\nbiga,0,\circ} @>>> \nbigm_{\nbiga,0,\bullet} \\ @V{\simeq}VV @V{\simeq}VV \\ \nbigrtilde_{\cnum^m}\big/ \nbigi(\nbiga,-\vecc_0) @>>> \nbigrtilde_{\cnum^m}\big/ \nbigi(\nbiga,0) \end{CD} \end{equation} Here, the vertical arrows are given by (), () and Proposition , and the upper horizontal arrow is the morphism underlying the natural morphism of mixed twistor $\nbigd$-modules $\gbigt_{\nbiga,0,\circ} \lrarr \gbigt_{\nbiga,0,\bullet}$, and the lower horizontal arrow is given by the multiplication of $\prod_{i=1}^m(\lambda\del_{x_i})^{b_i}$. \subsubsection{Relation with the reduced quantum $\nbigd$-module of toric complete intersection} Inspired by the work of E. Mann and T. Mignon {\rm}, T. Reichelt and C. Sevenheck {\rm} introduced some systems of differential equations to describe the reduced quantum $\nbigd$-modules of complete intersections. Let $X$ be an $n$-dimensional projective toric manifold corresponding to a fan $\Sigma$. Let $D_i$ $(i=1,\ldots,m)$ be the hypersurfaces of $X$ corresponding to the one dimensional cones in $\Sigma$. Let $K_X$ denote the canonical bundle of $X$. Let $\nbigl_i$ $(i=1,\ldots,r)$ be nef line bundles on $X$ such that $(K_X\otimes\bigotimes_{i=1}^{r}\nbigl_i)^{\lor}$ is nef. We may assume that $\nbigl_j=\nbigo\bigl(\sum_{i=1}^m \beta_{ji}D_i\bigr)$ for some $\beta_{ji}\geq 0$. We can regard $\bigoplus \nbigl_i^{\lor}$ as a toric manifold. Let $\nbiga\subset\seisuu^{n+r}$ be the set of primitive vectors in the one dimensional cones of $\bigoplus \nbigl_i^{\lor}$. We have the systems of differential equations $\GKZ^{\ord}_{\nbigrtilde}(\nbiga,K(\nbiga),0)$ and $\GKZ^{\ord}_{\nbigrtilde}(\nbiga,K(\nbiga)^{\circ},-\vecc_0)$, where $\vecc_0= (\overbrace{0,\ldots,0}^n,\overbrace{-1,\ldots,-1}^r) \in\seisuu^{n+r}$. On $(\cnum^{\ast})^{m+r}$, the systems $\GKZ^{\ord}_{\nbigrtilde}(\nbiga,K(\nbiga),0)$ and $\GKZ^{\ord}_{\nbigrtilde}(\nbiga,K(\nbiga)^{\circ},-\vecc_0)$ are equivalent to the systems in Definition-Lemma 6.1 of , i.e., we have the following commutative diagram: \[ \begin{CD} \bigl( \nbigrtilde_{\cnum^{m+r}} \big/ \nbigi(\nbiga,-\vecc_0) \bigr)_{|(\cnum^{\ast})^{m+r}} @>{\simeq}>> {}^{\ast}_0\widehat{\nbign}^{(0,0,0)} \\ @VVV @V{b}VV \\ \bigl( \nbigrtilde_{\cnum^{m+r}} \big/ \nbigi(\nbiga,0) \bigr)_{|(\cnum^{\ast})^{m+r}} @>{\simeq}>> {}^{\ast}_0\widehat{\nbigm}^{(-r,0,0)} \end{CD} \] (See \cite[\S6]{Reichelt-Sevenheck2} for the $\nbigrtilde$-modules ${}^{\ast}_0\widehat{\nbign}^{(0,0,0)}$ and ${}^{\ast}_0\widehat{\nbigm}^{(-r,0,0)}$.) Hence, the commutative diagram () implies that the $\nbigrtilde$-modules ${}^{\ast}_0\widehat{\nbign}^{(0,0,0)}$ and ${}^{\ast}_0\widehat{\nbigm}^{(-r,0,0)}$ underlie mixed twistor $\nbigd$-modules. We also obtain that the image of the morphism $b$ underlies a pure twistor $\nbigd$-module, which is related with the reduced quantum $\nbigd$-module of the complete intersection. \begin{rem} In {\rm}, they constructed $\nbigrtilde$-modules by using the partial Fourier-Laplace transform of GKZ-hypergeometric systems and the Brieskorn lattices associated to the Hodge filtrations. (See also {\rm} and {\rm}.) They conjectured some relation of the $\nbigrtilde$-modules with ${}^{\ast}_0\widehat{\nbign}^{(0,0,0)}$ and ${}^{\ast}_0\widehat{\nbigm}^{(-r,0,0)}$ {\rm\cite[Conjecture 6.13]{Reichelt-Sevenheck2}}. We review the construction in {\rm\S}, and we compare it with the $\nbigrtilde$-modules underlying our mixed twistor $\nbigd$-modules in Proposition {\rm}. (See also Proposition {\rm} and Corollary {\rm}.) In this way, we can verify their conjecture. In {\rm}, they also verified their conjecture in a different way. \hfill\qed \end{rem} \subsection{Application to toric local mirror symmetry} Recall that GKZ-hypergeometric systems $\GKZ^{\ord}(\nbiga,\Gamma,0)$ and the variants $\GKZ^{\ord}_{\nbigrtilde}(\nbiga,\Gamma,0)$ play important roles in the study of mirror symmetry of toric manifolds. We shall apply the general theory of mixed twistor $\nbigd$-modules to the study of toric local mirror symmetry. One of the important goals in the study of mirror symmetry is to obtain an isomorphism of Frobenius manifolds associated to a mirror pair of an $A$-model and a $B$-model. Roughly, a Frobenius manifold is a complex manifold with a holomorphic multiplication and a holomorphic inner product on the tangent bundle satisfying some compatibility conditions. The Frobenius manifold associated to the $A$-model contains much information on the genus $0$ Gromov-Witten invariants. The Frobenius manifold associated to the $B$-model contains much information on the generalized Hodge structure of the Landau-Ginzburg model. Hence, it is interesting to have such an isomorphism of the Frobenius manifolds. One of the most celebrated results is due to A. Givental who established it in the case of toric weak Fano manifolds. (See also .) Pursuing an analogue of such an isomorphism in the study of local mirror symmetry, Konishi and Minabe \cite{Konishi-Minabe-cubic, Konishi-Minabe,Konishi-Minabe2, Konishi-Minabe3} introduced the concept of mixed Frobenius manifold as a generalization of Frobenius manifold. Note that Frobenius manifolds do not appear at least naively in the local case. Roughly, a mixed Frobenius manifold is a complex manifold with a holomorphic multiplication, a holomorphic filtration on the tangent bundle, and inner products on the graded pieces with respect to the filtration, satisfying some compatibility conditions. It looks natural to expect that mixed Frobenius manifolds appear widely. Konishi and Minabe particularly studied the case of any toric weak Fano surface $S$. On the local $A$-side, they constructed a mixed Frobenius manifold from the genus $0$ local Gromov-Witten invariants of $S$. On the local $B$-side, they suggested that the expected mixed Frobenius manifolds should be related with a variation of mixed Hodge structure associated to the corresponding Landau-Ginzburg model. Then, it is natural to ask whether there really exists a mixed Frobenius manifold on the local B-side. Recently, in his master thesis, Y. Shamoto proved a reconstruction theorem of mixed Frobenius manifolds as a generalization of a reconstruction theorem of Frobenius manifolds due to C. Hertling and Y. Manin . Together with the description of the variation of mixed Hodge structure in , he proved the existence of mixed Frobenius manifolds associated to some toric local $B$-models, up to the ambiguity of the choice of inner products on the graded pieces. It is still interesting to ask how to choose pairings on the graded pieces, and how to obtain an isomorphism of mixed Frobenius manifolds associated to the local mirror pair. \vspace{.1in} In this paper, we study the expectation in a rough level. Instead of mixed Frobenius manifolds, we shall study mixed TEP-structures which will be explained in the next subsection. We shall obtain an isomorphism of mixed TEP-structures associated to a mirror pair of a local $A$-model and a local $B$-model. \subsubsection{Mixed TEP-structure} Recall that a TE-structure on a complex manifold $M$ in the sense of is a locally free $\nbigo_{\cnum_{\lambda}\times M}$-module $\nbigv$ with a meromorphic flat connection $\nabla:\nbigv\lrarr\nbigv\otimes \Omega^1_{\cnum_{\lambda}\times M} \bigl(\log(\{0\}\times M)\bigr) \otimes \nbigo\bigl(\{0\}\times M\bigr)$. If it is equipped with a perfect pairing $P:\nbigv\otimes j^{\ast}\nbigv\lrarr \lambda^n\nbigo_{\cnum_{\lambda}\times M}$ such that $(j^{\ast}P)(j^{\ast}a\otimes b) =(-1)^n P(b\otimes j^{\ast}a)$, then $(\nbigv,\nabla,P)$ is called a TEP-structure or more precisely ${\rm TEP}(n)$-structure on $M$. Here, $j:\cnum_{\lambda}\times M \lrarr \cnum_{\lambda}\times M$ is given by $j(\lambda,Q)=(-\lambda,Q)$. We shall often omit to denote $\nabla$, i.e., $(\nbigv,\nabla,P)$ is denoted by $(\nbigv,P)$. A mixed TE-structure on a complex manifold $M$ consists of the following: \begin{itemize} \item A TE-structure $\nbigv$ on $M$. \item An increasing filtration $\Wtilde=\bigl(\Wtilde_m(\nbigv)\,|\,m\in\seisuu\bigr)$ on $\nbigv$ such that (i) $\Wtilde_m=0$ $(m<\!<0)$ and $\Wtilde_{m}=\nbigv$ $(m>\!>0)$, (ii) $\Gr^{\Wtilde}_m(\nbigv)$ are locally free $\nbigo_{\cnum_{\lambda}\times M}$-modules, (iii) $\Wtilde_m$ are preserved by the connection of $\nbigv$. \end{itemize} If each $\Gr^{\Wtilde}_m(\nbigv)$ is equipped with a non-degenerate pairing $P_m:\Gr^{\Wtilde}_m(\nbigv) \otimes j^{\ast}\Gr^{\Wtilde}_m(\nbigv) \lrarr \lambda^{-m}\nbigo_{\cnum_{\lambda}\times M}$ such that $(\Gr^{\Wtilde}_m\nbigv,P_m)$ is a ${\rm TEP}(-m)$-structure, then $(\nbigv,W,\{P_m\})$ is called a mixed TEP-structure. For example, a graded polarized variation of mixed Hodge structure naturally induces a mixed TEP-structure by the Rees construction. In that case, we can reconstruct the Hodge filtration, the weight filtration and the polarization on the graded pieces of the original graded polarized variation of mixed Hodge structure from the mixed TEP-structure. So, it still contains much interesting information. \subsubsection{Mixed twistor $\nbigd$-modules with a graded polarization} Let $(\nbigm,P_{\real},W)$ be a mixed twistor $\nbigd$-module on a complex manifold $Y$. If the $\nbigrtilde_Y$-module $\nbigm$ is a locally free $\nbigo_{\nbigy}$-module, then $(\nbigm,W)$ is a mixed TE-structure by definition. Suppose that $(\nbigm,P_{\real},W)$ is pure of weight $w$, i.e., $\Gr^W_j=0$ unless $j=w$. Then, a polarization of $(\nbigm,P_{\real},W)$ is equivalent to a morphism $\nbigm\lrarr \lambda^{-w}j^{\ast}\DDD(\nbigm)$, where $\DDD$ is the duality functor for mixed twistor $\nbigd$-modules. If $\nbigm$ is a locally free $\nbigo_{\nbigy}$-module, the morphism gives a pairing $\nbigp:\nbigm\otimes j^{\ast}\nbigm \lrarr \lambda^{-w+\dim Y}\nbigo_{\nbigy}$, and $(\nbigm,\nbigp)$ is a TEP$(-w+\dim Y)$-structure. If $(\nbigm,P_{\real},W)$ is not necessarily pure, a tuple of polarizations $\vecnbigp=(\nbigp_m\,|\,m\in\seisuu)$ on $\Gr^W_m(\nbigm,P_{\real})$ is called a graded polarization of $(\nbigm,P_{\real},W)$. If $\nbigm$ is a locally free $\nbigo_{\nbigy}$-module, each $\Gr^W_m(\nbigm)$ is a locally free $\nbigo_{\nbigy}$-module. We set $\Wtilde_{m-\dim Y}\nbigm:= W_{m}\nbigm$. Then, $(\nbigm,\Wtilde,\vecnbigp)$ is a mixed TEP-structure. \subsubsection{Mixed TEP-structures on $\GKZ(\nbiga,K(\nbiga),0)$ and $\GKZ(\nbiga,K(\nbiga)^{\circ},0)$} We use the notation in \S. We have the open subset $(\cnum^m)^{\reg}\subset \cnum^m$ determined by the regularity condition of the Laurent polynomials $F_{\nbiga}$. The restriction $M_{\nbiga,0,\star|(\cnum^m)^{\reg}}$ $(\star=\ast,!)$ are locally free $\nbigo$-modules. By the general theory of mixed twistor $\nbigd$-modules, $\nbigm_{\nbiga,0,\star}$ are locally free $\nbigo$-modules on $\cnum\times (\cnum^m)^{\reg}$. As previously remarked, we obtain the mixed TE-structure $(\nbigm_{\nbiga,0,\star|(\cnum^m)^{\reg}},W)$. If $0$ is an interior point of the convex hull of $\nbiga$, then it turns out that we have a natural isomorphism $(\nbigm_{\nbiga,0,!},P_{\real},W) \simeq (\nbigm_{\nbiga,0,\ast},P_{\real},W)$, that they are pure of weight $n+m$, and that they are equipped with a natural polarization. In particular, the restriction to $(\nbigm_{\nbiga,0,!},P_{\real},W)_{|(\cnum^m)^{\reg}}$ with the polarization give a TEP$(-n)$-structure. It is equivalent to the TEP-structure previously studied in the mirror symmetry . In the general case, we need to choose an additional datum for the construction of graded polarization on $(\nbigm_{\nbiga,0,!},P_{\real},W)$ and $(\nbigm_{\nbiga,0,\ast},P_{\real},W)$. It turns out that if we are given a point $\vecb\in\seisuu^n$ such that $0$ is an interior point of the convex hull of $\nbiga\cup\{\vecb\}$, then we obtain graded polarizations on $(\nbigm_{\nbiga,0,!},P_{\real},W)$ and $(\nbigm_{\nbiga,0,\ast},P_{\real},W)$ depending on $\vecb$. In particular, we obtain mixed TEP-structures $(\nbigm_{\nbiga,0,\star|(\cnum^m)^{\reg}}, \Wtilde,\vecnbigp_{\vecb,\star})$ on $(\cnum^m)^{\reg}$. We have the action of $T^n$ on $(\cnum^{\ast})^m$ given by $(s_1,\ldots,s_n)(z_1,\ldots,z_m) =(s^{\veca_1}z_1,\ldots,s^{\veca_m}z_m)$. Let $S_{\nbiga}:=(\cnum^{\ast})^m/T^n$. Let $S_{\nbiga}^{\reg}$ denote the image of $(\cnum^{\ast})^m\cap(\cnum^m)^{\reg}$ by the projection $(\cnum^{\ast})^m\lrarr S_{\nbiga}$. It turns out that $(\nbigm_{\nbiga,0,\star},\Wtilde,\vecnbigp_{\vecb,\star})$ on $(\cnum^{\ast})^m\cap(\cnum^m)^{\reg}$ is equivariant with respect to the action. So, we obtain mixed TEP-structures $(\nbigv_{\nbiga},\Wtilde,\vecnbigp_{\vecb,\star})$ $(\star=\ast,!)$ on $S_{\nbiga}^{\reg}$. \subsubsection{An isomorphism} Let $X$ be an $n$-dimensional smooth projective toric variety corresponding to a fan $\Sigma$. Let $D_i$ $(i=1,\ldots,m)$ be the hypersurfaces of $X$ corresponding to the one dimensional cones in $\Sigma$. Let $K_X$ denote the canonical bundle of $X$. Let $\nbigl_i$ $(i=1,\ldots,r)$ be nef line bundles on $X$ such that $(K_X\otimes\bigotimes_{i=1}^{r}\nbigl_i)^{\lor}$ is nef. We may assume that $\nbigl_j=\nbigo\bigl(\sum_{i=1}^m \beta_{ji}D_i\bigr)$ for some $\beta_{ji}\geq 0$. Let $Y$ be the projective completion of $\nbige^{\lor}:=\bigoplus_{i=1}^r\nbigl^{\lor}_i$, i.e., $Y=\proj\bigl( \bigoplus\nbigl_i\oplus\nbigo \bigr)$. We have the degenerated quantum products $\bullet_{\sigma}$ $(\sigma\in\nbigu_X)$ on $H^{\ast}(X,\cnum)$ induced by the Gromov-Witten invariants of $Y$, as explained in \S. Here, $\nbigu_X$ is an appropriate open subset in $H^{2}(X,\cnum)$. As in the ordinary case, we have the associated TE-structure $\QDM\bigl(X,\nbige^{\lor}\bigr)$ on $\nbigu_X$. Let $\gamma\in H^2(Y)$ be the first Chern class of the tautological line bundle of $Y$ over $X$. As explained in \S, we introduce a filtration $\Wtilde$ on $H^{\ast}(X,\cnum)$, and pairings on $\Gr^{\Wtilde}_jH^{\ast}(X,\cnum)$ by using the action of $\gamma$ on $H^{\ast}(Y)$ and inner product on $H^{\ast}(Y)$. The construction was motivated by both the general theory of mixed twistor $\nbigd$-modules and the construction of Konishi and Minabe . Thus, we obtain a mixed TEP-structure $(\QDM(X,\nbige^{\lor}),\Wtilde,\vecnbigp^{\nbige^{\lor}})$ on $\nbigu_X$. It is equivariant with respect to the translation action of $2\pi\sqrt{-1}H^2(X,\seisuu)$ on $\nbigu_X$. So, we obtain a mixed TEP-structure $(\QDM(X,\nbige^{\lor}),\Wtilde,\vecnbigp^{\nbige^{\lor}})'$ on $\nbigu_X/2\pi\sqrt{-1}H^2(X,\seisuu)$. \begin{rem} As explained in {\rm\S}, the degenerated quantum products are related with the local Gromov-Witten invariants in some special cases. It was essentially given in {\rm }, and used in {\rm} for the construction of the weight filtration and the pairings. \hfill\qed \end{rem} Let $\nbiga(\nbige^{\lor})\subset \seisuu^{n+r}$ be the set of the primitive vectors in the one dimensional cones in the fan corresponding to $\nbige^{\lor}$. Let $\nbiga(X)=\{[\rho_i]\,|\,i=1,\ldots,m\}\subset\seisuu^n$ be the set of the primitive vectors in the one dimensional cones in the fan $\Sigma$ corresponding to $X$. In this case, $\nbiga(\nbige^{\lor})$ consists of elements $\veca_i$ $(i=1,\ldots,m+r)$ given as follows: \[ \veca_i:= \left\{ \begin{array}{ll} [\rho_i]+\sum_{j=1}^r\beta_{ji}n_j & (i=1,\ldots,m)\\ n_{i-m} & (i=m+1,\ldots,m+r) \end{array} \right. \] Here, $n_i=(\overbrace{0,\ldots,0}^{m+i-1},1,0,\ldots,0)$. We also set $\veca_{m+r+1}:=-\sum_{i=1}^r n_i$. Then, we have the mixed TEP-structure $(\nbigm_{\nbiga(\nbige^{\lor})},\Wtilde,\vecnbigp_{\veca_{m+r+1},\ast})$ on $(\cnum^{m+r})^{\reg}$. We obtain a mixed TEP-structure $(\nbigv_{\nbiga(\nbige^{\lor})},\Wtilde,\vecnbigp_{\veca_{m+r+1},\ast})$ on $S_{\nbiga(\nbige^{\lor})}^{\reg}$ as the reduction. \begin{thm} We have the following. \begin{itemize} \item An open subset $U_1\subset H^2(X,\cnum)/2\pi\sqrt{-1}H^2(X,\seisuu)$ which contains a neighbourhood of the large radius limit point. \item An open subset $U_2\subset S_{\nbiga(\nbige^{\lor})}^{\reg}$ which contains a neighbourhood of the large radius point. \item A holomorphic isomorphism $\varphi:U_1\simeq U_2$. \item Isomorphism of mixed TEP-structures \[ \bigl( \QDM(X,\nbige^{\lor}), \Wtilde, \vecnbigp^{\nbige^{\lor}} \bigr)'_{|U_1}\,\, \quad\mbox{\rm and} \quad \varphi^{\ast} \bigl( \nbigv_{\nbiga(\nbige^{\lor}),\ast}, \Wtilde, \vecnbigp_{\veca_{m+r+1},\ast} \bigr)_{|U_2} \] up to shift of weights. \end{itemize} \end{thm} See Theorem for a more refined and precise claim. \begin{rem} In {\rm}, it is announced that a result in {\rm } implies the comparison of the TE-structures in Theorem {\rm}. In {\rm}, the mirror theorem of Givental is generalized for non-compact or non-semipositive toric manifolds. At this moment, it is not clear to the author if we could also deduce the comparison of the weight filtrations from {\rm }. \hfill\qed \end{rem} \subsubsection{Idea of the proof} Roughly, our isomorphism is obtained as the specialization of the isomorphism of Givental for the weak Fano toric manifold $Y$. We naturally have $H^2(Y,\cnum)=H^2(X,\cnum)\times\cnum\gamma$. We have the quantum $\nbigd$-module $(\QDM(Y),\nbigp_Y)$ on an appropriate open subset of $H^2(Y,\cnum)\big/2\pi\sqrt{-1}H^2(Y,\seisuu)$ associated to the Gromov-Witten invariants of $Y$. The mixed TEP-structure $\bigl(\QDM(X,\nbige^{\lor}), \Wtilde,\vecnbigp^{\nbige^{\lor}}\bigr)$ is described as the ``specialization'' of $(\QDM(Y),\nbigp_Y)$. (We explain the procedure ``specialization'' in \S.) Let $\nbiga(Y):= \nbiga(\nbige^{\lor})\cup\{\veca_{m+r+1}\}$ in $\seisuu^{n+r}$, which is the set of the primitive vectors in the one dimensional cones in a fan corresponding to $Y$. We naturally have $S_{\nbiga(Y)}=S_{\nbiga(\nbige^{\lor})}\times\cnum^{\ast}$. By using the results on the specialization of mixed twistor $\nbigd$-modules, we can describe the mixed TEP-structure $\bigl(\nbigv_{\nbiga(\nbige^{\lor})\ast}, \Wtilde,\vecnbigp_{\veca_{m+r+1}\ast} \bigr)$ as ``the specialization'' of the TEP-structure $(\nbigv_{\nbiga(Y)},\nbigp)$. We have the isomorphism of Givental between the TEP-structures $(\QDM(Y),\nbigp_Y)$ and $(\nbigv_{\nbiga(Y)},\nbigp)$. (See .) Hence, we can obtain the isomorphism of mixed TEP-structures in Theorem as the specialization of the isomorphism of Givental. \subsection{The mixed twistor $\nbigd$-modules associated to meromorphic functions} Motivated by the applications to the study of the twistor property of GKZ-hypergeometric systems, mentioned in the previous subsections, we shall consider the technical issues on the mixed twistor $\nbigd$-modules. Let $X$ be a complex manifold with a hypersurface $H$. Let $f$ be a meromorphic function on $X$ whose poles are contained in $H$. We have the naturally defined integrable mixed twistor $\nbigd$-modules with real structure $\nbigt_{\star}(f,H)$ $(\star=\ast,!)$ over $L_{\star}(f,H)$. We also have the naturally induced polarizations on $\Gr^W_w\nbigt_{\star}(f,H)$ which depend on the choice of an effective divisor $D$ whose support is $H$. The much part of this paper is devoted to the study of such kind of graded polarized mixed twistor $\nbigd$-modules with real structure. For instance, we consider the following issues. \begin{itemize} \item Let $X^{(1)}:=\proj^1_{\tau}\times X$. We set $H^{(1)}:=(\proj^1\times H)\cup(\{0,\infty\}\times X)$. Let $\iota:X\lrarr X^{(1)}$ be the morphism induced by $\{0\}\lrarr \proj^1_{\tau}$. Let $f,g\in \nbigo_X(\ast H)$. We have the graded polarized mixed twistor $\nbigd$-modules with real structure $\nbigt_{\star}(g+\tau f,H^{(1)})$ on $X^{(1)}$ and $\nbigt_{\star}(g,H)$ on $X$, where $\star=\ast,!$. Under some assumptions, we shall relate $\iota_{\dagger}\nbigt_{\star}(g,H)$ with the kernel and the cokernel of $\nbigt_{!}(g+\tau f,H^{(1)}) \lrarr \nbigt_{\ast}(g+\tau f,H^{(1)})$. It is not difficult to obtain a relation in the level of mixed twistor $\nbigd$-modules with real structures (Proposition ). We need more preliminaries to obtain the relation between polarizations on the graded pieces (Proposition ). \item Let $X^{(1)}$ and $H^{(1)}$ be as above. Let $\pi:X^{(1)}\lrarr X$ be the projection. We have the graded polarized mixed twistor $\nbigd$-modules $\pi^0_{\dagger}\nbigt_{\star}(g+\tau f,H^{(1)})$ by taking the push-forward via $\pi$. Let $Z_f$ be the zero-set of $f$. We have the graded polarized mixed twistor $\nbigd$-modules $\nbigt_{\ast}(g,H)[!(f)_0]$ and $\nbigt_{!}(g,H)[\ast(f)_0]$ on $X$. Under some assumptions, we shall relate $\pi^0_{\dagger}\nbigt_{\ast}(g+\tau f,H^{(1)})$ (resp. $\pi^0_{\dagger}\nbigt_{!}(g+\tau f,H)$) and $\nbigt_{\ast}(g,D)[!Z_f]$ (resp. $\nbigt_{!}(g,D)[\ast Z_f]$.) Again, it is not difficult to obtain a relation in the level of mixed twistor $\nbigd$-modules with real structure (Proposition ). We need more preliminaries to obtain the relation between polarizations on the graded pieces (Proposition ). \end{itemize} We also need the compatibility of various standard functors for $\nbigd$-modules and flat bundles, which are explained in the appendix sections \S and \S. \paragraph{Acknowledgement} This study grows out of my attempt to understand the works of H. Iritani , Y. Konishi, S. Minabe and T. Reichelt, C. Sevenheck . I am grateful to them for helpful discussions on various occasions. In particular, Reichelt and Sevenheck explained their notation, and remarked me the issue of the normality condition in the study of GKZ-hypergeometric systems and toric varieties. I thank C. Sabbah for his kindness and for discussions on many occasions. I thank Y. Shamoto for some discussion. I thank G. Wilkin for his interest to this study and his kindness. I am grateful to C. Hertling and U. Walther for some discussions. I thank A. Ishii, A. Moriwaki, M.-H. Saito, Y. Tsuchimoto, T. Xue and K. Vilonen for their kindness and supports. This work is partially supported by the Grant-in-Aid for Scientific Research (C) (No. 22540078), the Grant-in-Aid for Scientific Research (C) (No. 15K04843), the Grant-in-Aid for Scientific Research (A) (No. 22244003) and the Grant-in-Aid for Scientific Research (S) (No. 24224001), Japan Society for the Promotion of Science. This research was partially completed while the author was visiting the Institute for Mathematical Sciences, National University of Singapore in 2014. This research was partially completed while the author stayed in the Mathematisches Forschungsinstitut Oberwolfach, partially supported by the Simons Foundation and by the Mathematisches Forschungsinstitut Oberwolfach. \subsection{Purity condition} \subsubsection{$\nbigd$-modules associated to meromorphic functions} Let $X$ be a complex manifold with a hypersurface $D$. Let $\nbigo_X(\ast D)$ be the sheaf of meromorphic functions on $X$ whose poles are contained in $D$. Let $\DDD$ denote the duality functor on the category of holonomic $\nbigd$-modules on $X$. For any coherent $\nbigd_X$-module $M$, we set $M(\ast D):=\nbigo_X(\ast D)\otimes_{\nbigo_X}M$ and $M(!D):=\DDD\bigl(\DDD(M)(\ast D)\bigr)$. The restriction of $M(\star D)$ $(\star =\ast,!)$ to $X\setminus D$ is naturally isomorphic to $M_{|X\setminus D}$. We have the canonical morphism $M(!D)\lrarr M(\ast D)$ whose restriction to $X\setminus D$ is the identity. Let $f$ be a meromorphic function on $(X,D)$, i.e., a section of $\nbigo_X(\ast D)$. Let $(f)_0$ and $(f)_{\infty}$ denote the effective divisors obtained as the zeroes and the poles of $f$. The supports of the divisors are denoted by $|(f)_0|$ and $|(f)_{\infty}|$. We obtain the meromorphic flat bundle $L_{\ast}(f,D):=\nbigo_X(\ast D)v$ with $\nabla v=v\,df$. We naturally regard it as a $\nbigd_X$-module. We set $L_!(f,D):=L_{\ast}(f,D)(!D) =\DDD\bigl(L_{\ast}(-f,D)\bigr)$. The image of the canonical morphism $L_!(f,D)\lrarr L_{\ast}(f,D)$ is independent of the choice of $D$ such that $|(f)_{\infty}|\subset D$, and denoted by $L(f)$. We have natural isomorphisms $L_{\star}(f,D)\simeq L(f)(\star D)$ for $\star=!,\ast$. Indeed, $L(f)\lrarr L_{\ast}(f,D)$ naturally induces $L(f)(\ast D)\lrarr L_{\ast}(f,D)$ which is clearly an isomorphism. By using the duality, we obtain $L_!(f,D)\simeq L(f)(!D)$. When $D=|(f)_{\infty}|$, we set $L_{\star}(f):=L_{\star}(f,|(f)_{\infty}|)$. \subsubsection{Purity} We continue to use the notation in \S. We introduce a condition. \begin{df} We say that a meromorphic function $f$ on $X$ is pure at $P\in X$ if the canonical morphism $L_{!}(f)\lrarr L_{\ast}(f)$ is an isomorphism on a neighbourhood of $P$. We say that $f$ is pure if it is pure at any point of $X$. \hfill\qed \end{df} Because $\DDD \bigl(L_{\ast}(f)\bigr) \simeq L_!(-f)$ and $\DDD \bigl(L_{!}(f)\bigr) \simeq L_{\ast}(-f)$, we have the following easy lemmas. \begin{lem} Let $f$ be a meromorphic function on $X$. \begin{itemize} \item $f$ is pure if and only if $-f$ is pure. \item $f$ is pure if and only if the morphisms $L_!(f)\lrarr L_{\ast}(f)$ and $L_{!}(-f)\lrarr L_{\ast}(-f)$ are epimorphisms. \hfill\qed \end{itemize} \end{lem} Let $\varphi:Y\lrarr X$ be a proper morphism of complex manifolds such that $\varphi$ induces an isomorphism $Y\setminus \varphi^{-1}(D)\simeq X\setminus D$. \begin{lem} Let $f$ be a meromorphic function on $(X,D)$. Suppose that $\bigl|(\varphi^{\ast}f)_{\infty}\bigr|=\varphi^{-1}(D)$ and that $\varphi^{\ast}(f)$ is pure. Then, $f$ is also pure, and it satisfies $|(f)_{\infty}|=D$. \end{lem} \pf We set $f_Y:=\varphi^{\ast}(f)$ and $D_Y:=\varphi^{-1}(D)$. Because $D=\varphi\bigl(|(f_Y)_{\infty}|\bigr) \subset |(f)_{\infty}|\subset D$, we have $|(f)_{\infty}|=D$. We have $\varphi^i_{+}L_{\ast}(f_Y,D_Y)=0$ $(i\neq 0)$, and $\varphi^0_{+}\bigl( L_{\ast}(f_Y,D_Y)\bigr)(\ast D) \simeq \varphi^0_{+}\bigl( L_{\ast}(f_Y,D_Y)\bigr)$. Hence, we have $\varphi^0_{+}\bigl( L_{\ast}(f_Y,D_Y)\bigr) \simeq L_{\ast}(f,D)$. By the duality, we have $\varphi^0_{+}\bigl( L_!(f_Y,D_Y)\bigr) \simeq L_!(f,D)$. Then, from the purity of $f_Y$, we obtain that $L_!(f,D)\lrarr L_{\ast}(f,D)$ is an isomorphism. \hfill\qed \subsubsection{Vanishing of cohomology for pure functions} The purity condition sometimes implies the vanishing of the cohomology. We mention a typical case. Let $X$ be a complex manifold with a hypersurface $D$. Let $f$ be a meromorphic function on $(X,D)$ such that (i) $|(f)_{\infty}|=D$, (ii) $f$ is pure. Let $F:X\lrarr S$ be a proper morphism of complex manifolds. \begin{prop} Suppose that $R^iF_{\ast}\bigl( \Omega_X^j\otimes\nbigo(\ast D) \bigr)=0$ for $(i,j)\in\seisuu_{\geq 0}^2$ such that $i+j>\dim X$. Then, we have $F_{\dagger}^kL(f)=0$ $(k\neq 0)$. \end{prop} \pf Because $\DDD L(f)\simeq L(-f)$ and because $-f$ is also pure, it is enough to prove that $F_{\dagger}^kL(f)=0$ for $k>0$. We have only to prove the claim locally around any point of $S$. We set $\omega_X:=\Omega_X^{\dim X}$ and $\omega_S:=\Omega_S^{\dim S}$. Recall that $F_{\dagger}L(f)$ is obtained as $RF_{\ast}\Bigl( \bigl[ \omega_X \otimes_{F^{-1}\nbigo_S} F^{-1}\bigl( \nbigd_S\otimes \omega_S^{-1} \bigr) \bigr] \otimes^L_{\nbigd_X} L(f) \Bigr)$. We have the standard free resolution $\nbigd_X\otimes\Omega^{\bullet}_X[\dim X]$ of the right $\nbigd_X$-modules $\omega_X$. By the assumption, if $i+j>\dim X$, we have \[ R^jF_{\ast}\Bigl( \bigl[ \Omega^i_X \otimes_{F^{-1}\nbigo_S}(F^{-1}\nbigd_S\otimes\omega_S) \bigr] \otimes_{\nbigo_X} L(f) \Bigr) \simeq R^jF_{\ast}\bigl( \Omega^i_X\otimes L(f) \bigr) \otimes_{\nbigo_S} (\nbigd_S\otimes\omega_S) =0. \] Hence, we have the desired vanishing $F_{\dagger}^kL(f)=0$ for $k>0$. \hfill\qed \begin{example} The condition of Proposition {\rm} is satisfied if $F$ is factorized into the composite of morphisms of complex manifolds $X\stackrel{\rho}{\lrarr} X'\stackrel{F'}{\lrarr}S$ such that (i) $D'=\rho(D)$ is also a hypersurface of $X'$, (ii) $\rho$ induces $X\setminus D\simeq X'\setminus D'$, (iii) $\nbigo_{X'}(D')$ is relatively ample with respect to $F'$. \hfill\qed \end{example} \subsection{$\nbigd$-modules associated to non-degenerate meromorphic functions} \subsubsection{A non-degeneracy condition} Let $X$ be a complex manifold with a simple normal crossing hypersurface $D$. Let $f$ be a meromorphic function on $(X,D)$. In this paper, we shall often use the following non-degeneracy condition. \begin{df} $f$ is called non-degenerate along $D$ if the following holds for a small neighbourhood $N$ of $|(f)_{\infty}|$. \begin{itemize} \item $(f)_0\cap N$ is reduced and non-singular. \item $N\cap\bigl(|(f)_0|\cup D\bigr)$ is normal crossing. \hfill\qed \end{itemize} \end{df} Let $D=\bigcup_{i\in\Lambda}D_i$ and $|(f)_{\infty}|=\bigcup_{i\in\Lambda_f}D_i$ be the irreducible decompositions, where $\Lambda_f\subset\Lambda$. For any non-empty subset $I\subset \Lambda$, we set $D_I:=\bigcap_{i\in\Lambda}D_I$ and $D_I^{\circ}:=D_I\setminus \bigcup_{j\not\in I}D_j$. If $I=\emptyset$, we set $D_{\emptyset}=X$. Then, the second condition can be reworded that $D_I^{\circ}$ is transversal with $|(f)_0|$ for any $I\subset \Lambda$ with $I\cap \Lambda_f\neq\emptyset$. \begin{rem} Suppose that a meromorphic function $f$ on $(X,D)$ is non-degenerate along $D$. Let $D'$ be a hypersurface of $X$ such that $|(f)_{\infty}|\subset D'\subset D$. Then, $f$ is non-degenerate along $D'$. But, the converse does not hold in general. Namely, even if a meromorphic function $f$ on $(X,D')$ is non-degenerate along $D'$, it is not necessarily non-degenerate along $D$. For example, set $X=\cnum^2$, $f=(z_1-z_2)/z_2$, $D'=\{z_2=0\}$ and $D=\{z_1=0\}\cup\{z_2=0\}$. \hfill\qed \end{rem} We reword the condition in terms of local coordinate systems. We set $D^c_f:=\bigcup_{i\in\Lambda_f^c}D_i$, where $\Lambda_f^c:= \Lambda\setminus \Lambda_f$. We have $D=|(f)_{\infty}|\cup D_f^c$. Let $Q\in |(f)_{\infty}|$. We take a holomorphic coordinate neighbourhood $(X_Q;z_1,\ldots,z_n)$ around $Q$ such that \[ |(f)_{\infty}|\cap X_Q= \bigcup_{i=1}^{\ell_1}\{z_i=0\}, \quad D_f^c\cap X_Q= \bigcup_{i=\ell_1+1}^{\ell_1+\ell_2}\{z_i=0\}. \] Let $k_i$ denote the pole order of $f$ along $\{z_i=0\}$. Then, we have an expression $f=f_0\prod_{i=1}^{\ell_1}z_i^{-k_i}$ where $f_0$ is holomorphic. Let $I\subset\Lambda_f$ be determined by $Q\in D_I^{\circ}$. Note $I\cap \Lambda_f\neq\emptyset$. If $f$ is non-degenerate along $D$, then the divisor $(f_{0|D_I^{\circ}})$ is reduced and non-singular. Conversely, if the above holds for any $Q\in |(f)_{\infty}|$, then $f$ is non-degenerate. \begin{rem} See Definition {\rm} and Lemma {\rm} for the relation between the non-degeneracy condition in {\rm} and the condition in Definition {\rm}. We postpone to discuss the exact relation between the (cohomologically) tameness condition for algebraic functions (see {\rm}) and the conditions in Definitions {\rm} and {\rm}. \hfill\qed \end{rem} \subsubsection{Convenient coordinate systems} Suppose that a section $f$ of $\nbigo_X(\ast D)$ is non-degenerate along $D$. We have a holomorphic coordinate system $(X_Q;z_1,\ldots,z_n)$ around $Q\in|(f)_{\infty}|$ satisfying the following conditions. \begin{itemize} \item $|(f)_{\infty}|\cap X_Q =\bigcup_{i=1}^{\ell_1}\{z_i=0\}$ and $D_f^c\cap X_Q =\bigcup_{i=\ell_1+1}^{\ell_1+\ell_2}\{z_i=0\}$. \item If $Q\in |(f)_0|$, we have $f_{|X_Q}=z_n\prod_{i=1}^{\ell_1}z_i^{-k_i}$ for some $\veck\in\seisuu_{> 0}^{\ell_1}$. \item If $Q\not\in |(f)_0|$, we have $f_{|X_Q}=\prod_{i=1}^{\ell_1}z_i^{-k_i}$ for some $\veck\in\seisuu_{> 0}^{\ell_1}$. \end{itemize} In this paper, such a coordinate system is called a convenient coordinate system. \subsubsection{Families of non-degenerate functions} Let $X\lrarr S$ be a smooth morphism of complex manifolds. Let $D$ be a simple normal crossing hypersurface in $X$ with the irreducible decomposition $D=\bigcup_{i\in\Lambda}D_i$. Suppose that the induced morphisms $D^{\circ}_I\lrarr S$ are smooth for any $I\subset\Lambda$. For any $s\in S$, let $(X_s,D_s)$ denote the fibers of $(X,D)$ over $s\in S$. In such a situation, we sometimes consider a condition which is stronger than the non-degeneracy along $D$. \begin{df} Let $f$ be a meromorphic function on $(X,D)$ which is non-degenerate along $D$. If moreover $f_{|X_s}$ is non-degenerate along $D_s$ for any $s\in S$, we say that $f$ is non-degenerate along $D$ over $S$. \hfill\qed \end{df} \subsubsection{Purity in the non-degenerate case} Let $X$ be a complex manifold with a normal crossing hypersurface $D$. Let $f$ be a meromorphic function on $(X,D)$. \begin{lem} Suppose that $D=|(f)_{\infty}|$ and that $f$ is non-degenerate along $D$. Then, $f$ is pure. \end{lem} \pf We have only to check the claim locally around any point of $|(f)_{\infty}|$. We use a convenient coordinate system as in \S. We have a natural isomorphism $L_!(f)(\ast D) \simeq L_{\ast}(f)(\ast D)$. Hence, for a large $N$, $u:=\prod_{i=1}^{\ell_1}z_i^{N}v$ is a section of $L_!(f)$. We have a natural morphism $\varphi:\nbigd_X \lrarr L_!(f)$ given by $P\longmapsto P u$. Let us check that the composite of $\varphi$ and $L_!(f)\lrarr L_{\ast}(f)$ is an epimorphism. It is enough to observe that the image of $u$ in $L_{\ast}(f)$ generates $L_{\ast}(f)$. For $\vecm\in\seisuu^{m_1}$, let $z^{\vecm}=\prod_{i=1}^{\ell_1}z_i^{m_i}$. If $Q\in |(f)_{\infty}|$, note $\del_{z_n}(z^{\vecm}v) =z^{\vecm-\veck}v$. If $Q\not\in|(f)_{\infty}|$, note $z_1\del_{z_1}(z^{\vecm}v) =-k_iz^{\vecm-\veck}v +m_iz^{\vecm}v$. Then, the claim easily follows. Hence, the natural morphism $\kappa:L_!(f)\lrarr L_{\ast}(f)$ is an epimorphism. As remarked in Lemma , because the dual of $\kappa$ is the epimorphism $L_{!}(-f)\lrarr L_{\ast}(-f)$, we obtain that $\kappa$ is a monomorphism. \hfill\qed \subsubsection{Expressions of the $\nbigd$-modules associated to non-degenerate functions} Let $V_D\nbigd_X$ denote the sheaf of subalgebras of $\nbigd_X$ generated by $\Theta_X(\log D)$ over $\nbigo_X$. Let $f$ be a meromorphic function on $(X,D)$ which is non-degenerate along $D$. Then, $L(f)$ and $L(f)(D)=L(f)\otimes_{\nbigo_X}\nbigo_X(D)$ are naturally $V_D\nbigd_X$-modules. \begin{lem} We naturally have $L_!(f,D)\simeq \nbigd_X\otimes_{V_D\nbigd_X} L(f)$ and $L_{\ast}(f,D)\simeq \nbigd_X\otimes_{V_D\nbigd_X} L(f)(D)$. \end{lem} \pf Because $\bigl( \nbigd_X\otimes_{V_D\nbigd_X} L(f)\bigr)(\ast D) \simeq \bigl( \nbigd_X\otimes_{V_D\nbigd_X} L(f)(D)\bigr)(\ast D) \simeq L(f)(\ast D)$, we have natural morphisms: \begin{equation} \nbigd_X\otimes_{V_D\nbigd_X}L(f)(D) \stackrel{\alpha_1}{\lrarr} L_{\ast}(f,D) \end{equation} \begin{equation} L_!(f,D) \stackrel{\alpha_2}{\lrarr} \nbigd_X\otimes_{V_D\nbigd_X}L(f). \end{equation} It is enough to prove that the morphisms are isomorphisms locally around any point of $D$. We use a convenient coordinate system in \S. Let $v$ be a frame of $L(f)$ over $\nbigo_X(\ast(f)_{\infty})$ such that $\nabla v=v\,df$. Because $a_1:L_!(f,D)\lrarr L(f)$ is an epimorphism, we can locally take a section $v'$ of $L_!(f,D)$ which is mapped to $v$ via $a_1$. We consider the submodule $V_D\nbigd_X\,v'\subset L_!(f,D)$. It is coherent over $V_D\nbigd_X$. Note that $L(f)$ is coherent over $V_D\nbigd_X$. Because $\Ker a_1\cap V_D\nbigd_Xv' =\Ker\bigl( a_{1|V_D\nbigd_Xv'} \bigr)$, it is coherent over $V_D\nbigd_X$. In particular, it is locally finitely generated over $V_D\nbigd_X$. Take a generator $f_1,\ldots,f_m$ of $\Ker a_1\cap V_D\nbigd_Xv'$. The supports of $f_j$ are contained in $D$. We can take a large $N$ such that $\prod_{i=1}^{\ell_1+\ell_2} z_i^{N}\,f_j=0$ in $L_!(f,D)$, because $L_!(f,D)(\ast D)=L(f)(\ast D)$. Then, it is easy to see that $\prod_{i=1}^{\ell_1+\ell_2}z_i^N\, \bigl( V_D\nbigd_X\,v'\cap\Ker(a_1) \bigr)=0$. We set $v'':=\prod_{i=1}^{\ell_1+\ell_2}z_i^Nv'$, and then we have $V_D\nbigd_Xv''\cap\Ker(a_1)=0$. The morphism $V_D\nbigd_Xv''\lrarr L(f)$ induced by $a_1$ gives an isomorphism $V_D\nbigd_Xv'' \lrarr \nbigo_X(\ast (f)_{\infty})\prod_{i=1}^{\ell_1+\ell_2}z_i^Nv$ of $V_D\nbigd_X$-modules. In particular, $V_D\nbigd_Xv''$ is naturally an $\nbigo_X\bigl(\ast(f)_{\infty}\bigr)$-module. We set $v^{(3)}:=\prod_{i=1}^{\ell_1}z_i^{-N}\,v''$. We have $z_i\del_iv^{(3)} =-k_iv^{(3)}\prod_{i=1}^{\ell_1}z_i^{-k_i}$ for $i=1,\ldots,\ell_1$, and $z_i\del_iv^{(3)} =Nv^{(3)}$ for $i=\ell_1+1,\ldots,\ell_1+\ell_2$. Take $0\leq p_{\ell_1+1},\ldots,p_{\ell_1+\ell_2}\leq N$. We set $\vecp:=(p_{\ell_1+1},\ldots,p_{\ell_1+\ell_2})$. We set $v_{\vecp}:= \prod_{i=\ell_1+1}^{\ell_1+\ell_2} \del_i^{p_i}v^{(3)}$. Note that $a_1(v_{\vecp}) =C_{\vecp}\prod_{i=\ell_1+1}^{\ell_1+\ell_2} z_i^{N-p_i}v$ for a non-zero constant $C_{\vecp}$. We consider the following morphism induced by $a_1$: \begin{equation} V_D\nbigd_X\cdot \prod_{i=\ell_1+1}^{\ell_1+\ell_2} \del_i^{p_i}v^{(3)} \lrarr L(f) \end{equation} Let us observe that () is a monomorphism. If $\vecp=(0,\ldots,0)$, it has already been observed. If $p_i>0$, set $p_j':=p_j$ $(j\neq i)$ and $p_i':=p_i-1$. We have $\del_iv_{\vecp'}=v_{\vecp}$. Let $s\in\Ker(a_1)\cap V_D\nbigd_Xv_{\vecp}$. We have $s=\del_is'$ for some $s'\in V_D\nbigd_Xv_{\vecp'}$. We have $0=a_1(z_is)=a_1(z_i\del_is')=0$. We have $z_i\del_is'=0$ in $V_D\nbigd_X v_{\vecp'} \simeq \nbigo_X(\ast (f)_{\infty}) \prod_{j=1}^{N-p_j'}v$. But, because $N-p_i'>0$, it is easy to see that if a section $s'$ of $\nbigo_X(\ast (f)_{\infty}) \prod_{j=1}^{N-p_j'}v$ satisfies $z_i\del_is'=0$ then $s'=0$. Hence, we obtain $s=0$. In particular, the induced morphism $V_D\nbigd_X\,v_{(N,\ldots,N)} \lrarr L(f)$ is a monomorphism. Because $a_1(v_{(N,\ldots,N)})$ is $v$ multiplied by a non-zero constant, $V_D\nbigd_X\,v_{(N,\ldots,N)} \lrarr L(f)$ is also an epimorphism, i.e., an isomorphism. Hence, we have a $V_D\nbigd_X$-homomorphism $a_2:L(f)\lrarr L_!(f,D)$ such that $a_1\circ a_2=\id$. It induces a $\nbigd_X$-homomorphism $\beta_2:\nbigd_X\otimes_{V_D\nbigd_X} L(f) \lrarr L_!(f,D)$. We set $g:=\beta_2\circ\alpha_2$ which is an endomorphism of $L_!(f,D)$. For the dual, $g$ induces the identity on $\DDD L_!(f,D)$. Hence, we obtain that $g$ is the identity. In particular, () is an epimorphism and $\beta_2$ is a monomorphism. Because $\nbigd_X\otimes_{V_D\nbigd_X}L(f)$ is generated by $v$, $\beta_2$ is an epimorphism. Hence, $\alpha_2$ and $\beta_2$ are isomorphisms. Let us study the morphism $\alpha_1$. Because $L(f)=L(f)(\ast (f_{\infty}))$, we naturally have \[ \nbigd_X\otimes_{V_D\nbigd_X}L(f)(D) \simeq \nbigd_X(\ast(f)_{\infty}) \otimes_{V_D\nbigd_X(\ast (f)_{\infty})}L(f)(D). \] For any non-negative integer $m$, we consider the $\nbigo_X(\ast(f)_{\infty})$-homomorphism $\gamma_m:L(f)\bigl((m+1)D\bigr) \lrarr \nbigd_X\otimes_{V_D\nbigd_X}L(f)(D)$ determined by \[ \gamma_m\Bigl(\prod_{i=1}^{\ell_1}z_i^{-1} \prod_{i=\ell_1+1}^{\ell_1+\ell_2}z_i^{-m-1}v\Bigr) =(-1)^{\ell_2m}(m!)^{-\ell_2} \prod_{i=\ell_1+1}^{\ell_1+\ell_2} \del_i^{m}\otimes \prod_{i=1}^{\ell_1+\ell_2}z_i^{-1}v. \] Let $\iota_{m}:L(f)(mD)\lrarr L(f)((m+1)D)$ be the natural inclusion. Then, we have $\alpha_1\circ\gamma_{m}\circ\iota_m =\alpha_1\circ\gamma_{m-1}$. We obtain an $\nbigo_X(\ast (f)_{\infty})$-homomorphism $\gamma: L(f)(\ast D) \lrarr \nbigd_X\otimes_{V_D\nbigd_X}L(f)(D)$. By the construction, $\gamma$ is an epimorphism, and $\alpha_1\circ\gamma$ is the identity. Then, we obtain that $\alpha_1$ is an isomorphism. \hfill\qed \subsubsection{De Rham complexes} We give some complexes which are quasi-isomorphic to the de Rham complexes of $L_{\star}(f,D)$ when $f$ is non-degenerate along $D$. (See , , for the case $|(f)_0|\cap|(f)_{\infty}|=\emptyset$.) We set $d_X:=\dim X$. We have the natural complex of right $\nbigd_X$-modules $\Omega_X^{\bullet}\otimes_{\nbigo_X} \nbigd_X[d_X]$ which is a right locally $\nbigd_X$-free resolution of $\Omega_X:=\Omega^{d_X}_X$. We have the subcomplex $\Omega_X^{\bullet}(\log D)(-D)\otimes_{\nbigo_X} V_D\nbigd_X[d_X]$, which is a right locally $V_D\nbigd_X$-free resolution of $\Omega_X$. Indeed, the natural morphism $\Omega_X\otimes V_D\nbigd_X \lrarr \Omega_X$ induces a quasi-isomorphism $\Omega_X^{\bullet}(\log D)(-D)\otimes V_D\nbigd_X[d_X] \simeq \Omega_X$ of complexes of right $V_D\nbigd_X$-modules. Suppose that $f$ is non-degenerate along $D$. Because $L_{\ast}(f,D)\simeq \nbigd_X\otimes_{V_D\nbigd_X} L(f)(D)$ according to Lemma , we have the following natural isomorphisms: \begin{multline} \Omega_X\otimes^L_{\nbigd_X} L_{\ast}(f,D) \simeq \Omega_X\otimes^L_{V_D\nbigd_X} L(f)(D) \simeq \Omega_X^{\bullet}(\log D)(-D)\otimes L(f)(D)[d_X] \\ \simeq \Bigl( \Omega_X^{\bullet}(\log D)(\ast (f)_{\infty}),d+df \Bigr)[d_X] \end{multline} Because $L_!(f,D)\simeq \nbigd_X\otimes_{V_D\nbigd_X} L(f)$ according to Lemma , we have the following isomorphisms: \begin{multline} \Omega_X\otimes^L_{\nbigd_X} L_!(f,D) \simeq \Omega_X\otimes^L_{V_D\nbigd_X} L(f) \simeq \Omega_X^{\bullet}(\log D)(-D) \otimes L(f)[d_X] \\ \simeq \Bigl( \Omega_X^{\bullet}(\log D)(-D)(\ast (f)_{\infty}), d+df \Bigr)[d_X] \end{multline} \paragraph{Coherent expression} Let $H$ be any divisor in $X$. Let us consider the complex \[ \Omega^{\ell-1}_X(\log D)\bigl(H\bigr) \stackrel{a_1}{\lrarr} \Omega^{\ell}_X(\log D)\bigl(H+(f)_{\infty}\bigr) \stackrel{a_2}{\lrarr} \Omega^{\ell+1}_X(\log D)\bigl(H+2(f)_{\infty}\bigr), \] where $a_i$ are induced by the multiplication of $df$. \begin{lem} If $f$ is non-degenerate along $D$, we have $\Image a_1=\Ker a_2$ on a neighbourhood of $|(f)_{\infty}|$. \end{lem} \pf We have only to check the claim locally around any point $Q$ of $|(f)_{\infty}|$ by using a convenient local coordinate system. Let us consider the case $Q\in |(f)_0|\cap|(f)_{\infty}|$. Then, $f=z_n\prod_{i=1}^{\ell_1}z_i^{-k_i}$. We have the following local section of $\Omega_X^1(\log D)$: \[ \tau:= df\cdot \prod_{i=1}^{\ell_1}z_i^{k_i} =dz_n-\sum_{i=1}^{\ell_1} k_i z_n\frac{dz_i}{z_i} \] We have $\tau_{|Q}\neq 0$. We consider the following on a neighbourhood of $Q$: \[ \Omega^{\ell-1}_X(\log D) \stackrel{b_1}{\lrarr} \Omega^{\ell}_X(\log D) \stackrel{b_2}{\lrarr} \Omega^{\ell+1}_X(\log D) \] Here, $b_i$ are induced by the multiplication of $\tau$. Because $\tau_{|Q}\neq 0$, we have $\Image(b_1)=\Ker(b_2)$. Then, in the case $Q\in|(f)_0|$, the claim of the lemma follows. The case $Q\not\in|(f)_0|$ can be argued similarly, and it is well known. \hfill\qed \vspace{.1in} We set $\Omega_X^{\ell}(\log D,f):= \Omega_X^{\ell}(\log D)(\ell(f)_{\infty})$. Together with the differential $d+df$, we obtain a complex $\bigl( \Omega_X^{\bullet}(\log D,f),d+df \bigr)$. We also have $\bigl( \Omega_X^{\bullet}(\log D,f)(-D),d+df \bigr)$. We obtain the following from the previous lemma. \begin{lem} The following natural morphisms are quasi-isomorphisms: \[ \bigl( \Omega_X^{\bullet}(\log D,f),d+df \bigr) \lrarr \Bigl( \Omega_X^{\bullet}(\log D)(\ast (f)_{\infty}),d+df \Bigr) \] \[ \bigl( \Omega_X^{\bullet}(\log D,f)(-D),d+df \bigr) \lrarr \Bigl( \Omega_X^{\bullet}(\log D)(-D)(\ast (f)_{\infty}),d+df \Bigr) \] \hfill\qed \end{lem} Hence, we have the following natural quasi-isomorphisms: \[ \Omega_X\otimes^L_{\nbigd_X}L_{\ast}(f,D) \simeq \bigl( \Omega_X^{\bullet}(\log D,f),d+df \bigr)[d_X] \] \[ \Omega_X\otimes^L_{\nbigd_X}L_{!}(f,D) \simeq \bigl( \Omega_X^{\bullet}(\log D,f)(-D),d+df \bigr)[d_X] \] \paragraph{Kontsevich complexes} Let $\Omega_{X,f,D}^k$ denote the kernel of the following morphism induced by the multiplication of $df$: \[ \Omega^k_X(\log D) \stackrel{df}{\lrarr} \frac{\Omega^{k+1}_X(\log D)\bigl(!(f)_{\infty}\bigr)} {\Omega^{k+1}_X(\log D)}. \] \begin{lem} $\Omega_{X,f,D}^k$ are locally free. \end{lem} \pf We have $\Omega_{X,f,D}^k=\Omega_X^k(\log D)$ outside $|(f)_{\infty}|$. Locally around any point of $(f)_{\infty}$, we have local decompositions $\Omega_X^k(\log D) =A_k\oplus B_k$ such that the multiplication of $df$ induces $B_k\simeq A_{k+1}((f)_{\infty})$. Hence, we have $\Omega_{X,f,D}^k=A_k\oplus B_k(-(f)_{\infty})$, and the claim of the lemma follows. \hfill\qed \vspace{.1in} The multiplication of $df$ induces $df:\Omega^{k}_{X,f,D}\lrarr \Omega^{k+1}_{X,f,D}$. By the commutativity $[d,df]=0$, the exterior derivative induces $d:\Omega^k_{X,f,D}\lrarr \Omega^{k+1}_{X,f,D}$. Hence, we obtain the complex $(\Omega^{\bullet}_{X,f,D},d+df)$. We also obtain the complex $(\Omega^{\bullet}_{X,f,D}(-D),d+df)$. \begin{lem} The following natural morphisms are quasi-isomorphisms: \[ \bigl( \Omega^{\bullet}_{X,f,D},d+df \bigr) \lrarr \bigl( \Omega_X^{\bullet}(\log D,f),d+df \bigr) \] \[ \bigl( \Omega^{\bullet}_{X,f,D}(-D),d+df \bigr) \lrarr \bigl( \Omega_X^{\bullet}(\log D,f)(-D),d+df \bigr) \] \hfill\qed \end{lem} \subsubsection{The push-forward by a projection} Let us consider the case $(X,D)=(X_0,D_0)\times S$. Suppose that $X_0$ is compact. Let $f$ be a meromorphic function on $(X,D)$. Let $\pi:X\lrarr S$ denote the projection. Let $n:=\dim X_0$. \begin{lem} Suppose that $f$ is non-degenerate along $D$. We have the following natural isomorphisms: \begin{equation} \pi_{+}^i\bigl( L_{\ast}(f,D) \bigr) \simeq \hyperr^i\pi_{\ast}\bigl( \Omega^{\bullet+n}_{X/S}(\log D)(\ast (f)_{\infty}), d+df \bigr) \end{equation} \begin{equation} \pi_{+}^i\bigl( L_{!}(f,D) \bigr) \simeq \hyperr^i\pi_{\ast}\bigl( \Omega^{\bullet+n}_{X/S}(\log D)(-D)(\ast (f)_{\infty}), d+df \bigr) \end{equation} \end{lem} \pf Let us consider (). By Lemma , we have $L_{\ast}(f,D)\simeq \nbigd_X\otimes_{V_D\nbigd_X} L(f)(D)$. We have $\nbigd_X=\nbigd_{X_0}\boxtimes \nbigd_{S}$ and $V_D\nbigd_X=V_{D_0}\nbigd_{X_0}\boxtimes \nbigd_{S}$. Let $p:X\times X_0$ be the projection. We obtain \begin{multline} \pi_{+}\bigl( L_{\ast}(f,D) \bigr) \simeq R\pi_{\ast}\Bigl( p^{\ast}\Omega_{X_0} \otimes^L_{p^{\ast}\nbigd_{X_0}} L_{\ast}(f,D) \Bigr) \simeq R\pi_{\ast}\Bigl( p^{\ast}\Omega_{X_0} \otimes^L_{p^{\ast}V_D\nbigd_{X_0}} \bigl( L_{\ast}(f,D)\otimes\nbigo(D) \bigr) \Bigr) \\ \simeq R\pi_{\ast}\Bigl( \Omega^{\bullet+n}_{X/S}(\log D)(-D) \otimes_{\nbigo_X} \bigl( L_{\ast}(f,D)\otimes\nbigo(D) \bigr) \Bigr) \simeq R\pi_{\ast}\Bigl( \Omega^{\bullet+n}_{X/S}(\log D) \otimes_{\nbigo_X} L_{\ast}(f,D) \Bigr) \end{multline} It implies (). Similarly, we obtain () from the expression $L_!(f,D)\simeq \nbigd_X\otimes_{V_D\nbigo_X} \bigl(L(f)(-D)\bigr)$ in Lemma , as in (). \hfill\qed \vspace{.1in} Suppose moreover that $f$ is non-degenerate along $D$ over $S$. (See Definition for this stronger condition.) As in \S, we can naturally define the complexes $\bigl(\Omega^{\bullet}_{X/S}(\log D,f),d+df\bigr)$ and $\bigl( \Omega^{\bullet}_{X/S,f,D},d+df \bigr)$ in the relative setting. \begin{lem} If $f$ is non-degenerate along $D$ over $S$, then we have the following natural isomorphisms: \begin{equation} \pi_{+}^i\bigl( L_{\ast}(f,D) \bigr) \simeq \hyperr^i\pi_{\ast} \bigl(\Omega^{\bullet+n}_{X/S}(\log D,f),d+df\bigr) \simeq \hyperr^i\pi_{\ast} \bigl( \Omega^{\bullet+n}_{X/S,f,D},d+df \bigr) \end{equation} \begin{equation} \pi_{+}^i\bigl( L_{!}(f,D) \bigr) \simeq \hyperr^i\pi_{\ast} \bigl(\Omega^{\bullet+n}_{X/S}(\log D,f)(-D),d+df\bigr) \simeq \hyperr^i\pi_{\ast} \bigl( \Omega^{\bullet+n}_{X/S,f,D}(-D),d+df \bigr) \end{equation} \end{lem} \pf As in \S, we obtain the isomorphisms from Lemma . \hfill\qed \begin{cor} Suppose that $f$ is non-degenerate along $D$ over $S$. Then, $\pi^i_{+}\bigl( L_{\star}(f,D) \bigr)$ are flat bundles on $S$, i.e., locally free $\nbigo_S$-modules with an integrable connection. \end{cor} \pf The right hand sides of () and () are $\nbigo_S$-coherent. Then, the claim of this corollary follows from a well known result, i.e., if a $\nbigd_X$-module is coherent over $\nbigo_X$, then it is a flat bundle. \hfill\qed \subsection{Some functions satisfying the purity condition} We give some examples of meromorphic functions which are not necessarily non-degenerate but satisfy the purity condition. Let $X$ denote a complex manifold with a normal crossing hypersurface $D$. \subsubsection{Basic cases} We set $X^{(1)}:=X\times\proj^1$ and $D^{(1)}:=(D\times\proj^1) \cup (\{\infty\}\times X)$. Let $\tau$ be the standard coordinate on $\proj^1$. Take $f,g\in\nbigo_X(\ast D)$, and we consider a meromorphic function $F:=\tau f+g$ on $(X^{(1)},D^{(1)})$. We give some sufficient conditions for $F$ to be pure on an open subset in $X^{(1)}$. \begin{lem} Suppose the following. \begin{itemize} \item $g$ is non-degenerate along $D$. \item $|(f)_0|\cap |(f)_{\infty}|=\emptyset$, and $|(f)_0|\subset |(g)_{\infty}|$. In particular, $|(f)_0|\subset D$. \item $D=|(f)_{\infty}|\cup|(g)_{\infty}|$. \end{itemize} Then, $F$ is pure on $X^{(1)}$. \end{lem} \pf Let us consider the $\nbigd_{X^{(1)}}(\ast D^{(1)})$-module $\nbigv=\nbigo_{X^{(1)}}(\ast D^{(1)}) v$ with $\nabla v=v dF$. It is enough to prove that $\nbigv[!D^{(1)}] \lrarr \nbigv$ is an epimorphism around any point of $(P,\tau)\in X^{(1)}$. We take a holomorphic coordinate system $(x_1,\ldots,x_n)$ around $P$ such that $D=\bigcup_{i=1}^{\ell}\{x_i=0\}$. Let us consider the case $\tau\neq\infty$. For a large $N$, we have $v_1=\prod_{i=1}^{\ell} x_i^Nv \in\nbigv(\star D^{(1)})$ $(\star=\ast,!)$. It is enough to prove that $v_1$ generates $\nbigv$ as a $\nbigd_{X^{(1)}}$-module. For any $\vecm\in\seisuu^{\ell}$, we have \[ \del_i (x^{\vecm}v) =\bigl(x_i^{-1}m_i+\del_ig+\tau \del_if\bigr) x^{\vecm}v. \] We also have $\del_{\tau}v=fv$. For $i=1,\ldots,\ell$, we obtain \begin{equation} x_i\del_i(x^{\vecm}v) -(f^{-1}x_i\del_if)\cdot\tau\del_{\tau} (x^{\vecm}v) =x^{\vecm}(x_i\del_ig+m_i)v. \end{equation} For $i=\ell+1,\ldots,n$, we have \begin{equation} \del_i(x^{\vecm}v) -(f^{-1}\del_if)\tau\del_{\tau} (x^{\vecm}v) =x^{\vecm}\del_i(g)v. \end{equation} By using () and (), we obtain that $\nbigo_{X^{(1)}}(\ast (g)_{\infty})v_1$ is contained in $\nbigd_{X^{(1)}}v_1$. Then, by using $\del_{\tau}v=fv$, we obtain that $\nbigv$ is contained in $\nbigd_{X^{(1)}}v_1$ around $(P,\tau)$. Let us consider the case $\tau=\infty$. Let $\kappa:=\tau^{-1}$. It is enough to see that $v_1=\kappa^N\prod_{i=1}^{\ell}x_i^Nv$ generates $\nbigv$ around $(P,\infty)$. We have $\kappa\del_{\kappa}v =-\tau\del_{\tau}v =-\kappa^{-1}fv$. As in the case $\tau\neq\infty$, we obtain $\nbigo_{X^{(1)}}(\ast (g)_{\infty})v_1$ is contained in $\nbigd_{X^{(1)}}v_1$ by using () and (). Then, by using $\kappa\del_{\kappa}v =-\kappa^{-1}fv$, we obtain that $\nbigv$ is contained in $\nbigd_{X^{(1)}}v_1$. \hfill\qed \vspace{.1in} The following lemma is easy to see. \begin{lem} Suppose the following: \begin{itemize} \item $g=0$. \item $D=|(f)_{\infty}|$. \item $(f)_0$ is smooth and reduced, and $D\cup|(f)_0|$ is normal crossing. \end{itemize} Then, $F=\tau f$ is non-degenerate on $\{\tau\neq 0\}\times X$. In particular, $F$ is pure on $\{\tau\neq 0\}\times X$. \hfill\qed \end{lem} \subsubsection{A variant} Let $X$ be a complex manifold with a simple normal crossing hypersurface $D$. We set $X^{(2)}:=\proj^1_{\tau}\times \proj^1_t\times X$. Let $D^{(2)}$ denote the union of $\{\infty\}\times\proj^1_t\times X$ and $\proj^1_{\tau}\times\{0,\infty\}\times X$ and $\proj^1_{\tau}\times\proj^1_t\times D$. Let $h$ be a meromorphic function on $(X,D)$ such that (i) $D=|(h)_{\infty}|$, (ii) $h$ is non-degenerate along $D$, (iii) $(h)_0$ is reduced and non-singular on $X$. We have the meromorphic function $F=t^{-1}\tau+th$ on $(X^{(2)},D^{(2)})$. We set $D_1:=\{t=0\}\times X$ and $D_2:=(\{t=\infty\}\times X)\cup(\proj^1_t\times D)$ in $\proj^1_t\times X$. We take a projective birational morphism $\varphi:Y\lrarr \proj^1_t\times X$ such that (i) $g:=\varphi^{\ast}(th)$ is non-degenerate, (ii) $D_Y:=\varphi^{-1}(D_1\cup D_2)$ is normal crossing, (iii) $Y\setminus \varphi^{-1}(D_1)\simeq (\proj^1_t\times X)\setminus D_1$. We set $f:=\varphi^{\ast}(t^{-1})$. \begin{lem} $\Ftilde=\tau f+g$ is pure on $\Ytilde:=\proj^1_{\tau}\times Y$. \end{lem} \pf Because $D_2\setminus D_1\subset |(th)_{\infty}|$, we have $\varphi^{-1}(D_2\setminus D_1) \subset \bigl|(g)_{\infty}\bigr|$. We clearly have $\varphi^{-1}(D_1)= \bigl| (f)_{\infty} \bigr|$. Hence, we have $D_Y=|(f)_{\infty}|\cup|(g)_{\infty}|$. We also have $|(f)_0|\cap|(f)_{\infty}|=\emptyset$ and $|(f)_0|\subset|(g)_{\infty}|$. Then, we obtain the claim of Lemma from Lemma . \hfill\qed \begin{lem} $F$ is pure on $\proj^1_{\tau}\times\proj^1_t\times X$. \end{lem} \pf Let $\varphitilde:\Ytilde\lrarr X^{(2)}$ be the induced morphism. Let $\Dtilde_Y$ be the union of $\{\infty\}\times Y$ and $\proj^1_{\tau}\times D_Y$ in $\Ytilde$. By the previous lemma, $\Ftilde:=\tau f+g$ is pure on $\Ytilde$. We also have $\Dtilde_Y=|(\Ftilde)_{\infty}|$. Hence, the natural morphism $L_!(\Ftilde,\Dtilde_Y) \lrarr L_{\ast}(\Ftilde,\Dtilde_Y)$ is an isomorphism. Because $L_{\star}(F,D^{(2)})\simeq \varphitilde_{\dagger} L_{\star}(\Ftilde,\Dtilde_Y)$, we obtain that $L_{!}(F,D^{(2)}) \lrarr L_{\ast}(F,D^{(2)})$ is an isomorphism, i.e., $F$ is pure on $X^{(2)}$. \hfill\qed \subsection{Push-forward} \subsubsection{A statement} Let $X$ be a complex manifold with a simple normal crossing hypersurface $D$. We set $Y:=X\times\proj^1$ and $D_Y^{(0)}:= (X\times\{\infty\})\cup (D\times \proj^1)$ and $D_Y^{(1)}:=D_Y^{(0)}\cup(X\times\{0\})$. Let $f$ and $g$ be meromorphic functions on $(X,D)$. We assume the following. \begin{itemize} \item The divisor $(f)_0\cap (X\setminus D)$ is reduced and non-singular. \end{itemize} We set $Z_f:=|(f)_0|$. Let $[z_0:z_1]$ be a homogeneous coordinate system of $\proj^1$, and we set $t:=z_0/z_1$. We obtain the meromorphic function $tf$ on $Y$. The pull back of $g$ by the projection $Y\lrarr X$ is also denoted by $g$. We set $F:=tf+g$. We have the $\nbigd$-modules $L_{\ast}(F,D_Y^{(0)})$ and $L_{\ast}(F,D_Y^{(1)})$ on $Y$. We have the natural exact sequence: \begin{equation} 0\lrarr L_{\ast}(F,D_Y^{(0)}) \lrarr L_{\ast}(F,D_Y^{(1)}) \lrarr L_{\ast}(F,D_Y^{(1)}) \big/ L_{\ast}(F,D_Y^{(0)}) \lrarr 0 \end{equation} Let $\pi:Y\lrarr X$ be the projection. We shall prove the following proposition in \S. \begin{prop} We have $\pi^{i}_{+}(M)=0$ $(i\neq 0)$ for \[ M=L_{\ast}(F,D_Y^{(0)}),\,\, L_{\ast}(F,D_Y^{(1)}),\,\, L_{\ast}(F,D_Y^{(1)}) \big/L_{\ast}(F,D_Y^{(0)}). \] We have the following isomorphisms: \begin{equation} \pi_{+}^0L_{\ast}(F,D_Y^{(0)}) \simeq L_{\ast}(g,D)(\ast Z_f)/L_{\ast}(g,D) \simeq \Ker\Bigl( L_{\ast}(g)(!Z_f)(\ast D) \lrarr L_{\ast}(g,D) \Bigr) \end{equation} \begin{equation} \pi_{+}^0L_{\ast}(F,D_Y^{(1)}) \simeq L_{\ast}(g)(!Z_f)(\ast D) \end{equation} \begin{equation} \pi_{+}^0\bigl( L_{\ast}(F,D_Y^{(1)})\big/ L_{\ast}(F,D_Y^{(0)}) \bigr) \simeq L_{\ast}(g,D) \end{equation} The push-forward of {\rm()} is isomorphic to the following standard exact sequence \[ 0\lrarr \Ker\Bigl( L_{\ast}(g)(!Z_f)(\ast D) \lrarr L_{\ast}(g,D) \Bigr) \lrarr L_{\ast}(g)(!Z_f)(\ast D) \lrarr L_{\ast}(g,D) \lrarr 0 \] \end{prop} By the duality, we obtain the following as a corollary of Proposition . \begin{cor} We have $\pi_{+}^iM=0$ $(i\neq 0)$ for \[ M=L_{!}(F,D_Y^{(0)}),\,\, L_{!}(F,D_Y^{(1)}),\,\, \Ker\bigl( L_{!}(F,D_Y^{(1)}) \lrarr L_{!}(F,D_Y^{(0)}) \bigr). \] We have the following isomorphisms: \[ \pi_+^0\bigl( L_!(F,D_Y^{(0)}) \bigr) \simeq \Ker\Bigl( L_!(g,D)(!Z_f) \lrarr L_!(g,D) \Bigr) \simeq \Cok\Bigl( L_!(g,D) \lrarr L_!(g)(\ast Z_f)(!D) \Bigr) \] \[ \pi_+^0\bigl( L_!(F,D_Y^{(1)}) \bigr) \simeq L_!(g)(\ast Z_f)(!D) \] \[ \pi_+^0\bigl( \Ker\bigl( L_!(F,D_Y^{(1)}) \lrarr L_!(F,D_Y^{(0)}) \bigr) \bigr) \simeq L_!(g,D) \] The push-forward of \begin{equation} 0\lrarr \Ker\bigl( L_{!}(F,D_Y^{(1)}) \lrarr L_{!}(F,D_Y^{(0)}) \bigr) \lrarr L_{!}(F,D_Y^{(1)}) \lrarr L_{!}(F,D_Y^{(0)}) \lrarr 0 \end{equation} is isomorphic to the standard exact sequence: \begin{equation} 0\lrarr L_!(g,D) \lrarr L_!(g)(\ast Z_f)(!D) \lrarr \Cok\bigl( L_!(g,D) \lrarr L_!(g)(\ast Z_f)(!D) \bigr) \lrarr 0 \end{equation} \hfill\qed \end{cor} \begin{rem} Note that the restriction of the morphism $L_{!}(F,D_Y^{(0)}) \lrarr L_{\ast}(F,D_Y^{(0)})$ to $\proj^1\times (X\setminus D)$ is an isomorphism. Indeed, the restriction of $F$ to $\proj^1\times (X\setminus D)$ is non-degenerate along $\{\infty\}\times(X\setminus D)$. In particular, the restriction of the morphism $\pi^0_{+}L_!(F,D_Y^{(0)}) \lrarr \pi^0_{+}L_{\ast}(F,D_Y^{(0)})$ to $X\setminus D$ is an isomorphism. Indeed, $\pi^0_{+}L_{\star}(F,D_Y^{(0)})_{|X\setminus D}$ are isomorphic to the push-forward of the $\nbigd_{Z_f\setminus D}$-module $(\nbigo_{Z_f\setminus D},d+dg')$ via the inclusion $Z_f\setminus D\lrarr X\setminus D$, where $g'$ is the restriction of $g$ to $Z_f\setminus D$. \hfill\qed \end{rem} Let us consider the case that $f$ is moreover non-degenerate along $D$. In this case, $Z_f$ is smooth, and $Z_f\cup D$ is normal crossing. Let $\iota:Z_f\lrarr X$ denote the inclusion. We set $D_{Z_f}:=D\cap Z_f$. We set $g_0:=g_{|Z_f}$. We have the $\nbigd$-modules $L_{\star}(g_0,D_{Z_f})$ on $Z_f$. \begin{cor} If $f$ is moreover non-degenerate along $D$, we have the following isomorphisms: \begin{equation} \pi^0_{+}L_{\ast}(F,D_Y^{(0)}) \simeq L_{\ast}(g,D)(\ast Z_f)/L_{\ast}(g,D) \simeq \iota_+ L_{\ast}(g_0,D_{Z_f}) \end{equation} \begin{equation} \pi^0_{+}L_{!}(F,D_Y^{(0)}) \simeq \Ker\bigl( L_!(g,D)(!Z_f) \lrarr L_!(g,D) \bigr) \simeq \iota_+ L_!(g_0,D_{Z_f}) \end{equation} The image of $\pi^0_+L_!(F,D_Y^{(0)}) \lrarr \pi^0_+L_{\ast}(F,D_Y^{(0)})$ is naturally isomorphic to $\iota_+L(g_0)$. Note that it is $\iota_{+}\nbigo_{Z_f}$ if $g=0$. \end{cor} \pf The first isomorphism in () is given in Proposition . If $Z_f\cup D$ is normal crossing, we clearly have the second isomorphism in (). We obtain the isomorphisms in () by the duality. According to Remark , the restriction of the morphism $\pi^0_+L_!(F,D_Y^{(0)}) \lrarr \pi^0_+L_{\ast}(F,D_Y^{(0)})$ to $X\setminus D$ is an isomorphism. Hence, the image of $\pi^0_+L_!(F,D_Y^{(0)}) \lrarr \pi^0_+L_{\ast}(F,D_Y^{(0)})$ is identified with the image of a non-zero morphism $\iota_{+}L_{!}(g_0,D_{Z_f}) \lrarr \iota_+L_{\ast}(g_0,D_{Z_f})$, which is isomorphic to $\iota_{+}L(g_0)$. \hfill\qed \subsubsection{Extensions} We give a general preliminary which is a variant of Beilinson's construction . Let $h$ be any meromorphic function on a complex manifold $Z$. We set $D_h:=|(h)_0|\cup|(h)_{\infty}|$. Let $\nbigo_Z(\ast D_h)[s]:= \nbigo_Z(\ast D_h)\otimes_{\cnum}\cnum[s]$. For any pair of integers $(a,b)$ with $a< b$, we consider the meromorphic flat bundle \[ I_h^{a,b}:= s^a\nbigo_Z(\ast D_h)[s] \bigl/ s^b\nbigo_Z(\ast D_h)[s] \] with the connection $\nabla$ determined by $\nabla s^j=s^{j+1}dh/h$. For any holonomic $\nbigd_Z$-module $M$, we set $\Pi^{a,b}_hM:=I_h^{a,b}\otimes_{\nbigo_Z} M$. \vspace{.1in} We return to the situation in \S. We shall prove the following proposition in \S. \begin{prop} Suppose that $(f)_0\cap X\setminus D$ is smooth and reduced. Then, we have $\pi_{+}^j\bigl( \Pi^{a,b}_{t}L_{\star}(F)(\star t) \bigr)=0$ $(\star=\ast,!)$ for $j\neq 0$, and we have the following natural isomorphisms: \[ \Psi^{a,b}_{\ast}: \pi_{+}^0\bigl( \Pi^{a,b}_{t}L_{\ast}(F)(\ast t) \bigr) \simeq \Bigl( \Pi^{a,b}_{f^{-1}} L_{\ast}(g) \Bigr) (!(f)_0) (\ast D) \] \[ \Psi^{a,b}_{!}: \pi_{+}^0\bigl( \Pi^{a,b}_{t}L_{!}(F)(!t) \bigr) \simeq \Bigl( \Pi^{a,b}_{f^{-1}} L_!(g) \Bigr) (\ast (f)_0) (!D) \] The following diagram is commutative: \begin{equation} \begin{CD} \pi_{+}^0\bigl( \Pi^{a,b}_{t}L_{!}(F)(!t) \bigr) @>>> \pi_{+}^0\bigl( \Pi^{a,b}_{t}L_{\ast}(F)(\ast t) \bigr) \\ @V{\Psi^{a,b}_!}VV @V{\Psi^{a,b}_{\ast}}VV \\ \Bigl( \Pi^{a,b}_{f^{-1}} L_!(g) \Bigr) (\ast (f)_0) (!D) @>>> \Bigl( \Pi^{a,b}_{f^{-1}} L_{\ast}(g) \Bigr) (!(f)_0) (\ast D) \end{CD} \end{equation} Here, the upper horizontal arrow is the natural morphism, and the lower horizontal arrow is given by the multiplication of $-s$. \end{prop} \subsubsection{Proof of Proposition } Let us obtain the first isomorphism in (). We set $\nbigo_X(\ast D)[t]:= \nbigo_X(\ast D)\otimes_{\cnum}\cnum[t]$. Let $R\pi_{\ast}$ denote the ordinary push-forward of sheaves by $\pi$. We have $R^j\pi_{\ast}\bigl(L_{\ast}(F,D_Y^{(0)})\bigr)=0$ for $j>0$, and $\pi_{\ast}\bigl(L_{\ast}(F,D_Y^{(0)})\bigr) \simeq \nbigo_X(\ast D)[t]$ as an $\nbigo_X$-module. Hence, $\pi_{+}L_{\ast}(F)$ is represented by the following complex: \[ \nbigo_X(\ast D)[t] \stackrel{\del_t+f}\lrarr \nbigo_X(\ast D)[t] \] Here, the second term sits in the degree $0$. The action of vector fields $V$ on $X$ is given by $V(t^i)=t^{i}(tV(f)+V(g))$. It is easy to see that the morphism $\nbigo_X(\ast D)[t] \stackrel{\del_t+f}\lrarr \nbigo_X(\ast D)[t]$ is a monomorphism. Let us look at the cokernel. We consider the following morphism of $\nbigo_X(\ast D)$-modules: \[ \nbigo_X(\ast D)[t] \lrarr L_{\ast}(g,D)(\ast Z_f), \quad\quad t^i\longmapsto (-1)^ii!f^{-i-1} \] Here, we use the natural identification $L_{\ast}(g,D)(\ast Z_f)=\nbigo_X(\ast D)(\ast Z_f)$ as an $\nbigo_X$-module. It is a morphism of $\nbigd_X$-modules. Indeed, for a holomorphic vector field $V$ on $X$, we have \[ V(t^i)=t^{i+1}V(f)+t^iV(g), \quad V(f^{-i-1}(-1)^ii!) =(-1)^{i+1}(i+1)!f^{-i-2}V(f) +V(g)f^{-i-1}(-1)^i i!. \] For $i>0$, we have $\del_tt^i+ft^i=it^{i-1}+ft^i$ which is mapped to $i(-1)^{i-1}(i-1)!f^{-i} +f(-1)^ii!f^{-i-1} =0$. We also have $\del_tt^0+ft^0=f$ which is mapped to $1$. Hence, we obtain the morphism of the $\nbigd$-modules: \begin{equation} \Cok\Bigl( \nbigo_X(\ast D)[t] \stackrel{\del_t+f}{\lrarr} \nbigo_X(\ast D)[t] \Bigr) \lrarr L_{\ast}(g,D)(\ast Z_f)\big/ L_{\ast}(g,D) \end{equation} We consider filtrations $\nbigf_j(\nbigo_X(\ast D)[t]) =\bigoplus_{i\leq j}\nbigo_X(\ast D)t^iv$ and $\nbigf_jL_{\ast}(g,D)(\ast Z_f) =L_{\ast}(g,D)\bigl((j+1)Z_f\bigr)$. It is easy to see that the induced morphism on the graded modules is an isomorphism. Hence, () is an isomorphism. Thus, we obtain the first half of (). \vspace{.1in} Let us prove (). Let $\nbigo_X(\ast D)[t,t^{-1}]:= \nbigo_X(\ast D)\otimes_{\cnum} \cnum[t,t^{-1}]$. Let $\iota:X\times\{0\}\lrarr X\times\proj^1$ denote the inclusion. We identify $L_{\ast}(F,D^{(1)}_Y)\big/L_{\ast}(F,D^{(0)}_Y)$ with the $\nbigo_{Y}$-module \[ \nbigm_1:= \iota_{\ast} \Bigl( \nbigo_X(\ast D)[t,t^{-1}] \Big/ \nbigo_X(\ast D)[t] \Bigr)\,v \] and the connection $\nabla$ given by $\nabla v=v d(F)$. Then, $\pi_{+} \bigl(L_{\ast}(F,D^{(1)}_Y)\big/ L_{\ast}(F,D^{(0)}_Y) \bigr)$ is represented by the complex $\nbigm_1 \stackrel{\del_t+f}{\lrarr} \nbigm_1$. Here, the second term sits in the degree $0$. The kernel and the cokernel are denoted by $\Ker_1$ and $\Cok_1$, respectively. For the identification $\nbigm_1=\bigoplus_{j=1}^{\infty} \nbigo_X(\ast D)t^{-j}v$, we set $\nbigf_n:=\bigoplus_{j\leq n} \nbigo_X(\ast D)t^{-j}v$. We have $\del_t+f: \nbigf_n\lrarr\nbigf_{n+1}$, which induces an isomorphism $\Gr^{\nbigf}_n\simeq\Gr^{\nbigf}_{n+1}$ for $n\geq 1$. Hence, it is easy to check $\Ker_1=0$, and that the natural $\nbigo_X$-morphism $L_{\ast}(g,D)\lrarr \nbigo_X(\ast D)t^{-1}v=\nbigf_1$ given by $\varphi\longmapsto \varphi (t^{-1}v)$ induces an isomorphism $L_{\ast}(g,D)\lrarr \Cok_1$ which is compatible with the flat connection. \vspace{.1in} Let us observe (). We consider $\nbigm_2=\nbigo_X(\ast D)[t,t^{-1}]v$ with $\nabla v=vdF$. As before, $\pi_{+}\bigl( L_{\ast}(F,D_Y^{(1)}) \bigr)$ is represented by $\nbigm_2\stackrel{\del_t+f}\lrarr \nbigm_2$, where the second term sits in the degree $0$. The kernel and the cokernel are denoted by $\Ker_2$ and $\Cok_2$, respectively. We set $\nbigm_0:=\nbigo_X(\ast D)[t]v$. We have the natural exact sequence $0\lrarr \nbigm_0 \lrarr \nbigm_2 \lrarr \nbigm_1 \lrarr 0$. Because $\nbigm_i\stackrel{\del_t+f}{\lrarr}\nbigm_i$ $(i=0,1)$ are monomorphisms, we obtain that $\Ker_2=0$. We also have the following exact sequence of $\nbigd$-modules: \[ 0\lrarr L_{\ast}(g,D)(\ast Z_f)/L_{\ast}(g,D) \lrarr \Cok_2 \lrarr L_{\ast}(g,D) \lrarr 0 \] It implies that $\Cok_2(\ast Z_f)\simeq L_{\ast}(g,D)(\ast Z_f)$. Hence, we have a uniquely induced morphism $ L_{\ast}(g,D)(!Z_f) \lrarr \Cok_2$. Because $\Cok_2(\ast D)\simeq \Cok_2$, we have a uniquely induced morphism \[ \rho: L_{\ast}(g)(!Z_f)(\ast D) \simeq L_{\ast}(g,D)(!Z_f)(\ast D) \lrarr \Cok_2. \] Let us prove that the morphism is an isomorphism. Around any point of $Z_f\setminus D$, we have $\psi_f(\Cok_2)\simeq \psi_f\bigl(L_{\ast}(g)\bigr)$ and $\phi_f(\Cok_2)\simeq \phi_f\bigl( L_{\ast}(g)(\ast Z_f)/L_{\ast}(g) \bigr)$. By a direct computation, we can check that the natural morphism $\psi_f(\Cok_2)\lrarr \phi_f(\Cok_2)$ is non-zero. It implies that $\psi_f(\Cok_2)\lrarr \phi_f(\Cok_2)$ is an isomorphism. Hence, $\rho$ is an isomorphism on $X\setminus D$, and then it follows that $\rho$ is an isomorphism on $X$. We also obtain the second isomorphism in (), and that the push-forward of () is isomorphic to (). \hfill\qed \subsubsection{Proof of Proposition } Let us prove the claim in the case $\star=\ast$. We have the following exact sequence: \[ 0\lrarr \Pi^{a-1,b}_t L_{\ast}(F)(\ast t) \lrarr \Pi^{a,b}_t L_{\ast}(F)(\ast t) \lrarr L_{\ast}(F)(\ast t)\,s^a \lrarr 0 \] Then, we obtain $\pi_{+}^j \Pi^{a,b}_t L_{\ast}(F)(\ast t)=0$ $(j\neq 0)$ by applying Proposition and an easy induction. We also have the exact sequence: \[ 0\lrarr \pi_{+}^0 \Pi^{a-1,b}_t L_{\ast}(F)(\ast t) \lrarr \pi_{+}^0 \Pi^{a,b}_t L_{\ast}(F)(\ast t) \lrarr L_{\ast}(g)(!(f)_0)(\ast D) \lrarr 0 \] By an easy induction, we obtain that the following natural morphisms are isomorphisms: \[ \pi_{+}^0 \bigl( \Pi^{a,b}_t L_{\ast}(F)(\ast t) \bigr) \stackrel{\simeq}{\lrarr} \pi_{+}^0 \bigl( \Pi^{a,b}_t L_{\ast}(F)(\ast t) \bigr)(\ast D) \stackrel{\simeq}{\llarr} \pi_{+}^0 \bigl( \Pi^{a,b}_t L_{\ast}(F)(\ast t) \bigr)(!(f)_0)(\ast D). \] Hence, for the construction of $\Psi^{a,b}_{\ast}$, it is enough to obtain an isomorphism $\pi_{+}^0\bigl( \Pi^{a,b}_tL_{\ast}(F)(\ast t) \bigr) \simeq \Pi_{f^{-1}}^{a,b}L_{\ast}(g)$ on $X\setminus (D\cup Z_f)$. We have the following representative of $\pi_{+}\bigl( \Pi_{t}^{a,b}L_{\ast}(F)(\ast t) \bigr)$: \[ \bigoplus_{j=a}^{b-1} \nbigo_X(\ast D) [t]s^j \stackrel{\kappa}{\lrarr} \bigoplus_{j=a}^{b-1} \nbigo_X(\ast D)[t]t^{-1}s^j \] Here, the morphism $\kappa$ is given by $\del_{t}+f+st^{-1}$. The action of any vector field $V\in\Theta_X$ is given by $V(s^j)=\bigl(V(g)+tV(f)\bigr)s^j$. The kernel of $\kappa$ is clearly $0$. The natural inclusion of $\bigoplus_{j=a}^{b-1} L_{\ast}(g)t^{-1}s^j$ into $\bigoplus_{j=a}^{b-1} L_{\ast}(g)[t]t^{-1}s^j$ induces an isomorphism of $\bigoplus_{j=a}^{b-1} L_{\ast}(g)t^{-1}s^j$ and the cokernel of $\kappa$. We have $V(s^jt^{-1}) =V(g)s^jt^{-1}+V(f)s^j$. Because $\kappa(s^j) =fs^j+s^{j+1}t^{-1}$, we have $V(f)s^j\equiv fV(f^{-1})s^{j+1}t^{-1}$ on $X\setminus \bigl(Z_f\cup D\bigr)$. Thus, we obtain the isomorphism $\Psi^{a,b}_{\ast}: \pi_{+}^0\bigl( \Pi^{a,b}_tL_{\ast}(F)(\ast t) \bigr) \simeq \Pi_{f^{-1}}^{a,b}L_{\ast}(g)$ on $X\setminus (D\cup Z_f)$ by setting $\Psi^{a,b}_{\ast}(t^{-1}s^j):=s^j$. Let us consider the case $\star=!$. We can deduce it from the claim in the case $\star=\ast$ by using the duality. But, we give a more explicit construction which would be useful for our study later. As in the case of $\star=\ast$, we have $\pi_{+}^j \Pi_t^{a,b}L_!(F)(!t)=0$ $(j\neq 0)$ and the following natural isomorphisms: \[ \pi_{+}^0 \bigl( \Pi^{a,b}_t L_{!}(F)(!t) \bigr) \stackrel{\simeq}{\lrarr} \pi_{+}^0 \bigl( \Pi^{a,b}_t L_{!}(F)(!t) \bigr)(!D) \stackrel{\simeq}{\llarr} \pi_{+}^0 \bigl( \Pi^{a,b}_t L_{!}(F)(!t) \bigr)(\ast Z_f)(!D). \] Hence, it is enough to obtain an isomorphism $\pi_{+}^0\bigl( \Pi^{a,b}_tL_{!}(F)(!t) \bigr) \simeq \Pi_{f^{-1}}^{a,b}L_{!}(g)$ on $X\setminus (D\cup Z_f)$. We have the following representative of $\pi_{+}\bigl( \Pi_{t}^{a,b}L_{!}(F)(! t) \bigr)$ on $X\setminus\bigl(Z_f\cup D\bigr)$: \[ \bigoplus_{j=a}^{b-1} \nbigo_X[t]ts^j \stackrel{\kappa}{\lrarr} \bigoplus_{j=a}^{b-1} \nbigo_X[t]s^j \] The morphism $\kappa$ and the action of $V\in\Theta_X$ are given by the same formula. The natural inclusion of $\bigoplus_{j=a}^{b-1} L_{\ast}(g)s^j$ into $\bigoplus_{j=a}^{b-1} L_{\ast}(g)[t]s^j$ induces an isomorphism of $\bigoplus_{j=a}^{b-1} L_{\ast}(g)s^j$ and the cokernel of $\kappa$. On $X\setminus \bigl(Z_f\cup D\bigr)$, we have $V(s^j) =V(g)s^j+V(f)ts^j$ and $\kappa(s^{j}t)=fs^{j}t+s^{j+1}+s^j$. Hence, we have $V(fs^j)\equiv V(g)(fs^j)+fV(f^{-1})fs^{j+1}$. Thus, we obtain the isomorphism $\Psi^{a,b}_!$ by setting $\Psi^{a,b}_!(fs^j):=s^j$. Let us look at the restriction of the natural morphism $\pi_{+}^0 \Pi^{a,b}_tL_!(F)(!t) \lrarr \pi_{+}^0 \Pi^{a,b}_tL_{\ast}(F)(\ast t)$ to $X\setminus (D\cup Z_f)$. Because $\kappa(s^j)=fs^j+s^{j+1}t^{-1}$, we have $fs^j=-t^{-1}s^{j+1}$ in $\pi_{+}^0 \Pi^{a,b}_tL_{\ast}(F)(\ast t)$. Hence, under the above isomorphisms $\Psi^{a,b}_!$ and $\Psi^{a,b}_{\ast}$, it is identified with $\Pi^{a,b}_tL(g) \lrarr \Pi^{a,b}_tL(g)$ induced by the multiplication of $-s$, i.e., the diagram () is commutative. \hfill\qed \begin{rem} We also have the following representative of $\pi^0_+\bigl( \Pi^{a,b}_{t}L_{\ast}(F)(!t) \bigr)$: \[ \bigoplus_{j=a}^{b-1} \nbigo_X(\ast D)(!t)s^j \stackrel{\kappa}\lrarr \bigoplus_{j=a}^{b-1} \nbigo_X(\ast D)(!t)s^j \] For any local section $c$ of $\bigoplus_{j=a}^{b-1} \nbigo_X(\ast D)(!t)s^j$, let $[c]_!$ denote the induced section of $\pi^0_+\bigl( \Pi^{a,b}_{t}L_{\ast}(F)(!t) \bigr)$. Recall that we have a natural isomorphism $L(F)(!t)\simeq L(F)\otimes \nbigo_X(!t)$. \end{rem} \subsubsection{A consequence} We continue to use the notation in \S. For simplicity, we consider the case $g=0$. Motivated by the descriptions of some hypergeometric systems in and , we give a remark on another description of $\pi_{+}^0\bigl( L_{\ast}(tf,D_Y^{(0)}) \bigr)$. We set $Y_0:=X\times\proj^1_t\times\proj^1_s$. Let $p_i$ denote the projection of $Y_0$ onto the $i$-th component. We set $D_0:=p_1^{-1}(D) \cup p_2^{-1}(\{0,\infty\}) \cup p_3^{-1}(\{0,\infty\})$. We regard $t$ and $s$ as meromorphic functions on $(Y_0,D_0)$. Let us consider $L_{\ast}\bigl(t(f+s),D_0\bigr)$. Let $\pi_{ts}:X\times\proj^1_t\times\proj^1_s\lrarr X$, $\pi_t:X\times\proj^1_t\lrarr X$ and $\pi_s:X\times\proj^1_s\lrarr X$ denote the projections. Set $D_{X\times\proj^1_t}:= \bigl( X\times\{0,\infty\} \bigr) \cup \bigl( D\times\proj^1_t \bigr)$, and $D_{X\times\proj^1_s}:= \bigl( X\times\{0,\infty\} \bigr) \cup \bigl( D\times\proj^1_s \bigr)$. \begin{prop} We have the following natural isomorphisms: \begin{equation} \pi_{s+}\nbigo_{X\times\proj^1_s} \bigl(!(f+s)_0\bigr) \bigl(\ast D_{X\times\proj^1_s}\bigr) \simeq \pi_{ts+}L_{\ast}\bigl( t(f+s),D_0 \bigr) \simeq \pi_{t+}L_{\ast}(tf,D_{X\times\proj^1_t}) \simeq \nbigo_{X}(!(f)_0)(\ast D). \end{equation} \end{prop} \pf Let $p_{12}$ and $p_{13}$ denote the projection of $Y_0$ onto $X\times\proj^1_t$ and $X\times\proj^1_s$, respectively. By Proposition , we have \[ p^0_{13+}L_{\ast}\bigl(t(f+s),D_0\bigr) \simeq \nbigo_{X\times\proj^1_s} \bigl(!(f+s)_0\bigr) \bigl(\ast D_{X\times\proj^1_s}\bigr), \quad\quad p^i_{13+}L_{\ast}\bigl(t(f+s),D_0\bigr)=0\,\,\, (i\neq 0). \] We also have the following by Proposition : \[ p^0_{12+}L_{\ast}\bigl(t(f+s),D_0\bigr) \simeq L_{\ast}\bigl(tf,D_{X\times\proj^1_t}\bigr), \quad\quad p^i_{12+}L_{\ast}\bigl(t(f+s),D_0\bigr)=0\,\,\, (i\neq 0). \] Then, the claim of the proposition follows. \hfill\qed \subsubsection{Complement for a non-resonant case} We give a remark on a non-resonant case to compare the resonant case above. Take $\alpha\in\cnum\setminus\seisuu$. We consider the line bundle $L_{\alpha,Y}(tf):=\nbigo_{Y}(\ast D_Y^{(1)})\,e$, with the flat connection $\nabla$ determined by $\nabla e=e(d(tf)+\alpha dt/t)$, where $e$ is a global frame. Similarly, we consider the line bundle $L_{-\alpha,X}:= \nbigo_X\bigl(\ast (D\cup (f)_0)\bigr)\,v$ with the connection $\nabla$ determined by $\nabla v=v\,(-\alpha df/f)$. The following is essentially contained in Theorem {\rm 1.5} of {\rm}. \begin{prop} We have a natural isomorphism $\pi^0_{+}L_{\alpha,Y}(tf)\simeq L_{-\alpha,X}$. \end{prop} \pf We use the notation in {\rm\S}. Indeed, $\pi_{+}^0L_{\alpha,Y}(tf)$ is represented by the complex \[ \nbigo_X(\ast D)[t,t^{-1}]\,e \stackrel{a}{\lrarr} \nbigo_X(\ast D)[t,t^{-1}]\,e. \] Here, the second term sits in the degree $0$, and $a(g)=(t\del_t+\alpha+tf)g$. The action of a holomorphic vector field $V$ on $t^j e$ is given as $V(t^je)=t^{j+1}(Vf)e$. We define an $\nbigo_X(\ast D)$-homomorphism $\Phi:\nbigo_X(\ast D)[t,t^{-1}]\,e \lrarr \nbigo_X(\ast D)\,v$ by $\Phi(t^je)= \Gamma(-\alpha+1)\, \Gamma(-\alpha-j+1)^{-1} f^{-j}v$. Then, we can check that $\Phi$ is compatible with the connections, and $\Phi\circ a=0$. It induces a morphism of $\nbigd$-modules $\pi^0_{+}L_{\alpha,Y}(tf) \lrarr L_{-\alpha,X}$. We can check that the induced morphism is an isomorphism. \hfill\qed \subsection{Specialization} \subsubsection{Statement} Let $X$ be a complex manifold with a simple normal crossing hypersurface $D$. Let us consider meromorphic functions $f$ and $g$ on $(X,D)$. We set $X^{(1)}:=X\times\cnum_{\tau}$ and $D^{(1)}:=D\times\cnum_{\tau}$. We have the meromorphic function $\tau f+g$ on $(X^{(1)},D^{(1)})$ and the associated $\nbigd$-modules $\nbigm_{\star,f,g}:=L_{\star}(\tau f+g,D^{(1)})$ on $X^{(1)}$ for $\star=\ast,!$. Let $K_{\star,f,g}$ and $C_{\star,f,g}$ denote the kernel and the cokernel of $\nbigm_{\star,f,g}(!\tau)\lrarr \nbigm_{\star,f,g}(\ast \tau)$ for $\star=\ast,!$. Let $\iota_0:X\lrarr X^{(1)}$ be given by $\iota_0(Q)=(Q,0)$. We shall prove the following proposition in \S--\S. \begin{prop} If $|(f)_0|\cap |(f)_{\infty}|=\emptyset$, we have the following: \[ C_{\ast,f,g}\simeq \iota_{0+}L_{\ast}(g,D), \quad K_{\ast,f,g}\simeq \iota_{0+}L_{\ast}(g,D)(!(f)_{\infty}). \] \[ K_{!,f,g}\simeq \iota_{0+}L_!(g,D), \quad C_{!,f,g}\simeq \iota_{0+}L_!(g,D)(\ast (f)_{\infty}). \] \end{prop} \subsubsection{The case $g=0$ and $D=|(f)_{\infty}|$} First, we consider the case $g=0$ and $D=|(f)_{\infty}|$. In this case, $\tau f$ is non-degenerate along $D^{(1)}$. Hence, we have $\nbigm_{\ast,f,0}=\nbigm_{!,f,0}$ which we denote by $\nbigm_f$. We also have $K_{\ast,f,0}\simeq K_{!,f,0}$ and $C_{\ast,f,0}\simeq C_{!,f,0}$. We have a global section $v$ of $\nbigm_f$ such that $\nbigm_f= \nbigo_{X^{(1)}}(\ast D^{(1)})v$ with $\nabla v=v\,d(\tau f)$. We described the $V$-filtration $U_{\bullet}(\nbigm_f)$ of $\nbigm_f$ along $\tau=0$ in , which we recall here. We use the convention that $\tau\del_{\tau}+\alpha$ is locally nilpotent on $U_{\alpha}\big/U_{<\alpha}$. We describe the filtration locally around any point of $(Q,0)\in D\times\{0\}$. Suppose $Q\not\in |(f)_{\infty}|$. We have $\Gr^U_j(\nbigm_f(\ast \tau))=0$ unless $j\in\seisuu$, and $U_j\bigl( \nbigm_f(\ast\tau)\bigr) =\tau^{-j}\nbigm_f$ for $j\in\seisuu$. Suppose $Q\in |(f)_{\infty}|$. We take a holomorphic coordinate system $(z_1,\ldots,z_n)$ around $Q$ such that $D=\bigcup_{i=1}^{\ell}\{z_i=0\}$ and $f=z^{-\veck}$ for some $\veck\in\seisuu_{> 0}^{\ell}$. For $0\leq\alpha<1$, we set $\vecp=[\alpha\veck]:= \bigl( [\alpha k_1],\ldots,[\alpha k_{\ell}] \bigr)$, where $[a]:=\max\{n\in\seisuu\,|\,n\leq a\}$. Let $\vecdelta=(1,\ldots,1)\in\seisuu_{\geq 0}^{\ell}$. Let $\pi:X^{(1)}\lrarr X$ denote the projection. We may naturally regard $\pi^{\ast}\nbigd_X$ as a sheaf of subalgebras in $\nbigd_{X^{(1)}}$. Locally around $(Q,0)$, we have \[ U_{\alpha}\bigl(\nbigm_f(\ast \tau)\bigr) =\pi^{\ast}\nbigd_X\bigl( \nbigo_{X^{(1)}} x^{-\vecdelta-\vecp}v \bigr) =\pi^{\ast}\nbigd_X\Bigl( \sum_{j=0}^{\infty} \nbigo_{X^{(1)}}x^{-\vecdelta-\vecp}(\tau f)^j\,v \Bigr) \] in $\nbigo_{X^{(1)}}(\ast D^{(1)})v$. We have $U_1\bigl( \nbigm(\ast \tau) \bigr) =\tau^{-1}U_0\nbigm$. We obtain \[ U_{<0}\nbigm+\tau\del_{\tau}U_0\nbigm =\pi^{\ast}\nbigd_X \Bigl( \sum_{j=1}^{\infty} \nbigo_{X^{(1)}}x^{-\vecdelta}(\tau f)^j\,v \Bigr). \] Hence, $U_0\nbigm\big/ (U_{<0}\nbigm+\tau\del_{\tau}U_0\nbigm) \simeq \iota_{0\ast}\Bigl( \nbigd_X\bigl( \nbigo_X(D) \bigr) \Bigr) \simeq \iota_{0\ast}\nbigo_X(\ast D)$. It implies $C_{\star,f,0}\simeq \iota_{0+}\nbigo_X(\ast D)$ if $D=|(f)_{\infty}|$. By using the duality, we obtain $K_{\star,f,0}\simeq\iota_{0+}\nbigo_X(! D)$ if $D=|(f)_{\infty}|$. \subsubsection{The case $|(g)_0|\cap|(g)_{\infty}|=\emptyset$} Let us consider the case that $|(g)_0|\cap|(g)_{\infty}|=\emptyset$. We define $\nbigm_0:=L(\tau f)$. Set $\nbigm:=\nbigm_{\ast,f,g}$. We put $D_0:=|(f)_{\infty}|\cup|(g)_{\infty}|$. We have the hypersurface $D_1\subset D$ such that $D_0\cup D_1=D$ and $\codim_X (D_0\cap D_1)\geq 2$. We set $D_2:=|(g)_{\infty}|\cup D_1$. We have $\nbigm\simeq \nbigm_0\otimes L_{\ast}(g,D_2^{(1)})$. We naturally have $\nbigm(\ast\tau) \simeq \nbigm_0(\ast\tau)\otimes L_{\ast}(g,D_2^{(1)})$. We shall observe that $\nbigm(!\tau) \simeq \nbigm_0(!\tau)\otimes L_{\ast}(g,D_2^{(1)})$. We have the $V$-filtration $U_{\bullet}(\nbigm_0(\ast \tau))$ of $\nbigm_0(\ast\tau)$. Set $U_{\alpha}(\nbigm(\ast\tau)) :=U_{\alpha}(\nbigm_0(\ast\tau))\otimes L_{\ast}(g,D_2^{(1)})$ for any $\alpha\in\real$. \begin{lem} $U_{\bullet}(\nbigm(\ast \tau))$ is the $V$-filtration of $\nbigm(\ast\tau)$. \end{lem} \pf By the construction, $\tau\del_{\tau}+\alpha$ is locally nilpotent on $U_{\alpha}\nbigm(\ast\tau) \big/ U_{<\alpha}\nbigm(\ast\tau)$. Let us prove that $U_{\alpha}\nbigm$ is $V\nbigd_{X^{(1)}}$-coherent. If $Q\not\in |(f)_{\infty}|$, the claim is clear. Let us consider the case $Q \in |(f)_{\infty}|$. We take a holomorphic coordinate system $(x_1,\ldots,x_n)$ around $Q$ such that $D=\bigcup_{i=1}^{\ell}\{x_i=0\}$ and $g=x^{-\veca}$ and $f=f_1 x^{-\vecb}$, where $f_1$ is nowhere vanishing. Here $\veca,\vecb\in\seisuu^{\ell}_{\geq 0}$. Put $\vecp:=[\alpha\vecb]$. Let $\vecdelta=(1,\ldots,1)\in\seisuu^{\ell}$. We use the identifications $\nbigm_0= \nbigo_X^{(1)}\bigl(\ast (f)^{(1)}_{\infty}\bigr)\,v$ with $\nabla v=v\,d(\tau f)$, and $L_{\ast}(g,D^{(1)}_2)= \nbigo_{X^{(1)}}\bigl(\ast D^{(1)}_2\bigr)\,e$ with $\nabla e=e\,dg$. It is enough to prove that $U_{\alpha}(\nbigm(\ast\tau)) \otimes L_{\ast}(g,D^{(1)}_2)$ is generated by $x^{-\vecdelta-\vecp}v\otimes e$ over $V\nbigd_{X^{(1)}}$. Set $\supp(\veca):=\{i\,|\,a_i\neq 0\}$ and $\supp(\vecb):=\{i\,|\,b_i\neq 0\}$. For $i\in \supp(\veca)\cup\supp(\vecb)$, and for any $\vecq\in\seisuu_{\geq 0}^{\ell}$, we have \[ \del_ix_i(x^{-\vecdelta-\vecp-\vecq}v\otimes e) =x^{-\vecdelta-\vecp-\vecq} \Bigl( -p_i-q_i -a_ix^{-\veca} -b_i\tau f_1 x^{-\vecb} +\tau x_ix^{-\vecb}\del_if_1 \Bigr)\,v\otimes e. \] Because $\tau\del_{\tau}(v\otimes e) =f_1 \tau x^{-\vecb}(v\otimes e)$, we have \begin{equation} \del_ix_i(x^{-\vecdelta-\vecp-\vecq}v\otimes e) +(b_i-\tau f_1^{-1}x_i\del_if_1) \tau\del_{\tau}(x^{-\vecdelta-\vecp-\vecq}v\otimes e) =x^{-\vecdelta-\vecp-\vecq} (-p_i-q_i-a_ix^{-\veca})v\otimes e. \end{equation} By using (), we obtain that $x^{-\vecdelta-\vecp}v\otimes \nbigo_{X^{(1)}}(\ast D^{(1)}_2)e \subset V\nbigd_{X^{(1)}}(v\otimes e)$. Let $\vecdelta_1:=(1,\ldots,1)\in\seisuu^{\supp(\vecb)}$. For any $Q\in \pi^{\ast}\nbigd_X$ and for any $h\in \nbigo_{X^{(1)}}(\ast D^{(1)}_2)$, we have $\del_i\bigl( Q(x^{-\vecdelta_1}v) \otimes he \bigr) =\del_iQ(x^{-\vecdelta_1}v) \otimes he +Q(x^{-\vecdelta_1}v) \otimes \del_i(he)$. Hence, we can easily deduce that $U_{\alpha}(\nbigm) \subset V\nbigd_{X^{(1)}}\,x^{-\vecdelta-\vecp}(v\otimes e)$. Thus, we obtain the lemma. \hfill\qed \vspace{.1in} We set $U_{\alpha}\bigl( \nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2) \bigr):= U_{\alpha}\bigl(\nbigm_0(!\tau)\bigr) \otimes L_{\ast}(g,D^{(1)}_2)$ for any $\alpha\in\real$. \begin{lem} $U_{\bullet}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr)$ is the $V$-filtration of $\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)$. \end{lem} \pf For $\alpha<1$, we have $U_{\alpha}\bigl( \nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2) \bigr) =U_{\alpha}(\nbigm_0)\otimes L_{\ast}(g,D^{(1)}_2)$, which is coherent over $V\nbigd_{X^{(1)}}$. For $\alpha\geq 1$, we have $U_{\alpha}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}\bigl(g,D^{(1)}_2)\bigr) =\sum_{\beta+n\leq \alpha,\,\,\beta<1} \del_{\tau}^n U_{\beta}\bigl( \nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr)$. We obtain that $U_{\alpha}(\nbigm_0(!\tau) \otimes L_{\ast}\bigl(g,D^{(1)}_2)\bigr)$ are coherent over $V\nbigd_{X^{(1)}}$ for any $\alpha$. We have $\tau\del\tau+\alpha$ are nilpotent on $U_{\alpha}/U_{<\alpha}$. Hence, $U_{\bullet}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr)$ is the $V$-filtration of $\nbigm_0(!\tau)\otimes L_{\ast}\bigl(g,D^{(1)}_2\bigr)$. \hfill\qed \vspace{.1in} Because the induced morphism \begin{multline} \del_{\tau}: U_{0}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr) \big/ U_{<0}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr) \\ \lrarr U_{1}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr) \big/ U_{<1}\bigl(\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2)\bigr) \end{multline} is an isomorphism, we have $\nbigm_0(!\tau)\otimes L_{\ast}(g,D^{(1)}_2) \simeq \nbigm(!\tau)$. We obtain $C_{\ast,f,g}\simeq \iota_{0+}\Bigl( \nbigo_X(\ast (f)_{\infty}) \otimes L_{\ast}(g,D_2) \Bigr) \simeq \iota_{0+}L_{\ast}(g,D)$. We also obtain $K_{\ast,f,g}\simeq \iota_{0+}\Bigl( \nbigo_X(!(f)_{\infty})\otimes L_{\ast}(g,D_2) \Bigr)$. Under the assumption $|(g)_0|\cap|(g)_{\infty}|=\emptyset$, it is naturally isomorphic to $\iota_{0+}L_{\ast}(g,D_2)(!(f)_{\infty})$. \vspace{.1in} Thus, we obtain the claims for $K_{\ast,f,g}$ and $C_{\ast,f,g}$ in the case $|(g)_{\infty}|\cap|(g)_0|=\emptyset$. By using the duality, we also obtain the claims for $C_{!,f,g}$ and $K_{!,f,g}$ in this case. \subsubsection{The general case} Let us consider the general case. We take a projective birational morphism of complex manifolds $G:X'\lrarr X$ such that (i) $D':=G^{-1}(D)$ is a simple normal crossing hypersurface, (ii) $X'\setminus D'\simeq X\setminus D$, (iii) $|(g')_0|\cap|(g')_{\infty}|=\emptyset$, where $f':=G^{\ast}(f)$. We set $g':=G^{\ast}(g)$. We have $(f')_{0}=G^{\ast}((f)_0)$ and $(f')_{\infty}=G^{\ast}((f)_{\infty})$. In particular, we have $|(f')_0|\cap|(f')_{\infty}|=\emptyset$. We have $C_{\ast,f',g'}\simeq L_{\ast}(g',D')$ and $K_{\ast,f',g'}\simeq L_{\ast}(g',D')(!(f')_{\infty})$. The induced morphism $X^{\prime(1)}\lrarr X^{(1)}$ is also denoted by $G$. We have $G^0_{+}(\nbigm_{\ast,f',g'}) \simeq \nbigm_{\ast,f,g}$ and $G^i_{+}(\nbigm_{\ast,f',g'})=0$ for $i\neq 0$. We obtain $G^0_{+}(\nbigm_{\ast,f',g'}(\star\tau)) \simeq \nbigm_{\ast,f,g}(\star\tau)$ and $G^i_{+}(\nbigm_{\ast,f',g'}(\star\tau))=0$ for $i\neq 0$. We have $G^0_{+}\bigl( L_{\ast}(g',D') \bigr) \simeq L_{\ast}(g,D)$ and $G^i_+\bigl( L_{\ast}(g',D') \bigr)=0$ for $i\neq 0$. We also have $G^0_+\bigl( L_{\ast}(g',D')\bigl(!(f')_{\infty}\bigr) \bigr) \simeq L_{\ast}(g,D)\bigl((f)_{\infty}\bigr) $ and $G^i_+\bigl( L_{\ast}(g',D')\bigl(!(f')_{\infty}\bigr) =0$ for $i\neq 0$. Let $I$ be the image of $\nbigm'_{\ast,f',g'}(!\tau) \lrarr \nbigm'_{\ast,f',g'}(\ast\tau)$. Because $G^i_+\bigl( \nbigm_{\ast,f,g}(\ast\tau) \bigr) =G^i_{+}\bigl( C_{\ast,f,g} \bigr)=0$ for $i\neq 0$, we obtain that $G^i_+(I)=0$ for $i\neq 0,-1$. Because $G^i_+\bigl( \nbigm_{\ast,f,g}(!\tau) \bigr) =G^i_{+}\bigl( K_{\ast,f,g} \bigr)=0$ for $i\neq 0$, we obtain that $G^i_+(I)=0$ for $i\neq 0,1$. Hence, we obtain $G^i_+(I)=0$ for $i\neq 0$. It implies that $G^0_+(C_{\ast,f',g'}) \simeq C_{\ast,f,g}$ and $G^0_+(K_{\ast,f',g'}) \simeq K_{\ast,f,g}$. Thus, we obtain the claims for $C_{\ast,f,g}$ and $K_{\ast,f,g}$. By using the duality, we obtain the claims for $C_{!,f,g}$ and $K_{!,f,g}$. Thus, the proof of Proposition is finished. \hfill\qed \subsection{Nearby cycle functor and Push-forward} \subsubsection{Beilinson's functors and variants} Let $h$ be a meromorphic function on a complex manifold $Y$. Let $M$ be a holonomic $\nbigd_Y$-module. Suppose $|(h)_0|\cap|(h)_{\infty}|=\emptyset$. We recall the functors of Beilinson : \[ \psi^{(a)}_h(M):= \varprojlim_{b} \Cok\bigl( \Pi^{a,b}_hM(!(h)_0) \lrarr \Pi^{a,b}_hM(\ast(h)_0) \bigr) \] \[ \Xi^{(a)}_h(M):= \varprojlim_{b} \Cok\bigl( \Pi^{a-1,b}_hM(!(h)_0) \lrarr \Pi^{a,b}_hM(\ast(h)_0) \bigr). \] (\S for $\Pi^{a,b}_hM$.) On any relatively compact subset $K$ in $Y$, if $b$ is sufficiently large, the cokernel of $\Pi^{a,b}_hM(!(h)_0) \lrarr \Pi^{a,b}_hM(\ast(h)_0)$ is isomorphic to $\psi^{(a)}_h(M)$ and the kernel is isomorphic to $\psi^{(b)}_h(M)$. On $K$, if $N$ is sufficiently large, $\Xi^{(a)}_h(M)$ is isomorphic to the cokernel of $\Pi^{a+1,a+N}_hM(!(h)_0) \lrarr \Pi^{a,a+N}_hM(\ast(h)_0)$ and the kernel of $\Pi^{a-N,a+1}_hM(!(h)_0) \lrarr \Pi^{a-N,a}_hM(\ast(h)_0)$. (See the argument in \cite[\S4.1]{Mochizuki-MTM}, for example.) \vspace{.1in} In the case $|(h)_0|\cap|(h)_{\infty}|\neq\emptyset$ and $M=\nbigo_Y(\ast D)$ for a normal crossing hypersurface $D\subset Y$, we also consider a variant. For simplicity, we assume that (i) $(h)_{\infty}$ is smooth and reduced, (ii) $D=|(h)_0|$, (iii) $D\cup|(h)_{\infty}|$ is normal crossing. We set $Z:=|(h)_{\infty}|$. \begin{lem} We have a canonical morphism $\Pi^{a+1,b+1}_{h}\nbigo_Y \lrarr \Pi^{a,b}_{h}\nbigo_Y(!Z)$ such that the composite with $\Pi^{a,b}_{h}\nbigo_Y(!Z) \lrarr \Pi^{a,b}_{h}\nbigo_Y$ is the canonical one. \end{lem} \pf Let $\nbigk$ and $\nbigc$ denote the kernel and the cokernel of the morphism $\Pi^{a,b+1}_{h}\nbigo_Y(!Z) \lrarr \Pi^{a,b+1}_{h}\nbigo_Y$. We set $D_Z:=D\cap Z$. Let $\iota:Z\lrarr Y$ denote the inclusion. By a direct computation, we can check $\nbigc\simeq \iota_+\nbigo_Z(\ast D_Z)s^a$ and $\nbigk\simeq \iota_+\nbigo_Z(\ast D_Z)s^b$. We have the induced monomorphism $\Pi^{a,b+1}_{h}\nbigo_X(!Z)/\nbigk \lrarr \Pi^{a,b+1}_{h}\nbigo_X$. Let us consider the canonical morphism $\Pi_{h}^{a+1,b+1}\nbigo_Y \lrarr \Pi_{h}^{a,b+1}\nbigo_Y$. The composite with $\Pi_{h}^{a,b+1}\nbigo_Y\lrarr \nbigc$ is $0$. Hence, we have an induced morphism $\Pi_{h}^{a+1,b+1}\nbigo_Y \lrarr \Pi_{h}^{a,b+1}\nbigo_Y(!Z)/\nbigk$. We have the following natural commutative diagram: \[ \begin{CD} \Pi^{a,b+1}_h\nbigo_Y(!Z)/\nbigk @>>> \Pi^{a,b+1}_h\nbigo_Y \\ @VVV @VVV \\ \Pi^{a,b}_h\nbigo_Y(!Z) @>>> \Pi^{a,b}_h\nbigo_Y \end{CD} \] Hence, we obtain $\Pi_{h}^{a+1,b+1}\nbigo_Y \lrarr \Pi_{h}^{a,b}\nbigo_Y(!Z)$ such that the composite with $\Pi_{h}^{a,b}\nbigo_Y(!Z) \lrarr \Pi_{h}^{a,b}\nbigo_Y$ is the canonical one. \hfill\qed \vspace{.1in} Let us consider the following induced morphism \begin{equation} \rho_N: \Pi^{a+1,a+N+1}_h\nbigo_Y(\ast (h)_{\infty})(!(h)_0) \lrarr \Pi^{a,a+N}_h\nbigo_Y(!(h)_{\infty})(\ast(h)_0). \end{equation} We set $\Xi^{\prime(a)}_h\bigl( \nbigo_X[\ast D] \bigr):= \varprojlim_N \Cok(\rho_N)$. \begin{lem} For any relatively compact subset $K\subset X$, there exists $N_0$ such that $\Cok(\rho_{N_2})_{|K} \lrarr \Cok(\rho_{N_1})_{|K}$ is an isomorphism if $N_0\leq N_1\leq N_2$. \end{lem} \pf As mentioned above, the claim is well known outside of $Z$. Hence, the induced morphism \[ \Cok(\rho_{N_2})_{|K}(\ast (h)_{\infty}) \lrarr \Cok(\rho_{N_1})_{|K}(\ast (h)_{\infty}) \] is an isomorphism. Take $Q\in Z\cap K$, and take a holomorphic local coordinate neighbourhood $(Y_Q;x_1,\ldots,x_n)$ of $Y$ around $Q$ such that $h_{|Y_Q}=x_n^{-1}\prod_{i=1}^{\ell}x_i^{k_i}$, where $k_i\in\seisuu_{>0}$. It is enough to prove that $\phi_{x_n}^{(0)} \Cok(\rho_{N_2})_{|K} \lrarr \phi_{x_n}^{(0)} \Cok(\rho_{N_1})_{|K}$ is an isomorphism. Set $h_1:=\prod_{i=1}^{\ell}x_i^{k_i}$ on $Z_Q:=Y_Q\cap Z$. Let $\iota_1:Z_Q\lrarr Y_Q$ be the inclusion. We have the following commutative diagram: \[ \begin{CD} \phi_{x_n}^{(0)} \Pi_h^{a+1,a+N+1} \nbigo_Y(\ast (h)_{\infty})(!(h)_0) @>>> \phi_{x_n}^{(0)} \Pi_h^{a,a+N} \nbigo_Y(! (h)_{\infty})(\ast(h)_0) \\ @V{\simeq}VV @V{\simeq}VV \\ \psi_{x_n}^{(0)} \Pi_h^{a+1,a+N+1} \nbigo_Y(\ast (h)_{\infty})(!(h)_0) @. \psi_{x_n}^{(1)} \Pi_h^{a,a+N} \nbigo_Y(! (h)_{\infty})(\ast(h)_0) \\ @V{b_1}V{\simeq}V @V{b_2}V{\simeq}V \\ \iota_{1+}\Pi^{a+1,a+N+1}_{h_1}\nbigo_{Z_Q}(!h_1) @>{c_1}>> \iota_{1+}\Pi^{a,a+N}_{h_1}\nbigo_{Z_Q} \end{CD} \] Here, $b_i$ are given as in \cite[Proposition 4.3.1]{Mochizuki-MTM}. We shall prove that $c_1$ is induced by the multiplication of $s^{-1}$ Lemma below. Hence, $\phi_{x_n}^{(0)} \Cok(\rho_{N_2})$ is identified with $\iota_{1+}\Cok\bigl( \Pi^{a,a+N}_{h_1}\nbigo_{Z_Q}(!h_1) \lrarr \Pi^{a,a+N}_{h_1}\nbigo_{Z_Q} \bigr)$. Then, the claim is reduced to the above standard case. \hfill\qed \begin{lem} $c_1$ is induced by the multiplication of $s^{-1}$. \end{lem} \pf It is enough to consider the case $D=\emptyset$. We may assume that $x_n=h_n^{-1}$. Taking the limit of $\rho_N$ for $N\to\infty$, we obtain $\Pi^{a+1,\infty}_{x_n^{-1}}\nbigo_Y \lrarr \Pi^{a,\infty}_{x_n^{-1}}\nbigo_Y[!x_n]$ such that the composite with the canonical morphism $\kappa_1: \Pi^{a,\infty}_{x_n^{-1}}\nbigo_Y[!x_n] \lrarr \Pi^{a,\infty}_{x_n^{-1}}\nbigo_Y$ is equal to the canonical morphism $\kappa_2: \Pi^{a+1,\infty}_{x_n^{-1}}\nbigo_Y \lrarr \Pi^{a,\infty}_{x_n^{-1}}\nbigo_Y$. We have the following commutative diagram: \[ \begin{CD} \phi_{x_n}^{(0)} \Pi_{x_n^{-1}}^{a+1,\infty}\nbigo_Y @>>> \phi_{x_n}^{(0)} \Pi_{x_n^{-1}}^{a,\infty}\nbigo_Y \\ @V{\simeq}VV @V{\simeq}VV \\ \iota_{1+} \Pi^{a+1,\infty}_{1}\nbigo_Z @>{i_1}>> \iota_{1+} \Pi^{a,\infty}_{1}\nbigo_Z \end{CD} \] Here, $i_1$ is the inclusion. We have the following commutative diagram: \[ \begin{CD} \phi_{x_n}^{(0)} \Pi_{x_n^{-1}}^{a,\infty}\nbigo_Y[!x_n] @>>> \phi_{x_n}^{(0)} \Pi_{x_n^{-1}}^{a,\infty}\nbigo_Y \\ @V{\simeq}VV @V{\simeq}VV \\ \iota_{1+} \Pi^{a,\infty}_{1}\nbigo_Z @>{i_2}>> \iota_{1+} \Pi^{a,\infty}_{1}\nbigo_Z \end{CD} \] Here, $i_2$ is induced by $-x_n\del_{x_n}$, and it is equal to the multiplication of $s$. Because $i_2\circ c_1=i_1$, we obtain that $c_1$ is induced by the multiplication of $s^{-1}$. \hfill\qed \begin{rem} In the proof of Lemma {\rm}, we also observed that $\phi_{x_n}^{(0)} \Xi^{\prime(a)}_h(\nbigo_X[\ast D])_{|Y_Q} \simeq \iota_{1+}\psi_{h_1}^{(a)}\nbigo_{Z_Q}$ around any point $Q\in Z\cap D$. \hfill\qed \end{rem} \subsubsection{The push-forward of the nearby cycle sheaf} Let $X$ be a complex manifold with a normal crossing hypersurface $D$. Let $f$ and $g$ be meromorphic functions on $(X,D)$. We set $Y:=\proj^1_{\tau}\times X$ and $D^{(1)}:=(\proj^1_{\tau}\times D)\cup(\{\infty\}\times X)$. Let $\pi:Y\lrarr X$ be the projection. We consider $F:=\tau f+g$ on $(Y,D^{(1)})$. We suppose the following. \begin{itemize} \item $F$ is pure on $\{\tau\neq 0\}\times X$. \item $g$ is pure on $X\setminus |(f)_{\infty}|$. \item $D=|(f)_{\infty}|\cup|(g)_{\infty}|$, $|(f)_0|\subset D$ and $|(f)_0|\cap |(f)_{\infty}|=\emptyset$. \end{itemize} Note that we have $\pi_+^0\psi_{\tau}^{(a)}L_{!}(F) =\pi_+^0\psi_{\tau}^{(a)}L_{\ast}(F)$ and $\Xi^{(a)}_{f^{-1}}L_{!}(g) \simeq \Xi^{(a)}_{f^{-1}}L_{\ast}(g)$ by the assumptions. \begin{prop} We have natural isomorphisms $\Lambda^{(a)}:\pi_{+}^0\psi_{\tau}^{(a)}L_{\ast}(F) \simeq \Xi^{(a)}_{f^{-1}}L_{\ast}(g)$. \end{prop} \pf By the assumption of the purity of $F$ on $\{\tau\neq 0\}\times X$, we have $\Pi^{a,N}_{\tau}L_{\ast}(F)(\star \tau) =\Pi^{a,N}_{\tau}L_!(F)(\star \tau)$. Let $\nbigi$ be the image of $\Pi^{a,b}_{\tau}L_{!}(F)(!\tau) \lrarr \Pi^{a,b}_{\tau}L_{\ast }(F)(\ast \tau)$. The support of the kernel $K^{a,b}$ and the cokernel $C^{a,b}$ are contained in $\tau=0$. Hence, we have $\pi_{+}^iK^{a,b}=0$ and $\pi_{+}^iC^{a,b}=0$ for $i\neq 0$. By Proposition , we have $\pi_{+}^i \Pi^{a,N}_{\tau}L_{\star_1}(F)(\star_2 \tau)=0$ for $i\neq 0$ and for $\star_1,\star_2\in\{\ast,!\}$. We can easily deduce that $\pi_{+}^i\nbigi=0$ unless $i=0$. Hence, we have the following exact sequence \[ 0\lrarr \pi_{+}^0 K^{a,b} \lrarr \pi_{+}^0 \Pi^{a,b}_{\tau} L_{!}(F)(!\tau) \lrarr \pi_{+}^0 \Pi^{a,b}_{\tau} L_{\ast}(F)(\ast\tau) \lrarr \pi_{+}^0 C^{a,b} \lrarr 0 \] If $|a-b|$ is sufficiently large, we have $C^{a,b}\simeq \psi^{(a)}_{\tau}L_{\ast}(F)$. Hence, we have \[ \pi_{+}^0 \psi^{(a)}_{\tau}L_{\ast}(F) \simeq \Cok\Bigl( \pi_{+}^0 \Pi^{a,b}_{\tau} L_{!}(F)(!\tau) \lrarr \pi_{+}^0 \Pi^{a,b}_{\tau} L_{\ast}(F)(\ast\tau) \Bigr) \] By Proposition , we have the following natural isomorphisms: \[ \Psi^{a,N}_{\star}: \pi_{+}^0 \Pi^{a,N}_{\tau}L_{\star}(F)(\star \tau) \simeq \Pi^{a,N}_{f^{-1}}L_{\star}(g)(\star D) \] By $|(g)_{\infty}|\cup|(f)_{\infty}|=D$ and the purity of $g$ on $X\setminus|(f)_{\infty}|$, we have \[ \Pi^{a,N}_{f^{-1}}L_{\ast}(g)(\ast D) \simeq \Pi^{a,N}_{f^{-1}}L_{\ast}(g)(\ast (f)_{\infty}), \quad\quad \Pi^{a,N}_{f^{-1}}L_{!}(g)(!D) \simeq \Pi^{a,N}_{f^{-1}}L_{\ast}(g)(!(f)_{\infty}). \] By the commutativity of (), the image of the morphism $\pi_{+}^0 \Pi^{a,N}_{\tau}L_{!}(F)(!\tau) \lrarr \pi_{+}^0 \Pi^{a,N}_{\tau}L_{\ast}(F)(\ast\tau)$ is identified with the image of \[ \Pi^{a+1,N+1}_{f^{-1}}L_{\ast}(g)(!(f)_{\infty}) \lrarr \Pi^{a,N}_{f^{-1}}L_{\ast}(g)(\ast (f)_{\infty}). \] Hence, we obtain the desired isomorphism. \hfill\qed \vspace{.1in} Let $\iota_0:X\lrarr \proj^1\times X$ be the inclusion given by $\iota_0(P)=(0,P)$. \begin{cor} Under the assumption, we have a natural isomorphism $\psi_{\tau}^{(a)}L_{\ast}(F) \simeq \iota_{0+} \Xi^{(a)}_{f^{-1}}L_{\ast}(g)$. \hfill\qed \end{cor} We have the canonical nilpotent map $N$ on $\psi_{\tau}^{(a)}L_{\ast}(F)$. The following proposition is clear by the above description. \begin{prop} Under the same assumption, we have $\Ker N\simeq \iota_{0\dagger}L_!(g,D)$ and $\Cok N\simeq \iota_{0\dagger}L_{\ast}(g,D)$. The canonical morphism $\Ker N\lrarr \Cok N$ is identified with the canonical morphism $L_!(g,D)\lrarr L_{\ast}(g,D)$. \hfill\qed \end{prop} \begin{rem} Note that $\Ker N$ and $\Cok N$ are isomorphic to the kernel and the cokernel of $L_{\star}(F)(!\tau) \lrarr L_{\star}(F)(\ast \tau)$. The isomorphisms in Proposition {\rm } and Proposition {\rm} are consistent. \hfill\qed \end{rem} \subsubsection{A variant} Let us give a similar statement in a slightly different situation. We continue to use the notation in \S. Suppose that (i) $g=0$, (ii) $D=|(f)_{\infty}|$, (iii) $(f)_0$ is non-singular and reduced, and $|(f)_0|\cup D$ is normal crossing. Note that $F=\tau f$ is non-degenerate on $\{\tau\neq 0\}\times X$. We set $Z:=|(f)_0|$, and let $\iota_Z:Z\lrarr X$ denote the inclusion. \begin{prop} We have a natural isomorphism $\pi_{+}^0\psi^{(a)}_{\tau}L_{\ast}(F) \simeq \Xi^{\prime(a)}_{f^{-1}}\nbigo_X$. \end{prop} \pf As in the proof of Proposition , we have \begin{equation} \pi_{+}^0\psi^{(a)}_{\tau}L_{\ast}(F) \simeq \Cok\Bigl( \pi_{+}^0\Pi^{a,a+N}_{\tau}L_{!}(F)(!\tau) \stackrel{\kappa_N}{\lrarr} \pi_{+}^0\Pi^{a,a+N}_{\tau}L_{\ast}(F)(\ast \tau) \Bigr). \end{equation} According to Proposition , we have the following isomorphisms: \begin{equation} \pi_{+}^0\Pi^{a,a+N}_{\tau}L_{!}(F)(!\tau) \simeq \Pi^{a,a+N}_{f^{-1}}\nbigo_X(\ast Z)(!D) \end{equation} \begin{equation} \pi_{+}^0\Pi^{a,a+N}_{\tau}L_{\ast}(F)(\ast\tau) \simeq \Pi^{a,a+N}_{f^{-1}}\nbigo_X(!Z)(\ast D) \end{equation} Then, we obtain the claim of Proposition from the following lemma. \begin{lem} Under the isomorphisms {\rm()} and {\rm()}, the image of $\kappa_N$ is equal to the image $\rho_N$ in {\rm()} if $N$ is sufficiently large. \end{lem} \pf We identify $\Pi^{a,a+N}_{f^{-1}}\nbigo_X(!Z)(\ast D)$ and $\Pi^{a+1,a+N+1}_{f^{-1}}\nbigo_X(!Z)(\ast D)$ by the multiplication of $-s$. We can compare the restriction of $\rho_N$ and $\kappa_N$ to $X\setminus \bigl(Z\cup D\bigr)$ by the commutativity of (). Take any $N_1\geq N$. We have the morphisms $\kappa_{N_1}$ and $\rho_{N_1}$. The morphisms $\kappa_{N}$ and $\rho_{N}$ are induced by $\kappa_{N_1}$ and $\rho_{N_1}$. Because $(\kappa_{N_1}-\rho_{N_1})_{|X\setminus (D\cup Z)}=0$, the difference $\kappa_{N_1}-\rho_{N_1}$ factors through $\iota_{Z+}\nbigo_Z(\ast D_Z)s^{a+N_1}$, where $D_Z:=Z\cap D$. Hence, we obtain that $\kappa_{N}-\rho_{N}=0$. Thus, we finish the proof of Lemma and Proposition . \hfill\qed \begin{cor} If $Z\cap D=\emptyset$, we have $\pi_{+}^0\psi^{(0)}_{\tau}L_{\ast}(\tau f) \simeq \Xi^{(0)}_{f^{-1}}\nbigo_X$. \hfill\qed \end{cor} Take a holomorphic local coordinate $(x_1,\ldots,x_n)$ such that $f=x_n^{-1}\prod_{i=1}^{\ell} x_i^{k_i}$. Set $x^{\veck}:=\prod_{i=1}^{\ell}x_i^{k_i}=f\cdot x_n$. \begin{lem} We have the following commutative diagram: \[ \begin{CD} \phi^{(0)}_{x_n} \pi_+^0 \Pi^{a,a+N}_{\tau}L_!(F)(!\tau) @>{\phi^{(0)}_{x_n}(\kappa_N)}>> \phi^{(0)}_{x_n} \pi_+^0 \Pi^{a,a+N}_{\tau}L_{\ast}(F)(\ast\tau) \\ @VVV @VVV \\ \iota_{Z+} \Pi^{a,a+N}_{x^{\veck}}\nbigo_Z[!D] @>{c}>> \iota_{Z+} \Pi^{a,a+N}_{x^{\veck}}\nbigo_Z \end{CD} \] Here, $c$ is induced by the multiplication of $-1$. \end{lem} \pf As observed in Lemma , $\kappa_N$ is identified with $\rho$ under the isomorphism $\Pi^{a,a+N}_{f^{-1}}\nbigo_X(!Z)(\ast D)$ and $\Pi^{a+1,a+N+1}_{f^{-1}}\nbigo_X(!Z)(\ast D)$ given by the multiplication of $-s$. Then, the claim of the lemma follows from Lemma . \hfill\qed \subsection{Mixed twistor $\nbigd$-modules} We recall some operations for $\nbigr$-modules and twistor $\nbigd$-modules. See , and for more details. \subsubsection{$\nbigr$-modules} Let $X$ be a complex manifold. We set $\nbigx:=\cnum_{\lambda}\times X$. Let $p_{\lambda}:\nbigx\lrarr X$ denote the projection. Let $\nbigr_X\subset\nbigd_{\nbigx}$ be the sheaf of subalgebras generated by $\lambda p_{\lambda}^{\ast}\Theta_X$ over $\nbigo_{\nbigx}$. We set $d_X:=\dim X$. The pull back $p_{\lambda}^{\ast}\Omega_X$ is also denoted by $\Omega_X$. For any left $\nbigr_X$-module $\nbigm$, $\lambda^{-d_X}\Omega_X\otimes_{\nbigo_{\nbigx}} \nbigm$ is naturally a right $\nbigr_X$-module, by which the category of left $\nbigr_X$-modules and the category of right $\nbigr_X$-modules are equivalent. In this paper, $\nbigr_X$-modules mean left $\nbigr_X$-modules. Let $D^b_c(\nbigr_X)$ denote the derived category of cohomologically bounded coherent complexes of $\nbigr_X$-modules. An $\nbigr_X$-module is equivalent to an $\nbigo_{\nbigx}$-module $\nbigm$ with a meromorphic relative flat connection \[ \DD^f: \nbigm\lrarr \Omega_{\nbigx/\cnum_{\lambda}}^1(\nbigx^0)\otimes\nbigm \] where $\nbigx^0:=\{0\}\times X$. The operator $\DD:=\lambda\DD^f$ is also often used, and called a family of flat $\lambda$-connections. The easiest example of $\nbigr_X$-module is the line bundle $\nbigo_{\nbigx}$ with the meromorphic relative flat connection $\DD^f$ determined by $\DD^f(1)=0$. It is just denoted by $\nbigo_{\nbigx}$. Let $\nbigm_i$ $(i=1,2)$ be $\nbigr_X$-modules. Then, $\nbigm_1\otimes_{\nbigo_{\nbigx}}\nbigm_2$ and $\nbigm_1\oplus\nbigm_2$ are naturally $\nbigr_X$-modules. The tensor product $\nbigm_1\otimes_{\nbigo_{\nbigx}}\nbigm_2$ is denoted just by $\nbigm_1\otimes\nbigm_2$, if there is no risk of confusion. We define the duality functor $\DDD_X:D^b_c(\nbigr_X)\lrarr D^b_c(\nbigr_X)$ by \[ \DDD_X(\nbigm):= \nrhom_{\nbigr_X}\bigl(\nbigm, \nbigr_X\otimes\Omega_X^{-1}\bigr)[\dim X]. \] (We will review more details in \S.) Note that $\DDD_X\nbigo_{\nbigx}$ is naturally isomorphic to $\lambda^{d_X}\nbigo_{\nbigx} =\nbigo_{\nbigx}(-d_X\nbigx^0)$. If $\nbigm$ is an $\nbigr_X$-module underlying a mixed twistor $\nbigd$-module, then $\nbigh^j\bigl( \DDD_X\nbigm\bigr)=0$ unless $j=0$. In that case, we identify $\DDD_X\nbigm$ and $\nbigh^0\DDD_X\nbigm$. Let $f:X\lrarr Y$ be any morphism of complex manifolds. We set \[ \nbigr_{Y\larr X}:= \lambda^{-d_X}\Omega_X\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigr_Y\otimes \lambda^{d_Y}\Omega_Y^{-1} \bigr). \] It is naturally a left $f^{-1}\nbigr_Y$-module and a right $\nbigr_X$-module. For any object $\nbigm$ in $D^b_c(\nbigr_X)$, we set \[ f_{\dagger}(\nbigm):= Rf_{!} \bigl( \nbigr_{Y\larr X} \otimes^L_{\nbigr_X} \nbigm\bigr) \] in the derived category of $\nbigr_Y$-modules. (We shall review more details on the push-forward in \S.) The $j$-th cohomology sheaves of $f_{\dagger}(\nbigm)$ are denoted by $f^j_{\dagger}(\nbigm)$. If $\nbigm$ is good relative to $f$, i.e., for any compact subset $K\subset\nbigy$, then there exists good filtration of $\nbigm_{|f^{-1}(K)}$. Moreover, suppose that the support of $\nbigm$ is proper over $f$. Then, $f_{\dagger}\nbigm$ is an object in $D^b_c(\nbigr_Y)$, and we have a natural isomorphism $f_{\dagger}\DDD_X\nbigm \simeq \DDD_Y f_{\dagger}\nbigm$. Let $j:\cnum_{\lambda}\lrarr\cnum_{\lambda}$ be given by $j(\lambda)=-\lambda$. The induced morphism $j\times\id_X:\nbigx\lrarr\nbigx$ is also denoted by $j$. For any $\nbigr_X$-module $M$, we naturally regard $j^{\ast}M$ as an $\nbigr_X$-module by the natural isomorphism $j^{\ast}\nbigr_X\simeq\nbigr_X$. \vspace{.1in} Let $H$ be a hypersurface in $X$. We set $\nbigh:=\cnum_{\lambda}\times H$. For any $\nbigr_X$-module $\nbigm$, we set $\nbigm(\ast H):= \nbigm\otimes\nbigo_{\nbigx}(\ast\nbigh)$. In particular, $\nbigr_X(\ast H):= \nbigr_X\otimes\nbigo_{\nbigx}(\ast\nbigh)$. We may naturally consider $\nbigr_X(\ast H)$-modules. For any $\nbigm\in D^b_c(\nbigr_X(\ast H))$, we set \[ \DDD_{X(\ast H)}(\nbigm) :=\nrhom_{\nbigr_X(\ast H)}\bigl( \nbigm,\nbigr_X(\ast H)\otimes\Omega_X^{-1} \bigr)[\dim X]. \] Let $f:X\lrarr Y$ be a proper morphism of complex manifolds. Let $H_Y$ be a hypersurface in $Y$, and we set $H_X:=f^{-1}(H_Y)$. For any $\nbigm\in D^b_c(\nbigr_X(\ast H_X))$, we have \[ f_{\dagger}(\nbigm):= Rf_{!}\bigl( \nbigr_{Y\rarr X} \otimes_{\nbigr_X}^L\nbigm \bigr) \simeq Rf_{!}\bigl( \nbigr_{Y\rarr X}(\ast H_X) \otimes_{\nbigr_X(\ast H_X)}^L\nbigm \bigr) \] in the derived category of $\nbigr_Y(\ast H_Y)$-modules. If $f$ induces an isomorphism $X\setminus H_X\simeq Y\setminus H_Y$, we have $f^j_{\dagger}\nbigm=0$ unless $j=0$. We shall identify $f_{\dagger}\nbigm$ and $f^0_{\dagger}\nbigm$ in that case. \subsubsection{$\nbigr$-triples} We set $\vecS:=\bigl\{\lambda\in\cnum\,\big|\,|\lambda|=1\bigr\}$. Let $\distribution_{\vecS\times X/\vecS}$ denote the sheaf of distributions on $\vecS\times X$ which are continuous with respect to $\vecS$. (See .) For local sections $P\in\nbigr_{X|\vecS\times X}$ and $Q\in\distribution_{\vecS\times X/X}$, the local section $P\bullet Q:=PQ$ of $\distribution_{\vecS\times X/\vecS}$ is naturally defined. Let $\sigma:\vecS\times X\lrarr \vecS\times X$ be given by $\sigma(\lambda,x)=(-\lambda,x)$. For local sections $\sigma^{\ast}P \in \sigma^{\ast}\nbigr_{X|\vecS\times X}$ and $Q\in\distribution_{\vecS\times X/X}$, the local section $\sigma^{\ast}(P)\bullet Q:= \sigma^{\ast}(\overline{P})Q$ is defined. Thus, the sheaf $\distribution_{\vecS\times X/X}$ is naturally a $(\nbigr_{X|\vecS\times X}, \sigma^{\ast}\nbigr_{X|\vecS\times X})$-module. Let $\nbigm_i$ $(i=1,2)$ be $\nbigr_X$-modules. A sesqui-linear pairing of $\nbigm_1$ and $\nbigm_2$ is a $(\nbigr_{X|\vecS\times X}, \sigma^{\ast}\nbigr_{X|\vecS\times X})$-homomorphism $C: \nbigm_{1|\vecS\times X} \otimes \sigma^{\ast} \nbigm_{2|\vecS\times X} \lrarr \distribution_{\vecS\times X/\vecS}$. Such a tuple $\nbigt=(\nbigm_1,\nbigm_2,C)$ is called an $\nbigr_X$-triple. Let $\nbigt_i=(\nbigm_{i1},\nbigm_{i2},C_i)$ be $\nbigr_X$-triples. A morphism of $\nbigr_X$-triples $\varphi:\nbigt_1\lrarr \nbigt_2$ is a pair of $\nbigr_X$-homomorphisms $\varphi_1:\nbigm_{21}\lrarr\nbigm_{11}$ and $\varphi_2:\nbigm_{12}\lrarr\nbigm_{22}$ such that $C(\varphi_1(m_2),\sigma^{\ast}m_1) =C(m_2,\sigma^{\ast}\varphi_2(m_1))$. For any morphism, we set $\Ker(\varphi):=\bigl( \Cok(\varphi_1),\Ker(\varphi_2),C_{\Ker\varphi} \bigr)$, where $C_{\Ker\varphi}$ denotes the naturally induced sesqui-linear pairing. Similarly, we set \[ \Cok(\varphi):= \bigl( \Ker(\varphi_1),\Cok(\varphi_2), C_{\Cok(\varphi)} \bigr), \quad \Image(\varphi):= \bigl( \Image(\varphi_1), \Image(\varphi_2), C_{\Image(\varphi)} \bigr). \] For an $\nbigr_X$-triple $\nbigt=(\nbigm_1,\nbigm_2,C)$, we set $\nbigt^{\ast}:= (\nbigm_2,\nbigm_1,C^{\ast})$, where $C^{\ast}(m_2,\sigma^{\ast}(m_1)):= \overline{\sigma^{\ast}C(m_1,\sigma^{\ast}m_2)}$. For a morphism $\varphi=(\varphi_1,\varphi_2): \nbigt_1\lrarr\nbigt_2$ of $\nbigr_X$-triples, we set $\varphi^{\ast}=(\varphi_2,\varphi_1)$. We set $\nbigu_X(a,b):= \bigl( \lambda^{a}\nbigo_{\nbigx}, \lambda^{b}\nbigo_{\nbigx}, C_0 \bigr)$, where $C_0(f,\sigma^{\ast}g):=f\overline{\sigma^{\ast}g}$. For any $\nbigr$-triple $\nbigt:=(\nbigm_1,\nbigm_2,C)$, we define $\nbigt\otimes\nbigu_X(a,b):= \bigl( \lambda^a\nbigm_1,\lambda^b\nbigm_2,C \bigr)$, where $C$ is induced by the natural identification $(\lambda^{c}\nbigm_{i})_{|\vecS\times X} =(\nbigm_i)_{|\vecS\times X}$. Particularly, $\nbigu_X(-n,n)$ is denoted by $\newTate(n)$, called the $n$-th Tate object, and $\nbigt\otimes\newTate(n)$ is called the $n$-th Tate twist of $\nbigt$. We use the identification $\newTate(n)\simeq\newTate(-n)^{\ast}$ given by the morphism $((-1)^n,(-1)^n)$. A morphism $\nbigs:\nbigt\lrarr \nbigt^{\ast}\otimes\newTate(-n)$ such that $\nbigs^{\ast}=(-1)^n\nbigs$ is called a sesqui-linear duality of weight $n$. For example, we have $\nbigu_X(p,q) \lrarr \nbigu_X(p,q)^{\ast} \otimes \newTate(-(p-q))$ given by $\bigl((-1)^p,(-1)^p\bigr)$. Let $f:X\lrarr Y$ be any morphism of complex manifolds. Let $\nbigt$ be an $\nbigr_X$-triple. We have a naturally defined push-forward $f_{\dagger}$ in the derived category of $\nbigr_Y$-triples. See . It induces a functor from the derived category of $\nbigr_X$-triples to the derived category of $\nbigr_Y$-triples. The $j$-th cohomology $\nbigr_Y$-triple of $f_{\dagger}\nbigt$ is denoted by $f_{\dagger}^j\nbigt$. For $\nbigt=(\nbigm_1,\nbigm_2,C)$, the $\nbigr$-triple $f_{\dagger}^j\nbigt$ consists of $(f_{\dagger}^{-j}\nbigm_1,f_{\dagger}^j\nbigm_2,f^j_{\dagger}C)$. \vspace{.1in} Let $H$ be a hypersurface. The sheaf $\distribution_{\vecS\times X/\vecS}(\ast H)$ is naturally $(\nbigr_{X}(\ast H)_{|\vecS\times X}, \sigma^{\ast}\nbigr_{X}(\ast H)_{|\vecS\times X})$-modules. For $\nbigr_X(\ast H)$-modules $\nbigm_i$ $(i=1,2)$, a sesqui-linear pairing of $\nbigm_1$ and $\nbigm_2$ is a $(\nbigr_{X}(\ast H)_{|\vecS\times X}, \sigma^{\ast}\nbigr_{X}(\ast H)_{|\vecS\times X})$-homomorphism $C:\nbigm_{1|\vecS\times X} \times \sigma^{\ast}\nbigm_{2|\vecS\times X} \lrarr \distribution_{\vecS\times X/\vecS}(\ast H)$. For a given $\nbigr_X$-triple $\nbigt=(\nbigm_1,\nbigm_2,C)$, we obtain an $\nbigr_X(\ast H)$-triple $\nbigt(\ast H):= \bigl( \nbigm_1(\ast H), \nbigm_2(\ast H), C(\ast H) \bigr)$, where $C(\ast H)$ is a naturally induced sesqui-linear pairing of $\nbigm_1(\ast H)$ and $\nbigm_2(\ast H)$. \subsubsection{Mixed twistor $\nbigd$-modules} A pure twistor $\nbigd_X$-module of weight $n$ is an $\nbigr_X$-triple $\nbigt=(\nbigm_1,\nbigm_2,C)$ satisfying some conditions. For example, we impose that $\nbigm_i$ are flat over $\nbigo_{\cnum_{\lambda}}$ and that the characteristic variety of $\nbigm_i$ are contained in the product of $\cnum_{\lambda}$ and Lagrangian varieties of $T^{\ast}X$. A polarization of $\nbigt$ is a sesqui-linear duality $\nbigs:\nbigt\lrarr\nbigt^{\ast}\otimes\newTate(-n)$ satisfying some conditions. The precise conditions are given based on the strategy due to M. Saito . We refer the detail to and . A pure twistor $\nbigd$-module which admit a polarization is called a polarizable pure twistor $\nbigd$-module. In this paper, we consider only polarizable pure twistor $\nbigd$-modules, and hence we often omit ``polarizable''. A mixed twistor $\nbigd$-module is an $\nbigr$-triple $\nbigt$ with a finite increasing filtration $W$ indexed by integers satisfying some conditions. For example, we impose that $\Gr^W_n(\nbigt)$ are polarizable pure twistor $\nbigd$-modules of weight $n$. The precise conditions are given based on the strategy due to M. Saito . We refer the detail to . We recall some important properties of mixed twistor $\nbigd$-modules. Let $\MTM(X)$ denote the category of mixed twistor $\nbigd$-modules on $X$. \begin{prop} The category $\MTM(X)$ is abelian. More concretely, we have the following. \begin{itemize} \item Let $\varphi:(\nbigt_1,W_1)\lrarr (\nbigt_2,W_2)$ be a morphism in $\MTM(X)$. Then, the $\nbigr$-triples $\Ker(\varphi)$ and $\Image(\varphi)$ with the induced filtrations are mixed twistor $\nbigd$-modules. Moreover, $\varphi$ is strict with respect to the filtration $W$, and $\Image(\varphi)$ with the induced filtration is a mixed twistor $\nbigd$-module. \end{itemize} Moreover, the category of polarizable pure twistor $\nbigd$-modules of weight $w$ is abelian and semisimple. \hfill\qed \end{prop} \begin{prop} Let $f:X\lrarr Y$ be a projective morphism of complex manifolds. \begin{itemize} \item Let $\nbigt$ be a polarizable pure twistor $\nbigd_X$-module of weight $w$. Then, $f^j_{\dagger}\nbigt$ are polarizable pure twistor $\nbigd_Y$-modules of weight $w+j$. \item Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. Let $W_kf^j_{\dagger}\nbigt$ be the image of $f^j_{\dagger}W_{k-j}\nbigt \lrarr f^j_{\dagger}\nbigt$. Then, $(f^j_{\dagger}\nbigt,W)$ are mixed twistor $\nbigd$-modules. \hfill\qed \end{itemize} \end{prop} \begin{prop} Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. Let $(\nbigm_1,\nbigm_2,C)$ be the underlying $\nbigr_X$-triple. \begin{itemize} \item $\DDD_X\nbigm_i \simeq \nbigh^0\bigl(\DDD_X\nbigm_i\bigr)$. \item We have the naturally induced sesqui-linear pairing $\DDD C$ of $\DDD_X\nbigm_1$ and $\DDD_X\nbigm_2$, and the $\nbigr_X$-triple $\DDD\nbigt=(\DDD\nbigm_1,\DDD\nbigm_2,\DDD C)$ with the induced filtration is a mixed twistor structure. \item For any projective morphism $f:X\lrarr Y$, we have a natural isomorphism $\DDD f^j_{\dagger}\nbigt \simeq f^{-j}_{\dagger}\DDD\nbigt$. \hfill\qed \end{itemize} \end{prop} \subsubsection{Underlying $\nbigd$-modules} Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. The $\nbigr$-triple $\nbigt$ is described as $(\nbigm_1,\nbigm_2,C)$. Let $\iota_1:X\lrarr \nbigx$ be given by $\iota_1(P)=(1,P)$. We obtain the $\nbigd$-module $\Xi_{DR}(\nbigt):= \iota_1^{-1}\bigl(\nbigm_2/(\lambda-1)\nbigm_2\bigr)$ which we call the $\nbigd$-module underlying $\nbigt$. It is naturally equipped with the filtration $W$. For any morphism of mixed twistor $\nbigt$-modules $\varphi: (\nbigt^{(1)},W) \lrarr (\nbigt^{(2)},W)$, we have the induced morphism of filtered $\nbigd$-modules $\Xi_{DR}(\varphi): \Xi_{DR}(\nbigt^{(1)},W) \lrarr \Xi_{DR}(\nbigt^{(2)},W)$. The following lemma is easy to see. \begin{lem} Let $\varphi: (\nbigt^{(1)},W) \lrarr (\nbigt^{(2)},W)$ be any morphism of mixed twistor $\nbigt$-modules. \begin{itemize} \item We have $\Ker\Xi_{DR}(\varphi) =\Xi_{DR}(\Ker(\varphi))$, $\Image\Xi_{DR}(\varphi) =\Xi_{DR}(\Image(\varphi))$, and $\Cok\Xi_{DR}(\varphi) =\Xi_{DR}(\Cok(\varphi))$. \item $\varphi$ is an epimorphism (resp. monomorphism) if and only if $\Xi_{DR}(\varphi)$ is an epimorphism (resp. monomorphism). \hfill\qed \end{itemize} \end{lem} The following lemma is easy to see by construction of the functors. \begin{lem} \mbox{{}} Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. \begin{itemize} \item Let $f:X\lrarr Y$ be a projective morphism of complex manifolds. Then, we have $f^i_{+}\Xi_{DR}(\nbigt) \simeq \Xi_{DR}(f^i_{\dagger}\nbigt)$. The induced filtrations $W$ are also equal. \item We have a natural isomorphism $\DDD_X\Xi_{DR}(\nbigt) \simeq \Xi_{DR}\DDD\nbigt$. The induced filtrations $W$ are also equal. \hfill\qed \end{itemize} \end{lem} \subsubsection{Integrable case} Let $\nbigd_{\nbigx}$ be the sheaf of holomorphic differential operators on $\nbigx$. Let $\nbigrtilde_X\subset\nbigd_{\nbigx}$ be the sheaf of subalgebras generated by $\nbigr_X$ and $\lambda^2\del_{\lambda}$. The correspondence $\nbigm\lrarr \lambda^{-\dim X}\Omega_X \otimes\nbigm$ also induces an equivalence between the category of left $\nbigrtilde_X$-modules and the category of right $\nbigrtilde_X$-modules. An $\nbigrtilde_X$-module is equivalent to an $\nbigo_{\nbigx}$-module with a meromorphic flat connection $\nabla: \nbigm\lrarr \Omega^1_{\nbigx}(\log \nbigx^0)\otimes\nbigo(\nbigx^0) \otimes\nbigm$. Let $\nbigm_i$ $(i=1,2)$ be $\nbigrtilde_X$-modules. Let $\theta$ be the polar coordinate of $\vecS$. We have $\del_{\theta}= \sqrt{-1}\bigl( \lambda\del_{\lambda} -\lambdabar\del_{\lambdabar} \bigr)$. It naturally acts on $\nbigm_i$. A sesqui-linear pairing of $\nbigm_1$ and $\nbigm_2$ is a $(\nbigr_{X|\vecS\times X}, \sigma^{\ast}\nbigr_{X|\vecS\times X})$-homomorphism $C:\nbigm_{1|\vecS\times X} \times \sigma^{\ast}\nbigm_{2|\vecS\times X} \lrarr \distribution_{\vecS\times X/\vecS}$ which is compatible with the action of $\del_{\theta}$. Such a tuple $(\nbigm_1,\nbigm_2,C)$ is called an $\nbigrtilde_X$-triple or integrable $\nbigr_X$-triple. For any morphism of complex manifolds $f:X\lrarr Y$, the $\nbigr_X$-triples $f^j_{\dagger}\nbigt$ are naturally integrable if $\nbigt$ is integrable. A mixed twistor $\nbigd$-module $(\nbigt,W)$ is called integrable if the $\nbigr$-triple $\nbigt$ is integrable and the filtration $W$ is compatible with the action of $\lambda^2\del_{\lambda}$. Propositions and are naturally extended to the integrable case. Let $(\nbigt,W)$ be an integrable mixed twistor $\nbigd_X$-module. Take an injective resolution $\nbigi$ of $\nbigr_X\otimes\Omega_X^{-1}[\dim X]$ as a left $(\nbigrtilde_X,\nbigr_X)$-module. Then, $\DDD_X\nbigm_i\simeq \nrhom_{\nbigr_X}(\nbigm,\nbigi)$ are naturally $\nbigrtilde_X$-modules. The action of $\lambda^2\del_{\lambda}$ is given by $(\lambda^2\del_{\lambda}f)(m) =\lambda^2\del_{\lambda}(f(m)) -f(\lambda^2\del_{\lambda}m)$. It is shown that $\DDD C$ is compatible with the action of $\del_{\theta}$. (See .) Hence, $\DDD(\nbigt,W)$ is also naturally integrable. We shall give some complements on the duality and the push-forward of $\nbigrtilde$-modules in \S. \subsection{Integrable mixed twistor $\nbigd$-modules associated to non-degenerate functions} \subsubsection{Pure twistor $\nbigd$-modules associated to meromorphic functions} Let $X$ be a complex manifold with a hypersurface $D$. We set $\nbigx:=\cnum_{\lambda}\times X$ and $\nbigd:=\cnum_{\lambda}\times D$. Let $f$ be a meromorphic function on $(X,D)$. We have a wild harmonic bundle $E(f)$ on $(X,D)$. It consists of a Higgs bundle $(\nbigo_{X\setminus D}\,e,df)$ with a pluri-harmonic metric $h$ determined by $h(e,e)=1$, where $e$ denotes a global frame of the line bundle. Recall that we have the associated polarizable pure twistor $\nbigd$-module $(\nbigl(f),\nbigl(f),C(f))$ of weight $0$ on $X$, where $\nbigl(f)$ is an $\nbigr_X$-module, and $C(f)$ is a sesqui-linear pairing $\nbigl(f)_{|\vecS\times X} \times \sigma^{\ast}\nbigl(f)_{|\vecS\times X} \lrarr \distribution_{\vecS\times X/\vecS}$. Here $\vecS:=\{|\lambda|=1\}\subset\cnum_{\lambda}$. A polarization is given by $(\id,\id)$. \vspace{.1in} Let $d_X:=\dim X$. We have the polarizable pure twistor $\nbigd$-module $\nbigt(f):=\bigl( \lambda^{d_X}\nbigl(f), \nbigl(f), C(f) \bigr)$ of weight $d_X$. The natural polarization is given by $\bigl((-1)^{d_X},(-1)^{d_X}\bigr)$. \vspace{.1in} The restriction $\nbigl(f)_{|\nbigx\setminus \nbigd}$ is equipped with a global frame $\upsilon$ such that $\nbigl(f)_{|\nbigx\setminus \nbigd}$ is isomorphic to $\nbigo_{\nbigx\setminus\nbigd}\upsilon$ with the family of flat $\lambda$-connections $\DD$ determined by $\DD\upsilon=\upsilon df$. We have \begin{equation} C(f)(\upsilon,\sigma^{\ast}\upsilon) =\exp\bigl( -\lambda \overline{f} +\lambda^{-1}f \bigr). \end{equation} In general, it is not easy to describe $\nbigl(f)$ explicitly. The following lemma is clear by the construction of $\nbigl(f)$ in , and it will be used implicitly. \begin{lem} $\upsilon$ is naturally extended as a section of $\nbigl(f)(\ast D)$, and it gives a global frame of $\nbigl(f)(\ast D)$ as an $\nbigo_{\nbigx}(\ast D)$-module. Moreover, the natural morphism $\nbigl(f)\lrarr\nbigl(f)(\ast D)$ is a monomorphism. \hfill\qed \end{lem} \begin{lem} If $|(f)_0|\cap D=\emptyset$ and $|(f)_{\infty}|=D$, then $\nbigl(f)$ is naturally isomorphic to $\nbigo_{\nbigx}(\ast D)\upsilon$. \end{lem} \pf If $D$ is normal crossing, the claim follows from the construction of $\nbigl(f)$ in . Although we are mainly interested in the case where $D$ is normal crossing, we give a proof in the general case. We take a projective morphism $\varphi:X'\lrarr X$ of complex manifolds such that (i) $D':=\varphi^{-1}(D)$ is normal crossing, (ii) $\varphi$ induces an isomorphism $X'\setminus D'\simeq X\setminus D$. We set $f':=\varphi^{\ast}(f)$. We have $\nbigl(f')=\nbigo_{\nbigx'}(\ast D')v'$ with a frame $v'$ such that $\DD v'=v'\,df'$. Because $\nbigl(f')(\ast D')=\nbigl(f')$, we obtain $\varphi_{\dagger}\bigl(\nbigl(f')\bigr) \simeq \varphi_{\dagger}\bigl(\nbigl(f')\bigr)(\ast D)$. We also have that $\nbigl(f)$ is a direct summand of $\varphi_{\dagger}\nbigl(f')$ and $\varphi_{\dagger}\bigl(\nbigl(f')\bigr) _{|X\setminus D} =\nbigl(f)_{|X\setminus D}$. Then, we can deduce the claim of the lemma. \hfill\qed \begin{prop} Suppose that (i) $D$ is normal crossing, (ii) $f$ is non-degenerate along $D$, (iii) $D=|(f)_{\infty}|$. Then, $\nbigl(f)$ is naturally isomorphic to $\nbigo_{\nbigx}(\ast D)\upsilon$. \end{prop} \pf We have only to check the claim locally around any point of $|(f)_{0}|\cap |(f)_{\infty}|$. Let $(z_1,\ldots,z_n)$ be a convenient coordinate system with $f=z_n\prod_{i=1}^{\ell}z_i^{-k_i}$ for some $k_i>0$. Note that $\prod_{i=1}^{\ell}z_i^N\upsilon$ is a section of $\nbigl(f)$. We have $\deldel_n\upsilon=\upsilon \prod_{i=1}^{\ell}z_i^{-k_i}$, where $\deldel_n=\lambda\del_{x_n}$. Then, it is easy to deduce that $\nbigl(f)\supset \nbigr_{X}\upsilon=\nbigo_{\nbigx}(\ast \nbigd)\upsilon$. \hfill\qed \subsubsection{Mixed twistor $\nbigd$-modules associated to meromorphic functions} For $\star=\ast,!$, we have the $\nbigr_X$-modules $\nbigl(f)[\star D]$ obtained as the localization $\nbigl(f)$. They are denoted by $\nbigl_{\star}(f,D)$ in this paper. If $D=|(f)_{\infty}|$, they are also denoted by $\nbigl_{\star}(f)$. We have the natural morphisms of $\nbigr_X$-modules $\nbigl_!(f,D) \lrarr \nbigl(f) \lrarr \nbigl_{\ast}(f,D)$. It induces $\nbigl_{\star}(f,D)(\ast D) \simeq \nbigl(f)(\ast D)$. For $\star=\ast,!$, we have the mixed twistor $\nbigd$-modules $\nbigt(f)[\star D]$ obtained as the localizations of $\nbigt(f)$. They are denoted by $\nbigt_{\star}(f,D)$ in this paper. If $D=|(f)_{\infty}|$, they are also denoted by $\nbigt_{\star}(f)$. The underlying $\nbigr_X$-triples of $\nbigt_{\ast}(f,D)$ and $\nbigt_!(f,D)$ are $\bigl( \lambda^{d_X}\nbigl_!(f,D), \nbigl_{\ast}(f,D), C(f)[\ast D] \bigr)$ and $\bigl( \lambda^{d_X}\nbigl_{\ast}(f,D), \nbigl_{!}(f,D), C(f)[!D] \bigr)$, respectively. The weight filtrations of $\nbigt_{\star}(f,D)$ are denoted by $W$. We have natural morphisms of mixed twistor $\nbigd$-modules $\nbigt_!(f,D)\lrarr \nbigt(f) \lrarr \nbigt_{\ast}(f,D)$. The meromorphic family of flat connections $\DD^f$ on $\nbigl(f)(\ast D)$ is naturally extended to a meromorphic flat connection $\nabla$ given by $\nabla \upsilon=\upsilon d(\lambda^{-1}f)$. Let $C(f)(\ast D)$ be the sesqui-linear pairing on $\nbigl(f)(\ast D)$ induced by $C(f)$. Due to (), $C(f)(\ast D)$ is compatible with the actions of $\lambda^{2}\del_{\lambda}$. \begin{lem} $\nbigt_{\star}(f,D)$ are integrable mixed twistor $\nbigd$-modules, and $\nbigt(f,D)$ is an integrable pure twistor $\nbigd$-module. The morphisms $\nbigt_{!}(f,D) \lrarr \nbigt(f,D) \lrarr \nbigt_{\ast}(f,D)$ are integrable. \end{lem} \pf By using \cite[Lemma 3.2.4]{Mochizuki-MTM}, we obtain that the underlying $\nbigr$-triples of $\nbigt_{\star}(f,D)$ are integrable. Then, we obtain that $\nbigt(f,D)$ is an integrable pure twistor $\nbigd$-module. By using \cite[Lemma 7.1.35]{Mochizuki-MTM}, we obtain that the filtrations $W$ of $\nbigt_{\star}(f,D)$ are also integrable. Hence, we obtain that $\nbigt_{\star}(f,D)$ are integrable mixed twistor $\nbigd$-modules. By using \cite[Lemma 7.1.37]{Mochizuki-MTM}, we obtain the integrability of the morphisms. \hfill\qed \begin{lem} The $\nbigd$-modules $\Xi_{DR}(\nbigt_{\star}(f,D))$ and $\Xi_{DR}(\nbigt(f))$ are naturally identified with $L_{\star}(f,D)$ and $L(f)$. The morphisms $\Xi_{DR}(\nbigt_!(f,D)) \lrarr \Xi_{DR}(\nbigt(f)) \lrarr \Xi_{DR}(\nbigt_{\ast}(f,D))$ are naturally identified with $L_!(f,D)\lrarr L(f)\lrarr L_{\ast}(f,D)$. \end{lem} \pf Let $\nbigm$ be an $\nbigr_X$-module underlying $\nbigt_{\star}(f,D)$ $(\star=\ast,!)$ or $\nbigt(f,D)$. Let $g$ be a holomorphic function on an open subset $U\subset X$. Because $\nbigt_{\star}(f,D)$ are integrable, the KMS-spectrum of $\nbigm_{|U}$ along $g$ are contained in $\real\times\{0\}$. Hence, the specialization of the $V$-filtration along $g$ at $\lambda=1$ is equal to the $V$-filtration of $\iota_1^{-1}(\nbigm/(\lambda-1)\nbigm)$, where $\iota_1:X\lrarr \nbigx$ given by $\iota_1(P)=(1,P)$. Then, the claim is clear. \hfill\qed \vspace{.1in} In the proof of Lemma , the KMS-spectrum of the $\nbigr_X$-modules are contained in $\rnum\times\{0\}$, indeed. We can directly check it by standard computations. \begin{lem} Suppose that (i) $D=|(f)_{\infty}|$, (ii) $f$ is pure at $P\in X$. Then, the canonical morphisms $\nbigt_{!}(f,D) \lrarr \nbigt(f) \lrarr \nbigt_{\ast}(f,D)$ are isomorphisms on a neighbourhood of $P$. \end{lem} \pf It follows from that the morphisms of the underlying $\nbigd$-modules are isomorphisms. \hfill\qed \begin{cor} Suppose that (i) $D$ is normal crossing, (ii) $D=|(f)_{\infty}|$, (iii) $f$ is non-degenerate. Then the canonical morphisms $\nbigt_{!}(f) \lrarr \nbigt(f) \lrarr \nbigt_{\ast}(f)$ are isomorphisms. In particular, we have isomorphisms $\nbigl_{\star}(f,D)\simeq \nbigo_{\nbigx}(\ast D)\,\upsilon$. \hfill\qed \end{cor} \subsubsection{Expression of $\nbigrtilde$-modules in non-degenerate cases} Suppose that $D$ is normal crossing. Let us describe $\nbigr_X$-modules $\nbigl_{\star}(f,D)$ in the case that $f$ is non-degenerate along $D$. We do not assume $D=|(f)_{\infty}|$. Let $V_D\nbigr_X$ denote the sheaf of subalgebras in $\nbigr_X$ generated by $\lambda\cdot p_{\lambda}^{\ast}\Theta_{X}(\log D)$ over $\nbigo_{\nbigx}$. Let $V_D\nbigrtilde_X$ denote the sheaf of subalgebras in $\nbigrtilde_X$ generated by $V_D\nbigr_X$ and $\lambda^2\del_{\lambda}$. We have the $V_D\nbigrtilde_X$-modules $\nbigl(f)$ and $\nbigl(f)(\nbigd) =\nbigl(f)\otimes_{\nbigo_{\nbigx}}\nbigo_{\nbigx}(\nbigd)$. \begin{prop} We have natural isomorphisms of $\nbigrtilde_X$-modules \[ \nbigl_{\ast}(f,D)\simeq \nbigr_X\otimes_{V_D\nbigr_X} \nbigl(f)\bigl(\nbigd\bigr), \quad \nbigl_{!}(f,D)\simeq \nbigr_X\otimes_{V_D\nbigr_X}\nbigl(f). \] \end{prop} \pf We have natural isomorphisms of $\nbigrtilde_X(\ast D)$-modules \[ \nbigl_{\ast}(f,D)(\ast D) \simeq \bigl( \nbigr_X\otimes_{V_D\nbigr_X} \nbigl(f)\bigl(\nbigd\bigr) \bigr)(\ast D), \quad \nbigl_{!}(f,D)(\ast D)\simeq \bigl( \nbigr_X\otimes_{V_D\nbigr_X}\nbigl(f) \bigr)(\ast D). \] Note that if we are given an $\nbigr_X$-endomorphism $g$ of $\nbigl_{\star}(f,D)$ such that $g_{|\cnum_{\lambda}\times (X\setminus D)}=0$, then we have $g=0$. Hence, it is enough to prove that there exist isomorphisms as $\nbigr_X$-modules. The claim is clear outside $|(f)_0|\cap|(f)_{\infty}|$ by the construction of $\nbigl_{\star}(f,D)$ in \cite[\S5.3]{Mochizuki-MTM}. We set \[ \nbigl'_{\ast}(f,D):= \nbigr_X\otimes_{V_D\nbigr_X} \nbigl(f)\bigl(\nbigd\bigr), \quad\quad \nbigl'_{!}(f,D):= \nbigr_X\otimes_{V_D\nbigr_X}\nbigl(f). \] Let $Q\in|(f)_0|\cap|(f)_{\infty}|$. We take a convenient coordinate system $(z_1,\ldots,z_n)$ with $D=\bigcup_{i=1}^{\ell_1+\ell_2}\{z_i=0\}$ and $f=z_n\prod_{i=1}^{\ell_1}z_i^{-k_i}$. We set $X_1:=\{(z_1,\ldots,z_{\ell_1},z_n)\}$ and $X_2:=\{(z_{\ell_1+1},\ldots,z_{n-1})\}$. We set $f_1:=z_n\prod_{i=1}^{\ell_1}z_i^{-k_i}$ and $D_1:=\bigcup_{i=1}^{\ell_1}\{z_i=0\}$ and $D_2:=\bigcup_{i=\ell_1+1}^{\ell_1+\ell_2}\{z_i=0\}$. We have $\nbigl'_{\star}(f_1,D_1) =\nbigl(f_1) =\nbigl_{\star}(f_1,D_1)$. In particular, we have $\nbigl'_{\star}(f_1,D_1)[\star z_j] =\nbigl'_{\star}(f_1,D_1)$ for $j=1,\ldots,\ell_1$. We also have $\nbigl'_{\star}(f_2,D_2) \simeq \nbigl_{\star}(f_2,D_2)$. Hence, $\nbigl'_{\star}(f_2,D_2)[\star z_i] \simeq \nbigl'_{\star}(f_2,D_2)$ for $i=\ell_1+1,\ldots,\ell_1+\ell_2$. We have $\nbigl'_{\star}(f,D) \simeq \nbigl'_{\star}(f_1,D_1) \boxtimes \nbigl'_{\star}(f_2,D_2)$. Then, it is easy to check that $\nbigl'_{\star}(f,D)[\star z_i] =\nbigl'_{\star}(f,D)$ for $i=\ell_1+1,\ldots,\ell_1+\ell_2$. We obtain the desired isomorphism by the characterization in \cite[Proposition 5.3.1]{Mochizuki-MTM}. \hfill\qed \subsubsection{De Rham complexes in non-degenerate cases} Let us consider the case that $(X,D)=S\times (X_0,D_0)$, where $X_0$ is a complex manifold with a normal crossing hypersurface $D_0$. Let $f$ be a meromorphic function on $(X,D)$ which is non-degenerate along $D$. Let $\pi:X\lrarr S$ denote the projection. We have the mixed twistor $\nbigd$-modules $\pi_{\dagger}^i\nbigt_{\star}(f,D)$ on $S$. Let us describe the underlying $\nbigr_{S}$-modules $\pi_{\dagger}^i\nbigl_{\star}(f,D)$, by assuming that $X_0$ is projective. Set $n:=\dim X_0$. Let $p_{\lambda}:\nbigx\lrarr X$ denote the projection. For any $\nbigr_X$-module $\nbigm$, we set \[ \DR_{X/S}^{m}\nbigm:= \lambda^{-m-n} p_{\lambda}^{\ast} \bigl( \Omega^{m+n}_{X/S} \bigr) \otimes_{\nbigo_{\nbigx}} \nbigm. \] Then, we obtain the complex $\DR_{X/S}^{\bullet}\nbigm$. We naturally have $\pi_{\dagger}^i\nbigm \simeq \hyperr^i\pi_{\ast} \DR^{\bullet}_{X/S}\nbigm$. \vspace{.1in} Suppose $|(f)_{\infty}|=D$. We have $\nbigl_{\star}(f,D)=\nbigl(f)$ which is isomorphic to $\nbigo_{\nbigx} (\ast D)\upsilon$ with $\DD\upsilon=\upsilon df$. We set $\Omegabar^{\ell}_{X/S}:= \lambda^{-\ell} p_{\lambda}^{\ast}\Omega_{X/S}^{\ell}$. Then, we immediately obtain the following natural isomorphism: \[ \DR^{\bullet}_{X/S}\nbigl(f) \simeq \Bigl( \Omegabar^{\bullet+n}_{X/S}(\ast D), d+\lambda^{-1}df \Bigr) \] Let us consider the case that $D$ is not necessarily $|(f)_{\infty}|$. We set $\Omegabar^{\ell}_{X/S}(\log D):= \lambda^{-\ell} p_{\lambda}^{\ast}\bigl( \Omega^{\ell}_{X/S}(\log D)\bigr)$. We obtain the complexes $\bigl( \Omegabar^{\bullet}_{X/S}(\log D),d+\lambda^{-1}df \bigr)$ and $\bigl( \Omegabar^{\bullet}_{X/S}(\log D)(-\nbigd), d+\lambda^{-1}df \bigr)$. \begin{lem} We have the following natural quasi-isomorphisms: \begin{equation} \DR^{\bullet}_{X/S}\nbigl_{\ast}(f,D) \simeq \Bigl( \Omegabar^{\bullet+n}_{X/S}(\log D)\bigl(\ast(f)_{\infty}\bigr) ,d+\lambda^{-1}df \Bigr) \end{equation} \begin{equation} \DR^{\bullet}_{X/S}\nbigl_{!}(f,D) \simeq \Bigl( \Omegabar^{\bullet+n}_{X/S}(\log D)(-\nbigd) \bigl(\ast(f)_{\infty}\bigr) ,d+\lambda^{-1}df \Bigr) \end{equation} Hence, $\pi^i_{\dagger} \nbigl_{\star}(f,D)$ $(\star=\ast,!)$ are expressed as the push-forward of the right hand side of {\rm ()} and {\rm()}. \end{lem} \pf It follows from the expression in Proposition . \hfill\qed \vspace{.1in} Suppose moreover that $f$ is non-degenerate along $D$ over $S$. We use the notation in \S. We set $\Omegabar^{\ell}_{X/S}(\log D,f):= \Omegabar^{\ell}_{X/S}(\log D)\bigl(\ell(f)_{\infty}\bigr)$, and we obtain the following complexes: \[ \bigl( \Omegabar^{\bullet}_{X/S}(\log D,f),d+\lambda^{-1}df \bigr), \quad \bigl( \Omegabar^{\bullet}_{X/S}(\log D,f)(-\nbigd) ,d+\lambda^{-1}df \bigr). \] Similarly, we set $\Omegabar^{\ell}_{X/S,f,D}:= \lambda^{-\ell} p_{\lambda}^{\ast} \Omega^{\ell}_{X/S,f,D}$, and we obtain the following complexes: \[ \bigl( \Omegabar^{\bullet}_{X/S,f,D}, d+\lambda^{-1}df \bigr), \quad\quad \bigl( \Omegabar^{\bullet}_{X/S,f,D}(-\nbigd), d+\lambda^{-1}df \bigr) \] As in the case of $\nbigd$-modules in \S and \S, we have the following lemma. \begin{lem} Suppose that $f$ is non-degenerate along $D$ over $S$. Then, we have the following natural quasi-isomorphisms: \begin{equation} \DR^{\bullet}_{X/S}\nbigl_{\ast}(f,D) \simeq \Bigl( \Omegabar^{\bullet+n}_{X/S}(\log D,f) ,d+\lambda^{-1}df \Bigr) \simeq \bigl( \Omegabar^{\bullet}_{X/S,f,D}, d+\lambda^{-1}df \bigr) \end{equation} \begin{equation} \DR^{\bullet}_{X/S}\nbigl_{!}(f,D) \simeq \Bigl( \Omegabar^{\bullet+n}_{X/S}(\log D,f)(-\nbigd) ,d+\lambda^{-1}df \Bigr) \simeq \bigl( \Omegabar^{\bullet}_{X/S,f,D}(-\nbigd), d+\lambda^{-1}df \bigr) \end{equation} Hence, $\pi_{\dagger}^i\nbigl_{\star}(f,D)$ are expressed as the push-forward of the complexes in {\rm()} and {\rm()}. \hfill\qed \end{lem} We also remark the following lemma. \begin{lem} Suppose that $f$ is non-degenerate along $D$ over $S$. Then, $\pi^i_{\dagger}\nbigl_{\star}(f,D)$ are locally free $\nbigo_{\cnum_{\lambda}\times S}$-modules. \end{lem} \pf The specialization at $\lambda=1$ are locally free $\nbigo_S$-modules by Corollary . So, by the general property of mixed twistor $\nbigd$-modules, we obtain that $\pi^i_{\dagger}\nbigl_{\star}(f,D)$ are locally free $\nbigo_{\cnum_{\lambda}\times S}$-modules. \hfill\qed \begin{cor} If $f$ is non-degenerate along $D$ over $S$, the $\nbigo_{\cnum_{\lambda}\times S}$-modules \[ \hyperr^i\pi_{\ast} \bigl( \Omegabar^{\bullet}_{X/S}(\log D)(\ast (f)_{\infty}),d+\lambda^{-1}df \bigr) \simeq \hyperr^i\pi_{\ast} \bigl( \Omegabar^{\bullet}_{X/S}(\log D,f),d+\lambda^{-1}df \bigr) \simeq \hyperr^i\pi_{\ast} \bigl( \Omegabar^{\bullet}_{X/S,f,D},d+\lambda^{-1}df \bigr) \] are locally free. The $\nbigo_{\cnum_{\lambda}\times S}$-modules \begin{multline*} \hyperr^i\pi_{\ast} \bigl( \Omegabar^{\bullet}_{X/S}(\log D) (\ast (f)_{\infty})(-\nbigd), d+\lambda^{-1}df \bigr) \simeq \hyperr^i\pi_{\ast} \bigl( \Omegabar^{\bullet}_{X/S}(\log D,f)(-\nbigd), d+\lambda^{-1}df \bigr) \\ \simeq \hyperr^i\pi_{\ast} \bigl( \Omegabar^{\bullet}_{X/S,f,D}(-\nbigd), d+\lambda^{-1}df \bigr) \end{multline*} are also locally free. \hfill\qed \end{cor} \subsection{The real structure of the localization of some mixed twistor $\nbigd$-modules} Let $X$ be a complex manifold. For a mixed twistor $\nbigd$-module $\nbigt=(\nbigm_1,\nbigm_2,C)$ on $X$, we have \[ j^{\ast}\nbigt=(j^{\ast}\nbigm_1,j^{\ast}\nbigm_2,j^{\ast}C), \quad \DDD\nbigt= (\DDD\nbigm_1,\DDD\nbigm_2,\DDD C). \] We set $\gammatilde^{\ast}\nbigt:= (j^{\ast}\DDD\nbigm_2,j^{\ast}\DDD\nbigm_1, j^{\ast}\DDD C^{\ast})$. We will naturally identify $j^{\ast}\DDD\nbigm_i$ and $\DDD j^{\ast}\nbigm_i$. Recall that a real structure of $\nbigt$ is an isomorphism $\kappa: \gammatilde^{\ast} \nbigt \simeq \nbigt$ satisfying $\gammatilde^{\ast}\kappa\circ\kappa=\id$. \subsubsection{Basic case} Let $d_X:=\dim X$. We have the isomorphism $\nu_X: \DDD\nbigo_{\nbigx}\simeq \nbigo_{\nbigx}\lambda^{d_X}$, whose specialization at $\{\lambda_0\}\times X$ $(\lambda_0\neq 0)$ is equal to the morphism in \S. We have the natural identification $j^{\ast}\nbigo_{\nbigx}= \nbigo_{\nbigx}$ given by the pull back of functions. We have $\DDD\nu_X=(-1)^{d_X}\nu_X$. As shown in , the isomorphism \[ \bigl( \nu_X^{-1}, (-1)^{d_X}\nu_X \bigr): \gammatilde^{\ast}\nbigu_X(d_X,0) \simeq \nbigu_X(d_X,0) \] gives a real structure. Let $Y$ be a smooth hypersurface of $X$. We set $d_Y:=\dim Y=d_X-1$. Let $\iota:Y\lrarr X$ be the inclusion. We have the following natural morphisms of integrable mixed twistor $\nbigd$-modules: \begin{equation} \nbigu_X(d_X,0)[\ast Y] \lrarr \iota_{\dagger}\nbigu_{Y}(d_Y,0) \otimes\newTate(-1) \end{equation} \begin{equation} \iota_{\dagger}\nbigu_{Y}(d_Y,0) \lrarr \nbigu_X(d_X,0)[!Y] \end{equation} The morphisms are induced by $\nbigr_X$-homomorphisms $\nbigo_{\nbigx}[\ast Y] \lrarr \iota_{\dagger}\nbigo_{\nbigy}\lambda^{-1}$ and $\iota_{\dagger}\nbigo_{\nbigy} \lrarr \nbigo_{\nbigx}[!Y]$. Locally, if $X$ is equipped with a holomorphic coordinate system $(x_1,\ldots,x_n)$ such that $Y=\{x_1=0\}$, the morphisms are given by $x_1^{-1} \longmapsto \lambda^{-1} \iota_{\ast}(dx_1/\lambda)^{-1}$ and $\iota_{\ast}(dx_1/\lambda)^{-1} \longmapsto -\deldel_{1}(1)$, where $\deldel_1=\lambda\del_{x_1}$. \begin{prop} The natural morphisms {\rm()} and {\rm()} are compatible with the real structures. \end{prop} \pf Let us look at the compatibility of {\rm()} which is the commutativity of the following diagrams: \[ \begin{CD} \nbigo_{\nbigx}[!Y] \lambda^{d_X} @<{\nu_X}<< \DDD\bigl(\nbigo_{\nbigx}[\ast Y]\bigr) \\ @AAA @AAA \\ \iota_{\dagger}\nbigo_{\nbigy}\lambda^{d_Y+1} @<{-\nu_Y}<< \DDD\bigl( \iota_{\dagger}\nbigo_{\nbigy}\lambda^{-1} \bigr) \end{CD} \quad\quad\quad \begin{CD} \nbigo_{\nbigx}[\ast Y] @<{(-1)^{d_X}\nu_X}<< \DDD\bigl(\nbigo_{\nbigx}[!Y]\lambda^{d_X}\bigr) \\ @VVV @VVV \\ \iota_{\dagger}\nbigo_{\nbigy}\lambda^{-1} @<{(-1)^{d_Y+1}\nu_Y}<< \DDD\bigl( \iota_{\dagger}\nbigo_{\nbigy}\lambda^{d_Y+1} \bigr) \end{CD} \] To check the commutativity, it is enough to compare the specialization along $\{\lambda_0\}\times X$ for any $\lambda_0\neq 0$. Then, it is reduced to Proposition . We can check the compatibility of () similarly. \hfill\qed \subsubsection{Mixed twistor $\nbigd$-modules associated to holomorphic functions} Let $F$ be any holomorphic function on $X$. We have the $\nbigr_X$-module $\nbigl(F)$ given by $\nbigo_{\nbigx}\,e$ with $\DD e=e\,dF$, where $e$ is a global frame. Let $\nbigl(F)^{\lor}= \nhom_{\nbigo_{\nbigx}}(\nbigl(F),\nbigo_{\nbigx})$ is naturally isomorphic to $\nbigo_{\nbigx}e^{\lor}$ with $\DD e^{\lor}=e^{\lor}(-dF)$, where $e^{\lor}$ is the dual frame of $e$. Hence, we have the isomorphism $j^{\ast}\nbigl(F)^{\lor} \simeq \nbigl(F)$ given by $j^{\ast}e^{\lor}\longleftrightarrow e$. We have the smooth $\nbigr_X$-triple $\nbigt_{sm}(F):=(\nbigl(F),\nbigl(F),C_F)$. The isomorphism $\gammatilde_{sm}^{\ast}\nbigt_{sm}(F):= (j^{\ast}\nbigl(F)^{\lor},j^{\ast}\nbigl(F),j^{\ast}C_F^{\lor}) \simeq \nbigt(F)$ is a real structure of $\nbigt_{sm}(F)$ as a smooth $\nbigr_X$-triple. (See \cite[\S2.1.7.2]{Mochizuki-MTM}.) Because $\nbigt(F)= \nbigt_{sm}(F)\otimes \nbigu_X(d_X,0)$, we have the induced real structure on $\nbigt(F)$ as a mixed twistor $\nbigd$-module \cite[Proposition 13.4.6]{Mochizuki-MTM}. More explicitly, as in the case of $\nbigo_{\nbigx}$, we have a natural isomorphism $\nu:j^{\ast}\DDD\nbigl(F)\simeq \lambda^{d_X}j^{\ast}\nbigl(F)^{\lor} \simeq \lambda^{d_X}\nbigl(F)$. The real structure of $\nbigt(F)=(\lambda^{d_X}\nbigl(F),\nbigl(F),C_F)$ is given by $(\nu^{-1},(-1)^{d_X}\nu)$. \begin{lem} We have natural isomorphisms of integrable mixed twistor $\nbigd$-modules with real structure \begin{equation} \nbigt(F)[\star Y] \simeq \nbigt_{sm}(F) \otimes\nbigu_X(d_X,0)[\star Y]. \end{equation} \end{lem} \pf We have the natural isomorphisms of the integrable $\nbigr_X$-modules $\nbigl(F)[\star Y] \simeq \nbigl(F)\otimes\nbigo_{\nbigx}[\star Y]$ $(\star=\ast,!)$ such that the restriction to $X\setminus Y$ is the identity. It is enough to check the claim locally around any point of $Y$. We may assume that $Y$ is given as $\{t=0\}$ for a holomorphic function $t$ such that $dt$ is nowhere vanishing. Then, it is easy to see that the $V$-filtration of $\nbigo_{\nbigx}[\star Y]$ along $t$ induces a $V$-filtration of $\nbigl(F)\otimes\nbigo_{\nbigx}[\star Y]$ along $t$, and that it satisfies the characterization condition for $\nbigl(F)[\star t]$ in \cite[\S3.1]{Mochizuki-MTM}. Because the sesqui-linear pairings of $\nbigt(F)[\star Y]$ and $\nbigt_{sm}(F) \otimes\nbigu_X(d_X,0)[\star Y]$ are the same on $X\setminus Y$, we obtain that they are the same on $X$. (See \cite[Proposition 3.2.1]{Mochizuki-MTM}.) Because the real structures of $\nbigt(F)[\star F]$ are uniquely determined by their restriction to $X\setminus Y$, the isomorphisms () are compatible with the real structures. \hfill\qed \vspace{.1in} We also obtain the integrable mixed twistor $\nbigd$-modules $\nbigt_{sm}(F)\otimes\iota_{\dagger}\nbigu_Y(d_Y,0)$ with real structure. We have the isomorphisms of integrable mixed twistor $\nbigd$-modules with real structure: \begin{equation} \nbigt(F)[\ast Y]/\nbigt(F) \simeq \nbigt_{sm}(F)\otimes\iota_{\dagger}\nbigu_Y(d_Y,0) \otimes\newTate(-1) \end{equation} \begin{equation} \Ker\bigl( \nbigt(F)[!Y]\lrarr \nbigt(F) \bigr) \simeq \nbigt_{sm}(F)\otimes \iota_{\dagger}\nbigu_Y(d_Y,0) \end{equation} Let $F_Y$ denote the restriction of $F$ to $Y$. We have the integrable mixed twistor $\nbigd$-module $\nbigt(F_Y)$ on $Y$ with real structure on $Y$. It is given as $\nbigt(F_Y)= \bigl(\lambda^{d_Y}\nbigl(F_Y),\nbigl(F_Y),C_{F_Y} \bigr)$. Let $\omega_{Y/X}$ denote the conormal bundle of $Y$ in $X$. We set $\omega_{\nbigy/\nbigx}:= \lambda \cdot p_{\lambda}^{\ast}\omega_{Y/X}$, where $p_{\lambda}:\nbigy\lrarr X$ denotes the projection. We have the $\nbigo_{\nbigx}$-submodules \[ \nbige_0:=\nbigl(F)\otimes \iota_{\ast}(\omega_{\nbigy/\nbigx}) \subset \nbigl(F)\otimes\iota_{\dagger}\nbigo_{\nbigx}, \quad \nbige_1:=\iota_{\ast}(\nbigl(F_Y)\otimes\omega_{\nbigy/\nbigx}) \subset \iota_{\dagger}\nbigl(F_Y). \] We have the natural isomorphism of $\nbigo_{\nbigx}$-modules $\varphi_0:\nbige_0\simeq\nbige_1$. \begin{lem} The $\nbigo$-isomorphism $\varphi_0$ is uniquely extended to an isomorphism of integrable $\nbigr_X$-modules $\varphi: \iota_{\dagger}\nbigl(F_Y) \simeq \nbigl(F)\otimes\iota_{\dagger}\nbigo_{\nbigy}$. They give the following isomorphism of integrable pure twistor $\nbigd$-modules with real structure: \[ \nbigt_{sm}(F) \otimes \iota_{\dagger}\nbigu_Y(d_Y,0) \simeq \iota_{\dagger}\nbigt(F_Y) \] \end{lem} \pf Let us observe that $\varphi_0$ is uniquely extended. Because $\nbige_0$ and $\nbige_1$ generate $\nbigl(F)\otimes\iota_{\dagger}\nbigo_{\nbigx}$ and $\iota_{\dagger}\nbigl(F_Y)$ over $\nbigr_X$, the uniqueness is clear. Hence, it is enough to obtain $\varphi$ locally around any point of $Y$. We may assume that $X$ is the product of $Y$ and a neighbourhood of $0$ in $\cnum_t$. We have $Y=\{t=0\}$. Let $V$ denote the $V$-filtration of $\nbigl(F)\otimes\iota_{\dagger}\nbigo_{\nbigy}$ and $\iota_{\dagger}\nbigl(F_Y)$. We may naturally regard $V_0\bigl(\nbigl(F)\otimes\iota_{\dagger}\nbigo_{\nbigy}\bigr)$ and $V_0\bigl( \iota_{\dagger}\nbigl(F_Y) \bigr)$ as integrable $\nbigr_Y$-modules, and the natural isomorphism $\varphi_0$ is an isomorphism of the integrable $\nbigr_Y$-modules $V_0\bigl(\nbigl(F)\otimes\iota_{\dagger}\nbigo_{\nbigy}\bigr) \simeq V_0\bigl( \iota_{\dagger}\nbigl(F_Y) \bigr)$. For both $\nbigrtilde_X$-modules, we have $V_m=\bigoplus_{j\leq m}\deldel_t^jV_0$. We can easily construct the isomorphism $\varphi$ inductively on the $V$-filtration. Because $\varphi_0$ is compatible with the sesqui-linear pairings on $V_0$, we obtain that $\varphi$ gives an isomorphism of the integrable $\nbigr_X$-triples. The comparison of the real structures is easy by construction. \hfill\qed \begin{prop} We have the following isomorphisms \[ \nbigt(F)[\ast Y]/\nbigt(F) \simeq \iota_{\dagger}\nbigt(F_Y)\otimes\newTate(-1) \] \[ \Ker\bigl( \nbigt(F)[! Y]\lrarr\nbigt(F) \bigr) \simeq \iota_{\dagger}\nbigt(F_Y) \] \end{prop} \pf It follows from (), () and Lemma . \hfill\qed \vspace{.1in} Let us describe the isomorphisms of the underlying integrable $\nbigr_X$-modules more explicitly. For simplicity, we assume that $X$ is the product of $Y$ and an open subset in $\cnum_t$. Let $\upsilon$ be the frame of $\nbigl(F)$ such that $\nabla\upsilon=\upsilon d(F/\lambda)$. Let $\upsilon_Y$ be the frame of $\nbigl(F_Y)$ obtained as the restriction of $\upsilon$ to $Y$. The morphism $\nbigl(F)[\ast t]\lrarr \iota_{\dagger}\nbigl(F_Y)$ is given by \begin{equation} t^{-1}\upsilon \longmapsto \iota_{\ast}(\upsilon_Y \lambda^{-1})(dt/\lambda)^{-1} =\iota_{\ast}(\upsilon_Y(dt)^{-1}). \end{equation} Note that we have a natural inclusion $\eta:\nbigl(F)\lrarr \nbigl(F)[!t]$ as $\nbigo_{\nbigx}$-module. The morphism $\iota_{\dagger}\nbigl(F_Y) \lrarr \nbigl(F)[!t]$ is given by \begin{equation} \iota_{\ast}(\upsilon_Y(dt/\lambda)^{-1}) \longmapsto -\deldel_t\eta(\upsilon) +\eta(\del_t(F)\cdot\upsilon) \end{equation} \subsection{Push-forward} We use the notation in \S. We have the twistor $\nbigd$-modules $\nbigt_{\star}\bigl(F,D_Y^{(0)}\bigr)$ and $\nbigt_{\star}\bigl(F,D_Y^{(1)}\bigr)$ $(\star=\ast,!)$. \begin{prop} For $\nbigt= \nbigt_{\ast}(F,D_Y^{(1)}), \nbigt_{\ast}(F,D_Y^{(0)}), \nbigt_{\ast}(F,D_Y^{(1)})\big/ \nbigt_{\ast}(F,D_Y^{(0)})$, we have $\pi_{\dagger}^i(\nbigt)=0$ $(i\neq 0)$. We also have the following natural isomorphisms of the integrable mixed twistor $\nbigd$-modules with real structure: \begin{equation} \pi_{\dagger}^0\nbigt_{\ast}(F,D_Y^{(1)}) \simeq \Bigl( \nbigt_{\ast}(g,D)\otimes \newTate(-1) \Bigr) [!(f)_0][\ast D] \end{equation} \begin{equation} \pi_{\dagger}^0\bigl( \nbigt_{\ast}(F,D_Y^{(1)}) \big/ \nbigt_{\ast}(F,D_Y^{(0)}) \bigr) \simeq \nbigt_{\ast}(g,D)\otimes \newTate(-1) \simeq \nbigt_{\ast}(g,D)\otimes \newTate(-1) [\ast D] \end{equation} \begin{equation} \pi_{\dagger}^0\bigl( \nbigt_{\ast}(F,D_Y^{(0)}) \bigr) \simeq \Ker\left( \Bigl( \nbigt_{\ast}(g,D)\otimes \newTate(-1) \Bigr) [!(f)_0][\ast D] \lrarr \nbigt_{\ast}(g,D)\otimes \newTate(-1) [\ast D] \right) \end{equation} \end{prop} \pf The vanishing $\pi_{\dagger}^i\nbigt=0$ $(i\neq 0)$ follows from Proposition and the twistor property. We set $D^{(2)}:D\times\proj^1$. Let $\iota_0:X\lrarr Y$ be induced by $\{0\}\lrarr \proj^1$. By the argument for Proposition , we have a natural isomorphism of integrable $\nbigr_Y(\ast D^{(2)})$-triples: \[ \bigl( \nbigt_{\ast}(F,D_Y^{(1)}) \big/\nbigt_{\ast}(F,D_Y^{(0)}) \bigr) (\ast D^{(2)}) \simeq \iota_{0\dagger} \nbigt_{\ast}(g,D)(\ast D) \] Hence, we have \[ \pi_{\dagger}^0 \bigl( \nbigt_{\ast}(F,D_Y^{(1)}) \big/\nbigt_{\ast}(F,D_Y^{(0)}) \bigr) (\ast D) \simeq \nbigt_{\ast}(g,D)(\ast D) \] Because we naturally have $\bigl( \nbigt_{\ast}(F,D_Y^{(1)}) \big/\nbigt_{\ast}(F,D_Y^{(0)}) \bigr) [\ast D^{(2)}] \simeq \nbigt_{\ast}(F,D_Y^{(1)}) \big/\nbigt_{\ast}(F,D_Y^{(0)})$ and $\nbigt_{\ast}(g,D)[\ast D] \simeq \nbigt_{\ast}(g,D)$, we obtain the isomorphism () of integrable mixed twistor $\nbigd$-modules. The comparison of the real structures is reduced to Proposition . We have the natural morphisms of integrable mixed twistor $\nbigd$-modules with real structure: \begin{equation} \pi_{\dagger}^0\nbigt_{\ast}\bigl(F,D_Y^{(1)}\bigr) \lrarr \pi_{\dagger}^0\nbigt_{\ast}\bigl(F,D_Y^{(1)}\bigr)[\ast D] \llarr \Bigl( \pi_{\dagger}^0\nbigt_{\ast}\bigl(F,D_Y^{(1)}\bigr) [!(f)_0] \Bigr) [\ast D]. \end{equation} According to Proposition , the induced morphisms of the underlying $\nbigd$-modules are isomorphisms. Hence, we obtain that the morphisms in () are isomorphisms. We have \[ \pi_{\dagger}^0\nbigt_{\ast}(F,D_Y^{(1)}) \bigl(\ast (f)_0\bigr)(\ast D) \simeq \pi_{\dagger}^0\Bigl( \nbigt_{\ast}(F,D_Y^{(1)}) \big/ \nbigt_{\ast}(F,D_Y^{(0)}) \Bigr) \bigl(\ast (f)_0\bigr)(\ast D). \] Hence, we have \[ \pi_{\dagger}^0\nbigt_{\ast}(F,D_Y^{(1)}) \simeq \Bigl( \pi_{\dagger}^0 \bigl( \nbigt_{\ast}(F,D_Y^{(1)}) \big/\nbigt_{\ast}(F,D_Y^{(0)}) \bigr) (\ast D) \Bigr)[!(f)_0][\ast D] \simeq \Bigl( \nbigt_{\ast}(g,D)\otimes\newTate(-1) \Bigr)[!(f)_0][\ast D] \] Thus, we obtain (). We obtain () from the others. \hfill\qed \begin{cor} For $\nbigt= \nbigt_{!}(F,D_Y^{(1)}), \nbigt_{!}(F,D_Y^{(0)}), \Ker\bigl( \nbigt_{!}(F,D_Y^{(1)})\lrarr \nbigt_{!}(F,D_Y^{(0)}) \bigr)$, we have $\pi_{\dagger}^i(\nbigt)=0$ $(i\neq 0)$. We also have the following natural isomorphisms of the integrable mixed twistor $\nbigd$-modules with real structure: \[ \pi_{\dagger}^0\nbigt_{!}(F,D_Y^{(1)}) \simeq \nbigt_{!}(g,D) [\ast(f)_0][!D] \] \[ \pi_{\dagger}^0\Bigl( \Ker\Bigl( \nbigt_{!}(F,D_Y^{(1)}) \lrarr \nbigt_{!}(F,D_Y^{(0)}) \Bigr) \Bigr) \simeq \nbigt_{!}(g,D) \simeq \nbigt_{!}(g,D)[!D] \] \[ \pi_{\dagger}^0\bigl( \nbigt_{!}(F,D_Y^{(0)}) \bigr) \simeq \Cok\Bigl( \nbigt_{!}(g,D)[!D] \lrarr \nbigt_{!}(g,D) [\ast (f)_0][!D] \Bigr) \] \end{cor} \pf Because $\nbigt_{\ast}(F,D_Y^{(a)})^{\ast} \simeq \nbigt_!(F,D_Y^{(a)}) \otimes\newTate(\dim Y)$, the claim follows from Proposition . \hfill\qed \vspace{.1in} Let us consider the case that $f$ is moreover non-degenerate along $D$. In this case $Z_f$ is smooth, and $Z_f\cup D$ is normal crossing. As in \S, let $\iota:Z_f\lrarr X$ denote the inclusion, and we set $D_{Z_f}:=D\cap Z_f$ and $g_0:=g_{|Z_f}$. We have the integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(g_0,D_{Z_f})$ with real structure on $Z_f$. By using Proposition , we obtain the following corollary. \begin{cor} If $f$ is moreover non-degenerate along $D$, we have the following isomorphisms of the integrable mixed twistor $\nbigd$-modules with real structure: \begin{equation} \pi^0_{\dagger}\nbigt_{\ast}(F,D_Y^{(0)}) \simeq \iota_{\dagger} \nbigt_{\ast}(g_0,D_{Z_f})\otimes\newTate(-1) \end{equation} \begin{equation} \pi^0_{\dagger}\nbigt_!(F,D_Y^{(0)}) \simeq \iota_{\dagger} \nbigt_{!}(g_0,D_{Z_f})\otimes\newTate(-1) \end{equation} The image of the morphism $\pi^0_{\dagger}\nbigt_!(F,D_Y^{(0)}) \lrarr \pi^0_{\dagger}\nbigt_{\ast}(F,D_Y^{(0)})$ is $\iota_{\dagger}\nbigt(g_0)\otimes\newTate(-1)$ under the isomorphisms. \hfill\qed \end{cor} \subsection{Specialization} Let $X$ be a complex manifold with a simple normal crossing hypersurface $D$. We set $X^{(1)}:=X\times\cnum_{\tau}$ and $D^{(1)}:=D\times\cnum_{\tau}$. Let $f$ and $g$ be meromorphic functions on $(X,D)$. We have the meromorphic function $F=\tau f+g$ on $(X^{(1)},D^{(1)})$. We have the associated integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(F,D^{(1)})$ $(\star=\ast,!)$ with real structure on $X^{(1)}$. Let $\nbigk_{\star,f,g}$ and $\nbigc_{\star,f,g}$ denote the kernel and the cokernel of $\nbigt_{\star}(F,D^{(1)})[!\tau] \lrarr \nbigt_{\star}(F,D^{(1)})[\ast\tau]$. Let $\iota_0:X\lrarr X^{(1)}$ be given by $\iota_0(Q)=(Q,0)$. \begin{prop} If $|(f)_0|\cap|(f)_{\infty}|=\emptyset$, we have the following isomorphism of the integrable mixed twistor $\nbigd$-modules with real structure: \begin{equation} \nbigc_{\ast,f,g}\simeq \iota_{0\dagger} \nbigt_{\ast}(g,D) \otimes\newTate(-1), \quad\quad \nbigk_{\ast,f,g}\simeq \iota_{0\dagger} \nbigt_{\ast}(g,D)[!(f)_{\infty}], \end{equation} \begin{equation} \nbigk_{!,f,g}\simeq \iota_{0\dagger} \nbigt_!(g,D), \quad\quad \nbigc_{!,f,g}\simeq \iota_{0\dagger} \nbigt_!(g,D)[\ast(f)_{\infty}] \otimes\newTate(-1). \end{equation} \end{prop} \pf By Proposition , if $D=\emptyset$, we already have the isomorphisms of integrable mixed twistor $D$-modules with real structure in (, ). In other words, we already have the isomorphisms of integrable mixed twistor $D$-modules with real structure on $X^{(1)}\setminus D^{(1)}$. It is enough to prove that they are extended isomorphisms of integrable mixed twistor $D$-modules with real structure on $X^{(1)}$. By construction of the morphisms in Proposition , we have the following isomorphisms of integrable filtered $\nbigr$-triples: \[ \Bigl( \nbigc_{\ast,f,g} \Bigr)(\ast D^{(1)}) \stackrel{a_1}{\simeq} \Bigl( \iota_{0\dagger} \nbigt_{\ast}(g,D) \otimes\newTate(-1) \Bigr)(\ast D^{(1)}), \quad\quad \Bigl( \nbigk_{\ast,f,g} \Bigr)(\ast D^{(1)}) \stackrel{a_2} \simeq \Bigl( \iota_{0\dagger} \nbigt_{\ast}(g,D)[!(f)_{\infty}] \Bigr)(\ast D^{(1)}), \] We have the following morphism of integrable mixed twistor $D$-modules induced by $a_1$: \begin{equation} \nbigc_{\ast,f,g} \stackrel{b_1}\lrarr \iota_{0\dagger} \nbigt_{\ast}(g,D) \otimes\newTate(-1) \end{equation} By Proposition , the morphism of the underlying $D$-modules is an isomorphism. Hence, () is an isomorphism. We have the following morphisms: \begin{equation} \nbigk_{\ast,f,g} \stackrel{c_1}{\llarr} \nbigk_{\ast,f,g}[!(f)_{\infty}] \stackrel{c_2}{\lrarr} \nbigk_{\ast,f,g}[\ast D^{(1)}][!(f_{\infty})] \stackrel{c_3}{\lrarr} \iota_{0\dagger}\nbigt_{\ast}(g,D)[!D^{(1)}][!(f)_{\infty}] \end{equation} Here, $c_i$ $(i=1,2)$ are canonical morphisms, and $c_3$ is induced by $a_2$. By Proposition , the morphisms of the underlying $D$-modules are isomorphisms. Hence, we obtain the isomorphism $\nbigk_{\ast,f,g} \simeq \iota_{0\dagger}\nbigt_{\ast}(g,D)[!(f)_{\infty}]$ from (). We obtain the isomorphisms in () similarly. \hfill\qed \subsection{$\cnum^{\ast}$-homogeneous $\nbigrtilde$-modules} \subsubsection{Homogeneity} Suppose that a complex manifold $X$ is equipped with a $\cnum^{\ast}$-action, i.e., a morphism $\mu:\cnum^{\ast}\times X\lrarr X$ satisfying $\mu(a_1a_2,x)=\mu(a_1,\mu(a_2,x))$ and $\mu(1,x)=x$. We consider the action of $\cnum^{\ast}$ on $\cnum_{\lambda}$ given by the multiplication. Set $\nbigx:=\cnum_{\lambda}\times X$. We have the diagonal $\cnum^{\ast}$-action $\mutilde$ on $\nbigx$. Let $\nbigm$ be an $\nbigrtilde_X$-module. It is equivalent to an $\nbigo_{\cnum_{\lambda}\times X}$-module $\nbigm$ with a meromorphic flat connection $\nabla:\nbigm\lrarr \nbigm\otimes\Omega^1_{\nbigx}(\log \nbigx^0) \otimes\nbigo_{\nbigx}(\nbigx^0)$, where $\nbigx^0:=\{0\}\times X$. We have the $\nbigo_{\cnum^{\ast}\times\nbigx}$-module $\mutilde^{\ast}\nbigm$ on $\cnum^{\ast}\times\nbigx$ equipped with the meromorphic flat connection $\mutilde^{\ast}\nabla$. We can easily check that \[ (\mutilde^{\ast}\nabla)\bigl( \mutilde^{\ast}\nbigm \bigr) \subset \mutilde^{\ast}\nbigm\otimes \Omega^1_{\cnum^{\ast}\times\nbigx} \bigl( \log(\cnum^{\ast}\times\nbigx^0) \bigr) \otimes\nbigo(\cnum^{\ast}\times\nbigx^0). \] Hence, $\mutilde^{\ast}\nbigm$ is naturally an $\nbigrtilde_{\cnum^{\ast}\times X}$-module. Let $p:\cnum^{\ast}\times \nbigx \lrarr \nbigx$ denote the projection. Let $p_i: \cnum^{\ast}\times \cnum^{\ast}\times \nbigx \lrarr \nbigx$ $(i=1,2,3)$ be given by \[ p_1(a_1,a_2,x)=x, \quad p_2(a_1,a_2,x)=\mutilde(a_2,x), \quad p_3(a_1,a_2,x)=\mutilde(a_1a_2,x). \] We have the morphisms $p_{23},\id\times\mutilde,\mu_{\cnum^{\ast}}\times\id: \cnum^{\ast}\times \cnum^{\ast}\times \nbigx \lrarr \cnum^{\ast}\times \nbigx$ given by \[ p_{23}(a_1,a_2,x)=(a_2,x), \quad (\id\times\mutilde)(a_1,a_2,x)=(a_1,\mutilde(a_2,x)), \quad (\mu_{\cnum^{\ast}}\times\id)(a_1,a_2,x) =(a_1a_2,x). \] \begin{df} An $\nbigrtilde_X$-module $\nbigm$ is called $\cnum^{\ast}$-homogeneous if we have an isomorphism of $\nbigrtilde_{\cnum^{\ast}\times X}$-modules $\kappa: p^{\ast}\nbigm \simeq \mutilde^{\ast}\nbigm$ satisfying the cocycle condition $(\id\times\mutilde)^{\ast}\kappa\circ p_{23}^{\ast}\kappa =(\mu_{\cnum^{\ast}}\times\id)^{\ast}\kappa$. \hfill\qed \end{df} The restriction of $\mutilde$ to $\cnum_{\lambda}^{\ast}\times X\subset\nbigx$ is a free action, and the quotient space is $X$. The condition in Definition implies that there exists a $\nbigd$-module $M$ such that the restriction $\nbigm_{|\cnum^{\ast}\times X}$ is isomorphic to the pull back of $M$ by the quotient map $\cnum^{\ast}_{\lambda}\times X\lrarr (\cnum^{\ast}_{\lambda}\times X)/\cnum^{\ast}\simeq X$. Indeed, $M$ is given as the specialization of $\nbigm$ to $\{1\}\times X$, i.e., $M=\Xi_{DR}(\nbigm):= \iota_1^{-1}\bigl(\nbigm/(\lambda-1)\nbigm\bigr)$, where $\iota_1:X\lrarr\nbigx$ is given by $\iota_1(P)=(1,P)$. \begin{lem} Suppose that $\nbigm$ is a locally free $\nbigo_{\nbigx}$-module, for simplicity. Then, the torus action is uniquely determined by the connection $\nabla$. \end{lem} \pf Let $\underline{\gminiv}$ be the holomorphic fundamental vector field of the action $\mu$ on $X$, i.e., $\underline{\gminiv}_{|Q}=T_{(1,Q)}\mu_{\ast}(\del/\del a)$ for any $Q\in X$. The holomorphic fundamental vector field of $\mutilde$ on $\nbigx$ is given by $\lambda\del_{\lambda}+\underline{\gminiv}$. The $\cnum^{\ast}$-action on $\nbigm$ induces a differential operator $L:\nbigm\lrarr \nbigm$ such that $L(fs)=fL(s) +(\lambda\del_{\lambda}+\underline{\gminiv})f\cdot s$. On $\cnum_{\lambda}^{\ast}\times X$, we can easily check that $L(s)=\nabla_{\lambda\del_{\lambda}+\underline{\gminiv}}(s)$. The equality holds on $\cnum_{\lambda}\times X$. Because the $\cnum^{\ast}$-action is determined by $L$, it is uniquely determined by the connection $\nabla$. \hfill\qed \subsubsection{$\nbigr$-modules associated to homogeneous meromorphic functions} Let $X$ be a complex manifold with a $\cnum^{\ast}$-action $\mu$ as above. For any $a\in\cnum^{\ast}$, the morphism $\mu(a,\bullet):X\lrarr X$ is denoted by $\mu_a$. Let $D$ be a hypersurface of $X$ such that $\mu_a(D)=D$ for any $a\in\cnum^{\ast}$. Let $f$ be a meromorphic function on $(X,D)$ such that $\mu_a^{\ast}(f)=a f$ for any $a\in\cnum^{\ast}$. \begin{prop} The $\nbigr$-modules $\nbigl_{\star}(f,D)$ are $\cnum^{\ast}$-homogeneous. \end{prop} \pf First, let us consider the $\nbigrtilde_{X}(\ast D)$-module $\nbigq(f)$ induced by $\nbigo_{\nbigx}(\ast D)$ with the meromorphic flat connection $d+d(\lambda^{-1}f)$. We set $X_1:=\cnum^{\ast}\times X$ and $D_1:=\cnum^{\ast}\times D$. We set $f_1:=p^{\ast}(f)$. We naturally have $p^{\ast}\nbigo_{\nbigx}(\ast D) \simeq \nbigo_{\nbigx_1}(\ast D_1) \simeq \mutilde^{\ast}\nbigo_{\nbigx}(\ast D)$. Because $\mutilde^{\ast}(\lambda^{-1}f) =p^{\ast}(\lambda^{-1}f) =\lambda^{-1}f_1$, the pull back of the connections are also equal. Hence, we have $p^{\ast}\nbigq(f) \simeq \mutilde^{\ast}\nbigq(f)$. Similarly, we can check the cocycle condition for $\nbigq(f)$ as in Definition . It is enough to check that we have natural isomorphisms $p^{\ast}\nbigl_{\star}(f) \simeq \mutilde^{\ast}\nbigl_{\star}(f)$ which induce the above isomorphism $p^{\ast}\nbigq(f) \simeq \mutilde^{\ast}\nbigq(f)$. Take $(a_0,\lambda_0,Q_0)\in \nbigx_1=\cnum^{\ast}\times \cnum_{\lambda}\times X$. Let us check that there exists a unique isomorphism $p^{\ast}\nbigl_{\star}(f) \simeq \mutilde^{\ast}\nbigl_{\star}(f)$ around $(a_0,\lambda_0,Q_0)$ which induces $p^{\ast}\nbigq(f) \simeq \mutilde^{\ast}\nbigq(f)$. Let $g$ be a holomorphic function defined on a neighbourhood of $\mu(a_0,Q_0)$ such that $g^{-1}(0)=D$. We set $g_1:=g\circ\mu$ defined on a neighbourhood of $(a_0,Q_0)$. We have $p^{\ast}\nbigl_{\star}(f)[\star g_1] =p^{\ast}\nbigl_{\star}(f)$ on a neighbourhood of $(\lambda_0,a_0,Q_0)$. It is enough to prove that $\mutilde^{\ast}\nbigl_{\star}(f)[\star g_1] =\mutilde^{\ast}\nbigl_{\star}(f)$ on a neighbourhood of $(\lambda_0,a_0,Q_0)$. We set $Q_1:=\mu(a_0,Q_0)$. We take a small neighbourhood $U=U_{\lambda_0}\times U_{Q_1}\subset \nbigx$ of $(\lambda_0,Q_1)$ such that $g$ is defined on $U_{Q_1}$. Let $\iota_{g}:U_{Q_1}\lrarr U_{Q_1}\times\cnum_t$ be the graph. We may assume that we have the $V$-filtration of $\iota_{g\dagger}\nbigl_{\star}(f)$ on $U\times\cnum_t$ as $\nbigr$-modules. Note that each $V_a\bigl( \iota_{g\dagger}\nbigl_{\star}(f) \bigr)$ are naturally $V\nbigrtilde_{X\times\cnum_t}$-modules. In the case $\star=\ast$, the induced morphism $t:\Gr^V_{0}\lrarr \Gr^{V}_{-1}$ is an isomorphism. In the case $\star=!$, the induced morphism $\deldel_t:\Gr^V_{-1}\lrarr \Gr^{V}_{0}$ is an isomorphism. Let $U'\subset \nbigx_1$ be a small neighbourhood of $(a_0,\lambda_0,Q_0)$ such that $\mutilde(U')\subset U$. Let $\mutilde_1:U'\times\cnum_t\lrarr U\times\cnum_t$ be the induced morphism. We naturally have $\iota_{g\dagger}\nbigl_{\star}(f) =\bigoplus_{j=0}^{\infty} \deldel_t^j\iota_{g\ast}\nbigl_{\star}(f)$ and $\iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f) =\bigoplus_{j=0}^{\infty} \deldel_t^j\iota_{g_1\ast}\mutilde^{\ast}\nbigl_{\star}(f)$. The natural isomorphisms $\mutilde_1^{\ast} \iota_{g\ast}\nbigl_{\star}(f) \simeq \iota_{g_1\ast}\mutilde^{\ast}\nbigl_{\star}(f)$ induce isomorphisms $\mutilde_1^{\ast}(\iota_{g\dagger}\nbigl_{\star}(f)) \simeq \iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f)$. We set $V_a\bigl( \iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f) \bigr) := \mutilde_1^{\ast} V_a\bigl( \iota_{g\dagger}\nbigl_{\star}(f) \bigr)$ as $\nbigo_{U'\times\cnum_t}$-modules. By the construction, $V_a\bigl( \iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f) \bigr)$ are $V\nbigrtilde_{X_1\times\cnum_t}$-modules on $U'\times\cnum_t$. Because $V_a\bigl( \iota_{g\dagger}\nbigl_{\star}(f) \bigr)$ are coherent over $V\nbigr_{X\times\cnum_t}$, we obtain that $V_a\bigl( \iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f) \bigr)$ are pseudo-coherent over $\nbigo_{U'\times\cnum_t}$, and we can easily check that they are finitely generated over $V\nbigr_{X_1\times\cnum_t}$. Hence they are coherent over $V\nbigr_{X_1\times\cnum_t}$. (See .) Then, we obtain that $V_{\bullet}\bigl( \iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f) \bigr)$ is a $V$-filtration of $\iota_{g_1\dagger}\mutilde^{\ast}\nbigl_{\star}(f)$ along $t$. In the case $\star=\ast$, the morphism $t:\Gr^V_0\lrarr \Gr^V_{-1}$ is an isomorphism. In the case $\star=!$, the morphism $\deldel_t:\Gr^V_{-1}\lrarr \Gr^V_{0}$ is an isomorphism. Hence, we have $\mutilde^{\ast}\nbigl_{\star}(f) \simeq \nbigq(f_1)[\star g_1] \simeq p^{\ast}\nbigl_{\star}(f)[\star g_1]$. Thus, the proof of Proposition is finished. \hfill\qed \subsubsection{Good filtrations on holonomic $\nbigd$-modules and $\cnum^{\ast}$-homogeneity} Let $X$ be a complex manifold. We regard that $X$ is equipped with the trivial $\cnum^{\ast}$-action $\mu_0$, i.e., $\mu_0(a,x)=x$ for any $(a,x)\in \cnum^{\ast}\times X$. The induced action $\mutilde_0$ on $\cnum_{\lambda}\times X$ is just the multiplication on the $\cnum_{\lambda}$-component. In this case, we have a relation between $\cnum^{\ast}$-homogeneity and good filtrations. \vspace{.1in} Let $\nbigc_1(X)$ denote the category of holonomic $\nbigd_X$-modules with a good filtration $(M,F)$. Morphisms $(M_1,F_1)\lrarr (M_2,F_2)$ in $\nbigc_1(X)$ are morphisms $\varphi:M_1\lrarr M_2$ of holonomic $\nbigd_X$-modules such that $\varphi(F_jM_1)\subset F_jM_2$. Let $\nbigc_2(X)$ denote the category of coherent $\nbigrtilde$-modules $\nbigm$ such that (i) it is $\cnum^{\ast}$-homogeneous with respect to $\mu_0$, (ii) $\nbigm$ is coherent as an $\nbigr_X$-module, (iii) $\nbigm$ is strict as an $\nbigr_X$-module i.e., the multiplication of $\lambda-\lambda_0$ is a monomorphism for any $\lambda_0\in\cnum$, (iv) $\Xi_{\DR}(\nbigm)$ is a holonomic $\nbigd_{X}$-module. Morphisms $\nbigm_1\lrarr\nbigm_2$ in $\nbigc_2(X)$ are morphisms of $\nbigrtilde_X$-modules. Let $(M,F)$ be an object in $\nbigc_1(X)$. We have the Rees module $R(M,F)=\sum F_jM\lambda^{j}$ of $(M,F)$, which is naturally an $\nbigo_X[\lambda]$-module. By taking the analytification we obtain an $\nbigo_{\cnum_{\lambda}\times X}$-module $\Rtilde(M,F)$ which is naturally an $\nbigrtilde_X$-module. This construction gives a functor $\Rtilde:\nbigc_1(X)\lrarr\nbigc_2(X)$. \begin{prop} The functor $\Rtilde$ is an equivalence. \end{prop} \pf Set $\nbigx:=\cnum_{\lambda}\times X$. Let $p_{\lambda}:\nbigx\lrarr X$ denote the projection. Let $\Hol(X)$ denote the category of holonomic $\nbigd$-modules on $X$. Let $\nbigc_3(X)$ denote the category of $\nbigd_{\nbigx}(\ast\lambda)$-modules $\nbigl$ satisfying the conditions (i) $\nbigl$ is $\cnum^{\ast}$-homogeneous with respect to $\mu_0$, (ii) $\nbigl$ is coherent over $p_{\lambda}^{\ast}(\nbigd_X)(\ast\lambda)$, (iii) $\nbigl$ is strict as an $\nbigr$-module, (iv) $\Xi_{\DR}(\nbigl)$ is a holonomic $\nbigd_X$-module. Morphisms $\nbigl_1\lrarr\nbigl_2$ in $\nbigc_3(X)$ are $\cnum^{\ast}$-equivariant morphisms of $\nbigd_{\nbigx}$-modules. For any $M\in \Hol(X)$, we set $G_1(M):=p_{\lambda}^{\ast}(M)(\ast\lambda)$. Then, we obtain a functor $G_1:\Hol(X)\lrarr \nbigc_3(X)$. Clearly $G_1$ is an exact functor. \begin{lem} $G_1$ is an equivalence. \end{lem} \pf Clearly $G_1$ is faithful. Let us prove that $G_1$ is full. Let $M_i\in\Hol(X)$. Let $f:G_1(M_1)\lrarr G_1(M_2)$ be a morphism in $\nbigc_3(X)$. By applying $\Xi_{\DR}$, we obtain a morphism $f_1:=\Xi_{\DR}(f):M_1\lrarr M_2$. Because the restriction of $G_1(f_1)-f:G_1(M_1)\lrarr G_1(M_2)$ to $\cnum^{\ast}\times X$ is $0$ because of the $\cnum^{\ast}$-equivariance of $f$. Hence, we obtain $G_1(f_1)-f=0$ on $\cnum\times X$, i.e., $G_1$ is full. Let us prove that $G_1$ is essentially surjective. Let $\nbigl$ be an object in $\nbigc_3(X)$. We set $L:=\Xi_{\DR}(\nbigl)$. It is enough to prove that $G_1(L)\simeq \nbigl$. Note that we have a natural isomorphism $\kappa:G_1(L)_{|\cnum^{\ast}\times X}\simeq \nbigl_{|\cnum^{\ast}\times X}$ induced by the $\cnum^{\ast}$-homogeneity. It is enough to prove that $\kappa$ is extended to an isomorphism on $\cnum\times X$. Take any point $P\in X$. It is enough to prove that for a small neighbourhood $X_P$ of $P$ in $X$, $\kappa_{|\cnum^{\ast}\times X_P}$ is extended on $\cnum\times X_P$. We have the subvariety $Z(\nbigl)\subset X_P$ such that the support of the sheaf $\nbigl_{|\nbigx_P}$ on $\nbigx_P$ is $\cnum\times Z(\nbigl)$. We use an induction on the dimension of $Z(\nbigl)$. To simplify the description, we set $Z:=Z(\nbigl)$. We have a stratification $Z=\coprod Z_i$ into locally closed smooth subvarieties such that the characteristic variety of $L$ is the union of the conormal bundle $T_{Z_i}^{\ast}X$ of $Z_i$ in $X$. By shrinking $X_P$, we may assume to have holomorphic function $g$ such that $Z_i\subset g^{-1}(0)$ if and only if $\dim Z_i<\dim Z$. We may assume to have a complex manifold $Y$ with a projective morphism $\varphi:Y\lrarr X$ such that (i) $\varphi(Y)\subset Z$ and $\dim(Z\setminus \varphi(Y))<\dim Z$, (ii) we set $g_Y:=g\circ\varphi$, and then $D_Y:=g_Y^{-1}(0)$ is a normal crossing, (iii) $\varphi$ induces $Y\setminus g_Y^{-1}(0)\simeq Z\setminus g^{-1}(0)$. We set $\gtilde:=g\circ p_{\lambda}$ and $\gtilde_Y:=g_Y\circ p_{\lambda}$. We have a meromorphic flat bundle $L_1$ on $(Y,D_Y)$ with an isomorphism $\varphi_{+}L_1\simeq L(\ast g)$ of $\nbigd_{X_P}$-modules. We have a $\nbigd_{\nbigy}(\ast \gtilde_Y)$-module $\nbigl_1$ with an isomorphism $(\id\times\varphi)_{+}\nbigl_1\simeq \nbigl(\ast \gtilde)$. Note that $\nbigl_1$ is a meromorphic flat bundle on $\nbigy$ with the pole $(\{0\}\times Y)\cup \gtilde_Y^{-1}(0)$. Because $\nbigl(\ast \gtilde)$ is $\cnum^{\ast}$-homogeneous with respect to the trivial action on $X$, $\nbigl_1$ is $\cnum^{\ast}$-homogeneous with respect to the trivial action on $Y$. It is easy to observe that $\nbigl_{1|\cnum\times (Y\setminus D_Y)}$ is regular along $\{0\}\times(Y\setminus D_Y)$. Hence, the isomorphism $p_{\lambda}^{\ast}(L_1)_{|\cnum^{\ast}\times(Y\setminus D_Y)} \simeq \nbigl_{1|\cnum^{\ast}\times (Y\setminus D_Y)}$ is extended to an isomorphism $G_1(L_1)_{|\cnum\times(Y\setminus D_Y)} \simeq \nbigl_{1|\cnum\times (Y\setminus D_Y)}$. By using the Hartogs property, the isomorphism is extended to $G_1(L_1) \simeq \nbigl_1$. We also obtain $\nbigl_1[!\gtilde_Y] \simeq G_1(L_1)[!\gtilde_Y] \simeq G_1(L_1[!g_Y])$. Hence, we have a natural isomorphism $G_1\bigl(L(\star g)\bigr) \simeq \nbigl(\star \gtilde)$. Considering the Beilinson functors, we have $G_1\bigl( \Pi^{a,b}_g(L)(\star g) \bigr) \simeq \Pi^{a,b}_{\gtilde}(\nbigl)[\star \gtilde]$. We obtain $G_1\bigl(\psi_g^{(a)}(L) \bigr) \simeq \psi_{\gtilde}^{(a)}(\nbigl)$ and $G_1\bigl( \Xi_g^{(a)}(L) \bigr) \simeq \Xi_{\gtilde}^{(a)}(\nbigl)$. By the assumption of the induction, we have $G_1\bigl( \phi_g^{(a)}(L)\bigr) \simeq \phi_{\gtilde}^{(a)}(\nbigl)$. The following is commutative: \[ \begin{CD} G_1\bigl( \psi^{(1)}_g(L) \bigr) @>>> G_1\bigl( \phi^{(0)}_g(L) \bigr) \oplus G_1\bigl( \Xi^{(0)}_g(L) \bigr) @>>> G_1\bigl( \psi^{(0)}_g(L) \bigr) \\ @V{\simeq}VV @V{\simeq}VV @V{\simeq}VV \\ \psi^{(1)}_{\gtilde}(\nbigl) @>>> \phi^{(0)}_{\gtilde}(\nbigl) \oplus \Xi^{(0)}_{\gtilde}(\nbigl) @>>> \psi^{(0)}_{\gtilde}(\nbigl) \end{CD} \] We obtain $G_1(L)\simeq \nbigl$. Thus, Lemma is finished. \hfill\qed \vspace{.1in} Let $\nbigc_4(X)$ denote the category whose objects are the same as those of $\nbigc_3(X)$, but morphisms $f:\nbigm_1\lrarr\nbigm_2$ in $\nbigc_4(X)$ are morphisms of $\nbigd_{\nbigx}$-modules, i.e., we do not assume the $\cnum^{\ast}$-equivariance. \begin{lem} The natural functor $\nbigc_3(X)\lrarr \nbigc_4(X)$ is an equivalence. Namely, any morphism $f:\nbigm_1\lrarr\nbigm_2$ in $\nbigc_4(X)$ is $\cnum^{\ast}$-equivariant. \end{lem} \pf Let $M_1\lrarr M_2$ be a morphism in $\Hol(X)$. Let $f:G_1(M_1)\lrarr G_1(M_2)$ be a morphism in $\nbigc_4(X)$. Applying $\Xi_{\DR}$, we obtain a morphism $f_0:M_1\lrarr M_2$ of $\nbigd_X$-modules. It is easy to see that $\DR_{\cnum^{\ast}\times X} G_1(f_0)_{|\cnum^{\ast}\times X}= \DR_{\cnum^{\ast}\times X}f_{|\cnum^{\ast}\times X}$. Hence, we obtain $G_1(f_0)_{|\cnum^{\ast}\times X}= f_{|\cnum^{\ast}\times X}$, and $G_1(f_0)=f$ on $\nbigx$. \hfill\qed \vspace{.1in} Let us return to the proof of Proposition . Let $\nbigm$ be an object in $\nbigc_2(X)$. Set $M:=\Xi_{\DR}(\nbigm)$. For any local section $s$ of $M$, we have the section $p_{\lambda}^{\ast}(s)$ of $\nbigm(\ast\lambda)\simeq G_1(M)$ by Lemma . The number $i(s):=\min\big\{i\,\big|\, \lambda^{i}p_{\lambda}^{\ast}(s)\in\nbigm \}$ exists because $\nbigm$ is a coherent $\nbigr_X$-module. It determines the filtration $F$ on $M$, i.e., $F_j(M)=\{s\in M\,|\,i(s)\leq j\}$. We have $F_j\nbigd_X \cdot F_k(M)\subset F_{j+k}(M)$, where $F_j\nbigd_X$ denote the sheaf of differential operators whose orders are less than $j$. \begin{lem} We have a natural isomorphism $\nbigrtilde$-modules $\Rtilde(M,F)\simeq \nbigm$. \end{lem} \pf By Lemma , we naturally have $\Rtilde(M,F)(\ast\lambda) =G_1(M) \simeq \nbigm(\ast\lambda)$. By the construction of $F$, we have $\Rtilde(M,F) \subset \nbigm$. Let $\overline{p}_{\lambda}: \proj^1_{\lambda}\times X\lrarr X$ be the projection. Let $\nbigm'$ be the sheaf on $\proj^1_{\lambda}\times X$ such that (i) the restriction to $\bigl(\proj^1_{\lambda} \setminus\{0\}\bigr)\times X$ is equal to the restriction of $\overline{p}_{\lambda}^{\ast}(M)(\ast\infty)$, (ii) the restriction to $\cnum_{\lambda} \times X$ is equal to $\nbigm$. Because $\nbigm$ is $\cnum^{\ast}$-equivariant, $\nbigm'$ is also $\cnum^{\ast}$-equivariant. It is easy to observe that $R^i\overline{p}_{\lambda\ast} \nbigm'=0$ for $i>0$, and that the natural morphism $\overline{p}_{\lambda\ast}\nbigm' \lrarr \nbigm_{|\{0\}\times X}$ is an epimorphism. For any local section $s$ of $\nbigm/\lambda\nbigm$, we take a section $\stilde$ of $\overline{p}_{\lambda\ast}\nbigm'$ which is mapped to $s$. Because $\nbigm'\subset \overline{p}_{\lambda}^{\ast}(M) \bigl(\ast(\{0,\infty\}\times X)\bigr)$, we have $\overline{p}_{\lambda\ast}\nbigm' \subset M[\lambda,\lambda^{-1}]$. Hence, we have $\stilde=\sum_{j=-N}^N\lambda^j\stilde_j$ for a large $N$, where $\stilde_j$ are sections of $M$. By the $\cnum^{\ast}$-equivariance of $\nbigm$, we obtain that each $\lambda^j\stilde_j$ is a section of $\nbigm$. By construction, $\lambda^j\stilde_j$ are sections of $\Rtilde(M,F)$. It implies that $\Rtilde(M,F)\lrarr \nbigm/\lambda\nbigm$ is an epimorphism, and hence we obtain $\Rtilde(M,F)= \nbigm$. \hfill\qed \vspace{.1in} Let $\iota_0:X\lrarr \nbigx$ be given by $\iota_0(P)=(0,P)$. Because $\Gr^F(M)\simeq \iota_0^{-1}\bigl( \nbigm/\lambda\nbigm\bigr)$, the $\Sym^{\bullet}\Theta_X$-module $\Gr^F(M)$ is coherent. Hence, $F$ is a good filtration, i.e., $(M,F)$ is an object in $\nbigc_1(X)$. We obtain that $\Rtilde$ is essentially surjective. \vspace{.1in} Clearly $\Rtilde$ is faithful. Let us prove that $\Rtilde$ is full. Let $(M^{(i)},F^{(i)})$ $(i=1,2)$ be objects in $\nbigc_1(X)$. Let $f:\Rtilde(M^{(1)},F^{(1)})\lrarr \Rtilde(M^{(2)},F^{(2)})$ be a morphism in $\nbigc_2(X)$. By applying $\Xi_{\DR}$, we have the morphism $f_1:M^{(1)}\lrarr M^{(2)}$ of $\nbigd_{X}$-modules. By Lemma , $f$ is the restriction of $G_1(f_1)$ to $\Rtilde(M^{(1)},F^{(1)})$. Then, it is easy to deduce that $f_1$ preserves the filtrations, i.e., $f_1$ gives a morphism $(M^{(1)},F^{(1)})\lrarr (M^{(2)},F^{(2)})$ in $\nbigc_1(X)$, and $f=\Rtilde(f_1)$. Thus, Proposition is proved. \hfill\qed \vspace{.1in} Let us give some complements. Let $\nbigc_2'(X)$ denote the category whose objects are the same as those of $\nbigc_2(X)$, and whose morphisms $\nbigm_1\lrarr\nbigm_2$ are $\cnum^{\ast}$-equivariant $\nbigrtilde_X$-homomorphisms. \begin{cor} The natural functor $\nbigc_2'(X)\lrarr \nbigc_2(X)$ is an equivalence, i.e., any morphism in $\nbigc_2(X)$ is $\cnum^{\ast}$-equivariant. \end{cor} \pf It follows from Lemma . It also follows from the equivalence of $\Rtilde$ and the $\cnum^{\ast}$-equivariance of any morphism $\Rtilde(f):\Rtilde(M^{(1)},F^{(1)})\lrarr\Rtilde(M^{(2)},F^{(2)})$ induced by $f:(M^{(1)},F^{(1)})\lrarr(M^{(2)},F^{(2)})$ in $\nbigc_1(X)$. \hfill\qed \subsubsection{$\cnum^{\ast}$-homogeneity and specialization (Appendix)} Let $X$ be a complex manifold with a $\cnum^{\ast}$-action $\mu$. Let $\nbigm$ be a coherent and strict $\nbigrtilde_X$-module which is $\cnum^{\ast}$-homogeneous with respect to $\mu$. Let $f$ be a holomorphic function on $X$ such that there exists an integer $m$ such that $f(\mu(a,x))=a^mx$ for any $x\in X$. Suppose that $\nbigm$ is strictly specializable along $f$. Recall that the $V$-filtration of any $\nbigrtilde_X$-modules are indexed by $\real\times\{0\}$, which is naturally identified with $\real$. So, we have the $\nbigrtilde_X$-modules $\psi_{t,u}(\nbigm)$. They are not necessarily $\cnum^{\ast}$-homogeneous. Let us observe that we can obtain homogeneous $\nbigrtilde$-modules by modifying $\psi_{f,b}(\nbigm)$ $(b\in\real)$. By considering the graph construction of $f$, it is enough to consider the case where $X=X_0\times\cnum_t$, $f=t$ and the action $\mu$ is given as the diagonal action of the $\cnum^{\ast}$-action $\mu_0$ on $X_0$ and the $\cnum^{\ast}$-action on $\cnum_t$ given by $a\bullet t=a^mt$, and to study the $\nbigrtilde_{X_0}$-modules $\psi_{t,b}(\nbigm)$. Let $\nabla$ denote the meromorphic flat connection of $\nbigm$ induced by the $\nbigrtilde_X$-module structure, i.e., $\nabla: \nbigm\lrarr \nbigm\otimes \lambda^{-1}\Omega_{\nbigx}(\log \nbigx^0)$. Let $\mutilde:\cnum^{\ast}\times \nbigx\lrarr\nbigx$ be given by $\mutilde(a,\lambda,x,t)=(a\lambda,\mu_0(a,x),a^mt)$. Let $p_2:\cnum^{\ast}\times\nbigx\lrarr\nbigx$ be the projection. We have the isomorphism $p_2^{\ast}\nbigm\simeq \mutilde^{\ast}\nbigm$ satisfying the cocycle condition under which $p_2^{\ast}(\nabla)=\mutilde^{\ast}\nabla$. Let $V_{\bullet}$ denote the $V$-filtration of $\nbigm$ along $t$. Clearly $p^{\ast}(V_{\bullet}\nbigm)$ is the $V$-filtration of $p^{\ast}\nbigm$ along $t$. It is easy to check that $\mutilde^{\ast}(V_{\bullet}\nbigm)$ is also the $V$-filtration of $\mutilde^{\ast}\nbigm=p^{\ast}\nbigm$ along $t$. We have $p^{\ast}V_{\bullet}= \mutilde^{\ast}(V_{\bullet})$. Hence, the $\nbigo_{\nbigx_0}$-modules $\psi_{t,b}(\nbigm):=V_b\nbigm/V_{<b}\nbigm$ are $\mutilde_0^{\ast}$-equivariant, i.e., we have the isomorphism $\mutilde_0^{\ast}\psi_{t,b}(\nbigm) \simeq p_{0,2}^{\ast}\psi_{t,b}(\nbigm)$ satisfying the cocycle condition, where $p_{0,2}:\cnum^{\ast}\times\nbigx_0\lrarr\nbigx_0$ denotes the projection. Note that it is not necessarily $\cnum^{\ast}$-homogeneous as an $\nbigrtilde_{X_0}$-module. We have $\nabla: V_b\nbigm\lrarr V_b\nbigm\otimes \lambda^{-1} \Omega^1_{\nbigx}\bigl(\log(\nbigx^0\cup\nbigx_0)\bigr)$, which induces \[ G^{(b)}: \psi_{t,b}(\nbigm) \lrarr \psi_{t,b}(\nbigm) \otimes \Bigl( \lambda^{-1} \Omega^1_{\nbigx_0}(\log \nbigx_0^0) \oplus \lambda^{-1}\nbigo_{\nbigx_0}dt/t \Bigr) \] We have the decomposition $G^{(b)}=\nabla^{(b)}+\nbigb^{(b)}dt/t$, where $\nabla^{(b)}$ is a meromorphic flat connection of the $\nbigrtilde_{X_0}$-module $\psi_{t,b}(\nbigm)$, and $\nbigb^{(b)}$ is an $\nbigrtilde_{X_0}$-endomorphism of $\psi_{t,b}(\nbigm)$. By the relation $\mutilde^{\ast}\nabla=p_2^{\ast}\nabla$, we obtain \[ \mutilde_0^{\ast}(\nabla^{(b)}) +\mutilde^{\ast}\nbigb^{(b)}a^{-1}da =p_{0,2}^{\ast}(\nabla^{(b)}), \quad\quad \mutilde_0^{\ast}\nbigb^{(b)} =p_{0,2}^{\ast}(\nbigb^{(b)}) \] Then, the meromorphic flat connection $\nabla^{(b)}+m\nbigb^{(b)}d\lambda/\lambda$ gives an $\nbigrtilde_{X_0}$-module structure on $\psi_{t,b}(\nbigm)$ which is $\cnum^{\ast}$-homogeneous. \begin{rem} Note that $\nbigb^{(-1)}$ on $\psi_{t,-1}(\nbigm)$ is nilpotent, and the induced action on $\Cok(\nbigb^{(-1)})$ and $\Ker(\nbigb^{(-1)})$ are $0$. Hence, $\Cok(\nbigb^{(-1)})$ and $\Ker(\nbigb^{(-1)})$ are naturally $\cnum^{\ast}$-homogeneous. It also follows from the fact that $\nbigm[\star t]$ $(\star=\ast,!)$ are $\cnum^{\ast}$-homogeneous, and the kernel and the cokernel of $\nbigm[!t]\lrarr\nbigm[\ast t]$ are also $\cnum^{\ast}$-homogeneous. \hfill\qed \end{rem} \subsection{Hodge modules} \subsubsection{Hodge modules and $\cnum^{\ast}$-homogeneity} Let $\MHM(X,\real)$ denote the category of graded polarizable mixed $\real$-Hodge modules on $X$. Let $\MTMint_{\rnum\times\{0\}}(X,\real)$ denote the category of integrable mixed twistor $\nbigd$-modules with real structure on $X$ whose KMS-spectrum are contained in $\rnum\times\{0\}$ . As explained in , we have the fully faithful functor $\Psi_X:\MHM(X,\real)\lrarr \MTMint_{\rnum\times\{0\}}(X,\real)$. It is exact. It is compatible with the other operations such as the duality and the push-forward. Let us identify the essential image by the $\cnum^{\ast}$-homogeneity condition. Because we consider only graded polarizable mixed Hodge modules, we omit to distinguish ``graded polarizable''. We shall prove the following theorem in \S--. \begin{thm} Let $(\nbigt,W)$ be an object in $\MTMint_{\rnum\times\{0\}}(X,\real)$. Then, $(\nbigt,W)$ is contained in the essential image of $\Psi_X$ if and only if the underlying $\nbigrtilde_X$-modules $W_k\nbigt$ $(k\in\seisuu)$ are $\cnum^{\ast}$-homogeneous with respect to the trivial $\cnum^{\ast}$-action on $X$. \end{thm} \subsubsection{Smooth case} Let $\nbigt\in\MTMint_{\rnum\times\{0\}}(X,\real)$. Suppose that the underlying $\nbigrtilde_X$-modules are $\cnum^{\ast}$-homogeneous with respect to the trivial action. Let us consider the case where the underlying $\nbigrtilde_X$-modules of $\nbigt$ are smooth, i.e., locally free $\nbigo_{\cnum\times X}$-modules. We set $\nbigt_0=(\nbigv_1,\nbigv_2,C_{\nbigv}) :=\nbigt\otimes\nbigu_X(-d_X,0)$. It is naturally a variation of integrable twistor structure. It is also equipped with the real structure as a variation of integrable twistor structure, i.e., an isomorphism $\kappa: \gammatilde^{\ast}_{sm}(\nbigt_0) \simeq \nbigt_0$ such that $\gammatilde_{sm}^{\ast}\kappa\circ\kappa=\id$, where $\gammatilde_{sm}^{\ast}(\nbigt_0):= (j^{\ast}\nbigv_2^{\lor},j^{\ast}\nbigv_1^{\lor},C_{\nbigv}^{\lor})$. (Note that the notion of real structure for variations of integrable twistor structure is not equal to the notion of real structure for integrable mixed twistor $\nbigd$-modules.) Recall that we have the $C^{\infty}$-vector bundle on $\proj^1\times X$ with some differential operator $\DD^{\sankaku}$ associated to the smooth $\nbigr_X$-triple $\nbigt_0$. We have the $\nbigr_X$-module $\nbige:=\nbigv_1^{\lor}$ on $\cnum_{\lambda}\times X$ which is a locally free $\nbigo_{\cnum_{\lambda}\times X}$-module. Let $X^{\dagger}$ denote the conjugate complex manifold. Let $\sigma:\cnum_{\mu}\lrarr \cnum_{\lambda}$ be given by $\sigma(\mu)=-\overline{\mu}$. We have the anti-holomorphic isomorphism $\sigma: \cnum_{\mu}\times X^{\dagger} \lrarr \cnum_{\lambda}\times X$. We have the sheaf $\nbige^{\dagger}:=\sigma^{\ast}\nbigv_2$ which is naturally an $\nbigr_{X^{\dagger}}$-module, and a locally $\nbigo_{\cnum_{\mu}\times X^{\dagger}}$-module. Note that $\nbige$ is naturally a $C^{\infty}$-bundle with a family of flat connections \[ \DD^f:\nbige \lrarr \nbige \otimes \Bigl( \lambda^{-1}p_{\lambda}^{\ast}\Omega^{1,0}_X \oplus p_{\lambda}^{\ast}\Omega^{0,1}_X \Bigr) \] induced by the structure of $\nbigr_X$-module, where $p_{\lambda}:\cnum_{\lambda}\times X\lrarr X$ denote the projection. The family of flat connections is holomorphic with respect to $\lambda$. Note also that $\nbige^{\dagger}$ is naturally a $C^{\infty}$-bundle with a family of flat connections \[ \DD^{\dagger\,f}: \nbige^{\dagger}_{|\cnum_{\mu}^{\ast}\times X^{\dagger}}\lrarr \nbige^{\dagger}_{|\cnum_{\mu}^{\ast}\times X^{\dagger}} \otimes \Bigl( \mu^{-1}p_{\mu}^{\ast}\Omega^{1,0}_{X^{\dagger}} \oplus \Omega^{0,1}_{X^{\dagger}} \Bigr) \] induced by the structure of $\nbigr_{X^{\dagger}}$-module. The family of flat connections is holomorphic with respect to $\mu$. We identify $\cnum^{\ast}_{\lambda}$ and $\cnum^{\ast}_{\mu}$ by $\lambda=\mu^{-1}$. Then, we have the unique isomorphism of family of flat connections $\Phi: (\nbige,\DD^{f})_{|\cnum_{\lambda}^{\ast}\times X} \simeq (\nbige^{\dagger},\DD^{\dagger\,f}) _{|\cnum_{\mu}^{\ast}\times X^{\dagger}}$, such that (i) it is holomorphic with respect to $\lambda$, (ii) $\Phi_{|\vecS\times X}$ is compatible with the sesqui-linear pairing $C_{\nbigv}$. Let $\nbige^{\sankaku}$ be the $C^{\infty}$-bundle on $\proj^1\times X$ obtained as the gluing of $\nbige$ and $\nbige^{\dagger}$ by $\Phi$. It is equipped with the holomorphic structure in the $\proj^1$-direction $d''_{\lambda}$, and the differential operator $\DD^{\sankaku}: \nbige^{\sankaku} \lrarr \nbige^{\sankaku} \otimes \Omega^1_{\proj^1\times X/\proj^1} \otimes\nbigo_{\proj^1}(1)$ induced by the $\nbigr$-modules structure on $\nbige$ and $\nbige^{\dagger}$. We have the commutativity of $d''_{\lambda}$ and $\DD^{\sankaku}$. We set $W_{k}(\nbigt_0):= W_{k+d_X}(\nbigt)\otimes\nbigu_X(-d_X,0)$. By applying the above construction to $W_k(\nbigt_0)$, we obtain $W_k(\nbige^{\sankaku})$. We have natural inclusions $W_k(\nbige^{\sankaku})\subset \nbige^{\sankaku}$. In this way, we obtain an increasing filtration $W$ on $\nbige^{\sankaku}$ compatible with $d''_{\lambda}$ and $\DD^{\sankaku}$. By the construction, $(\nbige^{\sankaku},W)$ gives a variation of mixed twistor structure in the sense of . Because the sesqui-linear pairings of the underlying $\nbigr$-modules are compatible with the actions of $\lambda\del_{\lambda}$, it is preserved by the action of $S^1=\bigl\{t\in\cnum^{\ast}\,\big|\,|t|=1\bigr\}$ obtained as the restriction of the $\cnum^{\ast}$-action. We obtain that the gluing $\Phi$ is $\cnum^{\ast}$-equivariant. Hence, $(\nbige^{\sankaku},\DD^{\sankaku})$ is a $\cnum^{\ast}$-equivariant variation of mixed twistor structure. Let $\gamma:\proj^1\lrarr\proj^1$ be given by $\gamma(\lambda)=(\lambdabar)^{-1}$. The induced map $\proj^1\times X\lrarr \proj^1\times X$ is also denoted by $\gamma$. We have the naturally induced variation of mixed twistor structure $\gamma^{\ast}(\nbige^{\sankaku},\DD^{\sankaku},W)$ as in . The isomorphism $\kappa:\gammatilde_{sm}^{\ast}\nbigt_0\simeq\nbigt_0$ induces an isomorphism $\kappa_1:\gamma^{\ast}(\nbige^{\sankaku},\DD^{\sankaku},W) \simeq (\nbige^{\sankaku},\DD^{\sankaku},W)$ such that (i) $\gamma^{\ast}\kappa_1\circ\kappa_1=\id$, (ii) $\kappa_1$ is equivariant with respect to $\cnum^{\ast}$-action. Let $\sigma:\proj^1\lrarr\proj^1$ be given by $\sigma(\lambda)=-(\lambdabar)^{-1}$. The induced morphism $\proj^1\times X\lrarr \proj^1\times X$ is also denoted by $\sigma$. We have the naturally defined variation of mixed twistor structure $\sigma^{\ast}(\nbige^{\sankaku},\DD^{\sankaku},W)$ as in . We have $\sigma=j\circ\gamma$. By the restriction of the $\cnum^{\ast}$-action, we naturally have $(\nbige^{\sankaku},\DD^{\sankaku},W) \simeq j^{\ast}(\nbige^{\sankaku},\DD^{\sankaku},W)$. Hence, we have the isomorphism $\kappa_2:\sigma^{\ast}(\nbige^{\sankaku},\DD^{\sankaku},W) \simeq (\nbige^{\sankaku},\DD^{\sankaku},W)$ such that (i) $\sigma^{\ast}(\kappa_2)\circ\kappa_2=\id$, (ii) $\kappa_2$ is equivariant with respect to the $\cnum^{\ast}$-action. Then, by using \cite[Corollary 3.72]{Mochizuki-tame}, we obtain a variation of mixed $\real$-Hodge structure $(L_{\real},F,W)$ which induces $(\nbige^{\sankaku},\DD^{\sankaku},W)$ by the Rees construction, where $L_{\real}$ denotes the local system over $\real$, $F$ denotes the Hodge filtration of $H\otimes_{\real}\nbigo_X$, and $W$ denotes the weight filtration. Because each integrable variation of pure twistor structure $\Gr^W_k(\nbigt_0)$ is assumed to have an integrable polarization compatible with the real structure, we obtain that each $\Gr^W_k(\nbige^{\sankaku},\DD^{\sankaku})$ has a polarization which is equivariant with respect to the $\cnum^{\ast}$-action. Hence, by \cite[Corollary 3.72]{Mochizuki-tame}, the variation of mixed $\real$-Hodge structure $(L_{\real},F,W)$ is polarizable. Let $(P_{\real},M,F,W)$ be the mixed Hodge module associated to the graded polarizable variation of mixed Hodge structure $(L_{\real},F,W)$, i.e., $P_{\real}=L_{\real}[d_X]$, $M=L_{\real}\otimes_{\real}\nbigo_X$, $F$ is the Hodge filtration, $W_j(P)=W_{j+d_X}(L)[d_X]$. Then, by construction, we can observe that $\Psi_X(P_{\real},M,F,W)$ is naturally isomorphic to $(\nbigt,W)$. \subsubsection{Admissible variations of mixed twistor structure} Let $H$ be a normal crossing hypersurface of $X$. Let $(\nbigt_0,W)$ be an admissible variation of integrable mixed twistor structure with an integrable real structure and integrable graded polarization. Suppose that the underlying $\nbigr_X$-modules of $(\nbigt_0,W)$ are $\cnum^{\ast}$-homogeneous with respect to the trivial $\cnum^{\ast}$-action. By the result in \S, we already have a graded polarizable variation of mixed $\real$-Hodge structure $(L_{\real},F,W)$ on $X\setminus H$ corresponding to $(\nbigt_0,W)_{|X\setminus H}$. By the assumption on the KMS-spectrum, the monodromy automorphisms of $L_{\real}$ along the loop around the irreducible components of $H$ are quasi-unipotent. In particular, the variation of pure twistor structure corresponding to $\Gr^W_k(\nbigt_0)$ comes from a polarized variation of pure Hodge structure of weight $k$. Hence, the underlying harmonic bundle of $\Gr^W_k(\nbigt_0)$ is tame, and the $\nbigr$-modules of $\Gr^W_k\nbigt_0$ are regular singular along $H$. We obtain that the $\nbigrtilde$-modules of $\nbigt_0$ are regular singular. Let $(\nbigm_1,\nbigm_2,C)$ be the underlying $\nbigrtilde_X(\ast H)$-triple of $\nbigt_0$. It is easy to observe that the $\nbigrtilde_X$-modules $\nbigm_2[\ast H]$ are also $\cnum^{\ast}$-homogeneous. It gives an object in $\nbigc_2(X)$ in Proposition , and hence we have a good filtration $F$ on $M:=\Xi_{\DR}(\nbigm_2[\ast H])$ with an isomorphism $\Rtilde(M,F)\simeq \nbigm_2[\ast H]$. Let $F_j^{\circ}(M):= F_j(M)\otimes\nbigo_X(\ast H)$. Then, we have $\Rtilde(M,F^{\circ}) \simeq \nbigm_2$. In particular, $F^{\circ}_jM$ are locally free $\nbigo_X(\ast H)$-submodules of $M$ such that $F^{\circ}_j(M)/F^{\circ}_k(M)$ are also locally free $\nbigo_X(\ast H)$-modules. We set $W_k(M):=\Xi_{\DR}(\nbigm_2)$. We obtain the filtration $W$ on $M$, for which we naturally have $\Gr^W_k(M)\simeq \Xi_{\DR}(\Gr^W_k\nbigm_2)$. By applying the above construction to $W_k(\nbigt_0)$ and $\Gr^W_k(\nbigt_0)$, we obtain good filtrations $F^{\circ}$ on $W_k(M)$ and $\Gr^W_k(M)$. Because $\nbigm_2$ is an integrable $\nbigrtilde_X(\ast H)$-module which is regular singular along $H$, $W_kM$ are meromorphic flat bundles obtained as the extension of $W_kL_{\real}\otimes_{\real}\nbigo_{X\setminus H}$. Similarly, $\Gr^W_k(M)$ are regular singular meromorphic flat bundles obtained as the extension of $\Gr^W_k(L_{\real})\otimes_{\real}\nbigo_{X\setminus H}$. The restriction of $F^{\circ}$ on $W_k(M)$ and $\Gr^W_k(M)$ to $X\setminus H$ are the Hodge filtrations $F$ on $W_k(L_{\real})\otimes_{\real}\nbigo_{X\setminus H}$ and $\Gr^W_k(L_{\real})\otimes_{\real}\nbigo_{X\setminus H}$. Let us prove that $(L_{\real},F,W)$ is admissible along $H$. It is enough to study it on a small neighbourhood $X_P$ of any smooth point $P\in H$, according to . We may also assume that the monodromy of $L_{\real}$ is unipotent, after the pull back by an appropriate covering ramified along $H$. Let $z$ denote the defining equation of $H$ on $X_P$. Let $V_{\bullet}$ denote the $V$-filtration along $z$ for $W_k(\nbigm_2)$, $\Gr^W_k(\nbigm_2)$, $W_k(M)$ and $\Gr^W_k(M)$. Because the monodromy is unipotent, $V_{\bullet}$ is indexed by $\seisuu$, and $V_{-1}M$ is equal to the Deligne extension of $M_{|X\setminus H}$. We set $F_j^{\circ}V_{-1}:=F_j^{\circ}\cap V_{-1}$ on $W_k(M)$ and $\Gr^W_k(M)$. Note that $V_{-1}W_k(\nbigm_2) =\Rtilde(V_{-1}W_k(M),F^{\circ})$. Because $V_{-1}W_k(\nbigm_2)$ and $V_{-1}\Gr^W_k(\nbigm_2)$ are locally free $\nbigo_{\nbigx_P}$-modules by the admissibility of $\nbigt_0$, we obtain that $\Gr^{F^{\circ}}V_{-1}W_k(M)$ and $\Gr^{F^{\circ}}V_{-1}\Gr^W_k(M)$ are locally free $\nbigo_{X_P}$-modules. Because the morphisms $V_{-1}W_k\nbigm_2 \lrarr V_{-1}\Gr^W_{k}\nbigm_2$ are surjective for any $k\in\seisuu$ by the admissibility of $\nbigt_0$, the morphisms $F^{\circ}_jV_{-1}W_kM \lrarr F^{\circ}_jV_{-1}\Gr^W_{k}M$ are surjective for any $k,j\in\seisuu$, i.e., $F^{\circ}$ on $V_{-1}\Gr^W_k(M)$ is equal to the filtration induced by $V_{-1}W_k(M)$. By the admissibility of $\nbigt_0$, we have the relative weight filtration of the action of the nilpotent part of $z\del_z$ on $(V_{-1}(M),W)_{|P}$. Hence, $(L_{\real},F,W)$ is admissible. (See .) \vspace{.1in} We have objects $\nbigt_{\star}:=\Bigl( (\nbigt_0,W)\otimes\nbigu_X(d_X,0) \Bigr)[\star H]$ in $\MTMint_{\rnum\times\{0\}}(X,\real)$ for $\star=\ast,!$. We also have the mixed $\real$-Hodge modules $(P_{\real\star},M(\star H),F,W)$ $(\star=\ast,!)$ on $X$ such that (i) $W_jP_{\real\star|X\setminus H} =W_{j+d_X}L_{\real}[d_X]$, (ii) $F_{|X\setminus H}$ is the Hodge filtration of $(L_{\real},F,W)$, (iii) $W_k(P_{\real\star})_{|X\setminus H} =W_{k-d_X}L_{\real}[d_X]$. We have objects $\Psi_X(P_{\real\star},M(\star H),F,W)$ in $\MTMint_{\rnum\times\{0\}}(X,\real)$. By the construction, we have $\Psi_X(P_{\real\star},M(\star H),F,W)(\ast H) \simeq \nbigt_0\otimes\nbigu_X(d_X,0) \simeq \nbigt_{\star}(\ast H)$. We also have $\Psi_X(P_{\real\star},M(\star H),F,W)[\star H] \simeq \Psi_X(P_{\real\star},M(\star H),F,W)$. Hence, we have $\Psi_X(P_{\real\star},M(\star H),F,W) \simeq (\nbigt_{\star},W)$ in $\MTMint_{\rnum\times\{0\}}(X,\real)$, i.e., $(\nbigt_{\star},W)$ are contained in the essential image of $\Psi_X$. \subsubsection{End of Proof of Theorem } Let us consider the general case. Let $(\nbigt,W)$ be an object in $\MTMint_{\rnum\times\{0\}}(X,\real)$ such that the underlying $\nbigrtilde$-modules are $\cnum^{\ast}$-homogeneous with respect to the trivial action. We shall construct a mixed $\real$-Hodge module $(P,M,F,W)$ with an isomorphism $\Psi_X(P,M,F,W)\simeq (\nbigt,W)$. Because the functor $\Psi_X$ is fully faithful, it is enough to study the issue locally around any point $P$ of $X$. Note that an integrable polarization compatible with real structure on $\Gr^W(\nbigt)$ induces a polarization of $\Gr^W(P,M,F)$. We take a small neighbourhood $X_P$ of $X$. We use an induction on the dimension of the support $\Supp(\nbigt)$ of $\nbigt$. We set $Z:=\Supp(\nbigt)$. By shrinking $X_P$, we may assume to have a holomorphic function $g$ on $X$, a complex manifold $Y$ with a projective morphism $\varphi:Y\lrarr X$ and a normal crossing hypersurface $H_Y\subset Y$, and a variation of mixed twistor structure $\nbigt_1$ on $(Y,H_Y)$ such that the following holds: \begin{itemize} \item $\varphi(Y)\subset Z$, $\dim(Z\setminus\varphi(Y))<\dim Z$, and $Z\setminus \varphi(Y)\subset g^{-1}(0)$. \item $H_Y=(g\circ \varphi)^{-1}(0)$. \item $\varphi_{\dagger}(\nbigt_1,W) \simeq (\nbigt,W)(\ast g)$. \end{itemize} We have the $\cnum^{\ast}$-homogeneity of $(\nbigt_1,W)$ with respect to the trivial action. We set $g_Y:=g\circ\varphi$. Applying the results in \S, $ \Pi^{a,b}_{g_Y}(\nbigt_1,W)[\star H_Y]$ are contained in the essential image of $\Psi_Y$. Hence, we obtain that $\Pi^{a,b}_g(\nbigt,W)[\star g]$ are contained in the essential image of $\Psi_X$. It follows that $\psi^{(a)}_g(\nbigt,W)$ and $\Xi^{(a)}_g(\nbigt,W)$ are contained in the essential image of $\Psi_X$. By construction, the underlying $\nbigrtilde_X$-modules of $\phi_g^{(0)}(\nbigt,W)$ are also $\cnum^{\ast}$-homogeneous. By the assumption of the induction, $\phi_g^{(0)}(\nbigt,W)$ is also contained in the essential image of $\Psi_X$. Because $(\nbigt,W)$ is reconstructed as the cohomology of $\psi^{(1)}_g(\nbigt,W) \lrarr \phi^{(0)}_g(\nbigt,W)\oplus \Xi^{(0)}_g(\nbigt,W) \lrarr \psi^{(0)}_g(\nbigt,W)$, we obtain that $(\nbigt,W)$ is also contained in the essential image of $\Psi_X$. Thus, Theorem is proved. \hfill\qed \subsection{Sesqui-linear dualities and graded sesqui-linear dualities} Let $X$ be a complex manifold. Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. \begin{itemize} \item A sesqui-linear duality of weight $w$ on $(\nbigt,W)$ is a morphism $\nbigs:(\nbigt,W)\lrarr (\nbigt,W)^{\ast}\otimes\newTate(-w)$ such that $\nbigs^{\ast}=(-1)^w\nbigs$. \item A graded sesqui-linear duality on $(\nbigt,W)$ is a tuple of sesqui-linear dualities $\nbigs_w$ $(w\in\seisuu)$ of weight $w$ on $\Gr^W_w\nbigt$. \item A graded sesqui-linear duality $(\nbigs_w\,|w\in\seisuu)$ on $(\nbigt,W)$ is called a graded polarization if each $\nbigs_w$ is a polarization of $\Gr^W_w(\nbigt)$. \end{itemize} \begin{rem} The notions of sesqui-linear duality and graded sesqui-linear duality are the same if $(\nbigt,W)$ is pure. But, in general, they are not directly related. A sesqui-linear duality $\nbigs$ of weight $w$ on a mixed twistor $\nbigd$-module $(\nbigt,W)$ induces just morphisms $\Gr^W_{\ell}(\nbigt)\lrarr \Bigl( \Gr^W_{-\ell+2w}(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w)$. \hfill\qed \end{rem} If $(\nbigt,W)$ is pure of weight $w$, a sesqui-linear duality of weight $w$ is called just a sesqui-linear duality. \subsection{Induced graded sesqui-linear dualities} \subsubsection{Pure case} Let $X$ be a complex manifold. Let $\nbigt$ be a pure twistor $\nbigd$-module of weight $w$ on $X$. Let $D$ be an effective divisor of $X$. Recall that we have the mixed twistor $\nbigd$-modules $\nbigt[\star D]$ $(\star=\ast,!)$ obtained as the localizations. Note that a polarization $\nbigs$ of $\nbigt$ induces a graded polarization $\nbigs[\star D]$ of $\nbigt[\star D]$ as explained in . By the same construction, from a sesqui-linear duality $\nbigs$ of $\nbigt$, we obtain a graded sesqui-linear duality $\nbigs[\star D]=(\nbigs[\star D]_j\,|\,j\in\seisuu)$ on $\nbigt[\star D]$. Let us recall the local construction, for which we can take a holomorphic function $f$ such that $D=(f)_0$. (The graded sesqui-linear duality is eventually independent from the choice of the function $f$. See .) Let $\nbign:\psi_f^{(a)}(\nbigt)\lrarr \psi_f^{(a-1)}(\nbigt)$ be the canonical morphism. Recall that the weight filtration of $\psi_f^{(a)}(\nbigt)$ is the shift of the monodromy weight filtration of $\nbign$, i.e., $W(\nbign)_j\psi_f^{(a)}(\nbigt) =W_{w+1-2a+j}\psi_f^{(a)}(\nbigt)$. In particular, the induced morphisms \[ \Gr^W_{w+1-2a+j}\psi_f^{(a)}(\nbigt) \stackrel{\nbign^j}{\lrarr} \Gr^W_{w+1-2a+j}\psi_f^{(a-j)}(\nbigt) = \Gr^W_{w+1-2a-j}\bigl(\psi_f^{(a)}(\nbigt)\bigr) \otimes\newTate(-j) \] are isomorphisms for $j\geq 0$. The primitive part $P\Gr^W_{w+1-2a+j}\psi_f^{(a)}(\nbigt)$ $(j\geq 0)$ is defined to be the kernel of $\nbign^{j+1}: \Gr^W_{w+1-2a+j}\psi_f^{(a)}(\nbigt) \lrarr \Gr^W_{w-1-2a+j}\psi_f^{(a-j-1)}(\nbigt)$. We formally set $P\Gr^W_{w+1-2a+j}\psi_f^{(a)}(\nbigt)=0$ for $j<0$, in this paper. We have natural isomorphisms: \[ \Gr^W_{w+j}\nbigt[\ast f] \simeq \left\{ \begin{array}{ll} 0 & (j<0)\\ \nbigt & (j=0)\\ P\Gr^W_{w+j}\psi^{(0)}_f\nbigt & (j>0) \end{array} \right. \] A sesqui-linear duality $\nbigs[\ast (f)_0]_{w+1+\ell}$ on $P\Gr^W_{w+1+\ell}\psi_f^{(0)}\nbigt$ $(\ell\geq 0)$ is induced by the composite of the following morphisms: \begin{multline} \Gr^W_{w+1+\ell}\psi_f^{(0)}\nbigt \stackrel{a_1}{\lrarr} \Gr^W_{w+1+\ell}\psi_f^{(-\ell)}\nbigt =\Gr^W_{w+1-\ell}\bigl( \psi_f^{(0)}\nbigt\bigr) \otimes\newTate(-\ell) \stackrel{a_2}{\lrarr} \Gr^W_{-w+1-\ell}\bigl( \psi_f^{(0)}(\nbigt^{\ast})\bigr) \otimes\newTate(-w-\ell) \\ \stackrel{a_3}{\simeq} \Bigl( \Gr^W_{w-1+\ell}\psi_f^{(1)}(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-\ell) \stackrel{a_4}{\simeq} \Bigl( \Gr^W_{w+1+\ell} \bigl( \psi_f^{(0)}(\nbigt) \bigr) \otimes \newTate(1) \Bigr)^{\ast} \otimes\newTate(-w-\ell) \\ \stackrel{a_5}{\simeq} \Bigl( \Gr^W_{w+1+\ell}\psi_f^{(0)}(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-1-\ell) \end{multline} Here, $a_1$ is induced by $(-\nbign)^{\ell}$, $a_2$ is induced by $\nbigs$, and $a_i$ $(i=3,4,5)$ are the isomorphisms in . For $j<0$, let $P'\Gr^W_{w+1-2a+j}\psi^{(a)}_f\nbigt$ denote the image of $P\Gr^W_{w+1-2a+j}\psi^{(a-j)}_f\nbigt \lrarr \Gr^W_{w+1-2a+j}\psi^{(a)}_f\nbigt$. We have the following natural isomorphisms: \[ \Gr^W_{w+j}\nbigt[!f] \simeq \left\{ \begin{array}{ll} 0 & (j>0)\\ \nbigt & (j=0)\\ P'\Gr^W_{w+j}\psi^{(1)}_f(\nbigt) & (j<0) \end{array} \right. \] Because $P'\Gr^W_{w+1-2a+j}\psi^{(a)}_f\nbigt \simeq P\Gr^W_{w+1-2a+j}\psi^{(a-j)}_f\nbigt$, the pure twistor $\nbigd$-modules $P'\Gr^W_{w-1-\ell}\psi^{(1)}_f(\nbigt)$ $(\ell\geq 0)$ are equipped with the induced sesqui-linear dualities $\nbigs[!(f)_0]_{w-1-\ell}$. They are induced by the composite of the following: \begin{multline} \Gr^W_{w-1-\ell}\psi_f^{(1)}\nbigt \stackrel{b_1}{\lrarr} \Gr^W_{-w-1-\ell}\psi_f^{(1)}(\nbigt^{\ast})\otimes\newTate(-w) \\ \stackrel{b_2}{\simeq} \Gr^{W}_{-w+1-\ell}\psi_f^{(0)}(\nbigt^{\ast})\otimes\newTate(-w+1) =\Bigl( \Gr^{W}_{w-1+\ell}\psi_f^{(1)}(\nbigt) \Bigr)^{\ast}\otimes\newTate(-w+1) \\ \stackrel{b_3}{\lrarr} \Bigl( \Gr^{W}_{w-1-\ell}\psi_f^{(1)}(\nbigt)\otimes\newTate(-\ell) \Bigr)^{\ast}\otimes\newTate(-w+1) \stackrel{b_4}{\simeq} \Bigl( \Gr^{W}_{w-1-\ell}\psi_f^{(1)}(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w+1+\ell) \end{multline} Here, $b_1$ is induced by $\psi_f^{(1)}(\nbigs)$, $b_3$ is the inverse of the induced morphism of $(-1)^{\ell}\bigl(\nbign^{\ell}\bigr)^{\ast}$, and $b_i$ $(i=2,4)$ are the natural isomorphisms. It is also induced by the composite of the following: \begin{multline} \Gr^W_{w-1-\ell}\psi_f^{(1)}(\nbigt) \stackrel{c_1}{\lrarr} \Gr^{W}_{w-1+\ell}\psi_f^{(1)}(\nbigt) \otimes\newTate(\ell) \stackrel{c_2}{\lrarr} \Gr^{W}_{-w-1+\ell}\psi_f^{(1)}(\nbigt^{\ast})\otimes\newTate(-w+\ell) \\ = \Gr^{W}_{-w+1+\ell}\psi_f^{(0)}(\nbigt^{\ast})\otimes\newTate(-w+1+\ell) \stackrel{c_3}{\simeq} \Bigl( \Gr^{W}_{w-1-\ell}\psi_f^{(1)}(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w+1+\ell) \end{multline} Here, $c_1$ is the inverse of the induced morphism of $\nbign^{\ell}$, $c_2$ is induced by $\nbigs$, and $c_3$ is the natural morphism. The graded sesqui-linear dualities $\nbigs[!D]$ and $\nbigs[\ast D]$ induce graded sesqui-linear dualities on the kernel and the cokernel of the morphism $\varphi: \nbigt[!D] \lrarr \nbigt[\ast D]$, denoted as $\vecnbigs_{\Ker(\varphi)}= \bigl(\nbigs_{\Ker(\varphi),j}\,\big|\,j\in\seisuu\bigr)$ and $\vecnbigs_{\Cok(\varphi)}= \bigl(\nbigs_{\Cok(\varphi),j}\,\big|\,j\in\seisuu\bigr)$. Suppose that the sesqui-linear duality $\nbigs:\nbigt\lrarr\nbigt^{\ast}\otimes\newTate(-w)$ is an isomorphism. Note that we have the induced isomorphism $\Cok(\varphi)\simeq \Ker(\varphi)^{\ast}\otimes\newTate(-w)$, and hence we have the following isomorphisms for $j>0$: \begin{equation} \Gr^W_{w+1+j}\Cok(\varphi) \simeq \Bigl( \Gr^W_{w-1-j}\Ker(\varphi) \Bigr)^{\ast} \otimes\newTate(-w) \end{equation} The following can be checked by a direct computation. \begin{lem} Under the isomorphism {\rm()}, we have $\nbigs_{\Cok(\varphi),w+1+j} =\bigl( \nbigs_{\Ker(\varphi),w-1-j} \bigr)^{-1}$ for $j\geq 0$. \hfill\qed \end{lem} \begin{rem} If $\nbigs$ is a polarization, then $\nbigs[\star D]$ are graded polarizations. \hfill\qed \end{rem} \subsubsection{Mixed case} The construction was also generalized in the mixed case . Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. Let $D$ be an effective divisor of $X$. A graded sesqui-linear duality $\nbigvecs=(\nbigs_w\,|\,w\in\seisuu)$ of $(\nbigt,W)$ induces graded sesqui-linear dualities $\vecnbigs[\star D]$ of the mixed twistor $\nbigd$-modules $(\nbigt[\star D],W)$. We recall the local construction. If we are given a holomorphic function $f$ such that $D=(f)_0$, then $\psi_f^{(a)}(\nbigt)$ is equipped with the two filtrations. One is the filtration $L$ induced by the weight filtration of $\nbigt$. The other is the relative monodromy weight filtration $W$ of $\nbign$ with respect to $L$, which is equal to the weight filtration of the mixed twistor $\nbigd$-module $\psi_f^{(a)}(\nbigt)$. The induced filtration $L$ on $\Gr^W\psi_f^{(a)}(\nbigt)$ has a canonical splitting due to Kashiwara: \[ \Gr^W\psi_f^{(a)}(\nbigt) =\bigoplus_w\Gr^L_w\Gr^W\psi_f^{(a)}(\nbigt) =\bigoplus_w\Gr^W\Gr^L_w\psi_f^{(a)}(\nbigt) \] Hence, the sesqui-linear dualities $\nbigs_w$ on $\Gr^L_w(\nbigt)$ $(w\in\seisuu)$ induce sesqui-linear dualities of $\Gr^W_j\psi_f^{(a)}(\nbigt)$ $(j\in\seisuu)$. We have the decomposition: \[ \Gr^L_k(\nbigt[\ast D])=A_{1,k}\oplus A_{2,k} \] Here, $A_{1,k}$ is the sum of the direct summands of $\Gr^L_k(\nbigt)$ whose strict supports are not contained in $D$, and the support of $A_{2,k}$ is contained in $D$. As shown in , $A_{2,k}$ is naturally isomorphic to a subobject in $\Gr^W_k\psi^{(0)}_f(\nbigt)$. Hence, it is equipped with the induced sesqui-linear duality, which is the $k$-th entry of $\vecnbigs[\star D]$. If $\vecnbigs$ is a graded polarization, then $\vecnbigs[\star D]$ are graded polarizations. \subsection{Push-forward} \subsubsection{A condition} We introduce a condition on the push-forward of mixed twistor $\nbigd$-modules equipped with a graded sesqui-linear duality. Let $F:X\lrarr Y$ be a projective morphism of complex manifolds. Let $(\gbigt,W)$ be a mixed twistor $\nbigd$-module on $X$ with a graded sesqui-linear duality $\vecnbigs=(\nbigs_j\,|\,j\in\seisuu)$. Recall that we have the induced complex \[ \begin{CD} F^{i-1}_{\dagger}\Gr^W_{j+1}\gbigt @>{a^{i-1}_{j+1}}>> F^i_{\dagger}\Gr^W_{j}\gbigt @>{a^i_j}>> F^{i+1}_{\dagger}\Gr^W_{j-1}\gbigt, \end{CD} \] and $\Ker a^i_j/\Image a^{i-1}_{j+1}$ is naturally isomorphic to $\Gr^W_{j}F^i_{\dagger}\gbigt$. Here, $a^i_j$ are induced by the extensions $0\lrarr \Gr^W_{j-1}\lrarr W_{j}/W_{j-2} \lrarr \Gr^W_{j}\lrarr 0$. We set $\alpha_j:=a_{j+1}^{-1}$ and $\beta_j:=a_{j}^0$. We have \[ \begin{CD} F^{-1}_{\dagger}\Gr^W_{j+1}\gbigt @>{\alpha_j}>> F^0_{\dagger}\Gr^W_{j}\gbigt @>{\beta_j}>> F^{1}_{\dagger}\Gr^W_{j-1}\gbigt. \end{CD} \] As the Hermitian adjoint, we have the following: \[ \begin{CD} \Bigl( F^{-1}_{\dagger}\Gr^W_{j+1}\gbigt \Bigr)^{\ast} @<{\alpha_j^{\ast}}<< \Bigl( F^0_{\dagger}\Gr^W_{j}\gbigt \Bigr)^{\ast} @<{\beta_j^{\ast}}<< \Bigl( F^{1}_{\dagger}\Gr^W_{j-1}\gbigt \Bigr)^{\ast} \end{CD} \] We also have the induced isomorphism $F_{\dagger}^0\nbigs_{j}: F_{\dagger}^0\Gr^W_j\gbigt \stackrel{\simeq}{\lrarr} \bigl( F_{\dagger}^0\Gr^W_j\gbigt \bigr)^{\ast}\otimes\newTate(-j)$. \begin{lem} Let $\nbigi_j$ denote the image of $\Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) \lrarr \Ker\beta_j/\Image\alpha_j$. Then, the morphism \begin{equation} \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F^0_{\dagger}\nbigs_{j} \bigr) \lrarr \Bigl( \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F^0_{\dagger}\nbigs_{j} \bigr) \Bigr)^{\ast} \otimes\newTate(-j) \end{equation} induced by $F_{\dagger}^0\nbigs_{j}$ is factorized as follows: \begin{equation} \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) \lrarr \nbigi_j \stackrel{\nu_j}{\lrarr} \nbigi_j^{\ast}\otimes\newTate(-j) \lrarr \Bigl( \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) \Bigr)^{\ast} \otimes\newTate(-j) \end{equation} Namely, we have an induced sesqui-linear duality $\nu_j$ of weight $j$ on $\nbigi_j$. \end{lem} \pf Because $\Bigl( \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) \Bigr)^{\ast}$ is the quotient of $(F_{\dagger}^0\Gr^W_j\gbigt)^{\ast}$ by $\Image\beta_j^{\ast} + \Image\bigl( F_{\dagger}^0\nbigs_{j}\circ\alpha_j \bigr)$. Hence, the morphism () factors through $\nbigi_j$. Because $\nbigi_j^{\ast}$ is the image of $\Ker(\alpha_j^{\ast})$ to the quotient, the morphism () factors through $\nbigi^{\ast}\otimes\newTate(-j)$. The condition $\nu_j^{\ast}=(-1)^{j}\nu_{j}$ is clear by the construction. \hfill\qed \begin{df} We say that Condition {\bf (A)} is satisfied for the morphism $F:X\lrarr Y$ and the mixed twistor $\nbigd$-module $(\gbigt,W)$ with the graded sesqui-linear duality $\vecnbigs$ if the following holds: \begin{description} \item[(A)] The morphisms $\Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F^0_{\dagger}\nbigs_{j} \bigr) \lrarr \Ker\beta_j/\Image\alpha_j$ are epimorphisms for any $j$. \hfill\qed \end{description} \end{df} \begin{rem} If $F_{\dagger}^0\nbigs_j$ is a polarization of $F_{\dagger}^0\Gr^W_j\gbigt$, then $\Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) \lrarr \Ker\beta_j/\Image\alpha_j$ is an isomorphism. \hfill\qed \end{rem} According to Lemma , if the condition {\bf(A)} is satisfied, we have the induced sesqui-linear duality $[F^0_{\dagger}\nbigs_j]$ of weight $j$ on $\Gr^W_jF^0_{\dagger}\gbigt$ such that the following diagram is commutative: \[ \begin{CD} F_{\dagger}^0\Gr^W_j\gbigt @>{F_{\dagger}^0\nbigs_j}>> \bigl(F_{\dagger}^0 \Gr^W_j\gbigt\bigr)^{\ast}\otimes\newTate(-j)\\ @AAA @VVV\\ \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) @>>> \Bigl( \Ker\beta_j\cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{j} \bigr) \Bigr)^{\ast}\otimes\newTate(-j) \\ @VVV @AAA\\ \Gr^W_jF^0_{\dagger}\gbigt @>{[F^0_{\dagger}\nbigs_j]}>> \Bigl( \Gr^W_jF^0_{\dagger}\gbigt \Bigr)^{\ast}\otimes\newTate(-j) \end{CD} \] The tuple $\bigl( [F_{\dagger}^0\nbigs_j]\,\big|\, j\in\seisuu \bigr)$ is denoted by $[F_{\dagger}^0\vecnbigs]$. \subsubsection{Statements} Let $F:X\lrarr Y$ be a projective morphism of complex manifolds. Let $\nbigt$ be a pure twistor $\nbigd$-module of weight $w$ on $X$ with a sesqui-linear duality $\nbigs_X$. Let $D_Y$ be an effective divisor of $Y$. We set $D_X:=F^{\ast}D_Y$. As explained in \S, we have the mixed twistor $\nbigd$-modules $\nbigt[\star D_X]$ with the induced graded sesqui-linear duality $\nbigs_X[\star D_X]= \bigl(\nbigs_X[\star D_X]_{m}\,\big|\,m\in\seisuu\bigr)$. We shall prove the following theorem in \S. \begin{thm} Condition {\bf(A)} is satisfied for the projective morphism $F$ and the mixed twistor $\nbigd$-module $\nbigt[\star D_X]$ with the graded sesqui-linear duality $\nbigs_X[\star D_X]$. \end{thm} The pure twistor $\nbigd$-module $F_{\dagger}^0\nbigt$ of weight $w$ is equipped with the induced sesqui-linear duality $\nbigs_Y:=F_{\dagger}^0\nbigs_X$. Moreover, it induces a graded sesqui-linear duality $\nbigs_Y[\star D_Y]=(\nbigs_Y[\star D_Y]_m\,|\,m\in\seisuu)$ of $F_{\dagger}^0(\nbigt)[\star D_Y] \simeq F_{\dagger}^0(\nbigt[\star D_X])$. We shall prove the following theorem in \S. \begin{thm} We have $\bigl[F_{\dagger}^0\nbigs_{X}[\star D_X]\bigr] =\nbigs_{Y}[\star D_Y]$. \end{thm} \begin{cor} Suppose that $\nbigs$ is a polarization, and that $F_{\dagger}^i\nbigt[\star D_X]=0$ $(i\neq 0)$. Then, $[F_{\dagger}^0\nbigs_X[\star D_X]]$ are graded polarizations. \end{cor} \pf The assumptions imply that $\nbigs_Y$ is a polarization. Then, the claim follows from $[F_{\dagger}^0\nbigs_X[\star D_X]] =\nbigs_Y[\star D_Y]$. \hfill\qed \vspace{.1in} Let $\nbigk$ and $\nbigc$ denote the kernel and the cokernel of the natural morphism $\nbigt[!D_X]\lrarr\nbigt[\ast D_X]$. They are naturally equipped with the induced graded sesqui-linear dualities $\vecnbigs_{\nbigk}=(\nbigs_{\nbigk,j}\,|\,j\in\seisuu)$ and $\vecnbigs_{\nbigc}=(\nbigs_{\nbigc,j}\,|\,j\in\seisuu)$. \begin{cor} Suppose that $F_{\dagger}^i\nbigt[\star D_X] =F_{\dagger}^i\nbigk =F_{\dagger}^i\nbigc=0$ for $i\neq 0$. Then, Condition {\bf(A)} is satisfied for the morphism $F$ and the mixed twistor $\nbigd$-modules $\nbigk$ (resp. $\nbigc$) with $\vecnbigs_{\nbigk}$ (resp. $\vecnbigs_{\nbigc}$). Moreover, we have $[F_{\dagger}^0\nbigs_{\nbigk,j}]=\nbigs_{Y}[!D_Y]_j$ for $j<w$ and $[F_{\dagger}^0\nbigs_{\nbigc,j}]=\nbigs_{Y}[\ast D_Y]_j$ for $j>w$. Under the assumptions, if $\nbigs$ is a polarization, $[F_{\dagger}^0\vecnbigs_{\nbigk}]$ and $[F_{\dagger}^0\vecnbigs_{\nbigc}]$ are graded polarizations. \end{cor} \pf Under the assumption of the corollary, we also have $F_{\dagger}^i\nbigt=0$ $(i\neq 0)$. Then, the claims immediately follow from Theorem . \hfill\qed \vspace{.1in} For the proof of the theorems we give an argument in the case $\star=\ast$, and the other case is similar. So, to simplify the description, we denote $\nbigs_{X}[\ast D_X]_j$ and $\nbigs_{Y}[\ast D_Y]_j$ by $\nbigs_{X,j}$ and $\nbigs_{Y,j}$ in the following proof. Because it is enough to consider the issue locally on $Y$, we may assume to have a holomorphic function $g_Y$ such that $D_Y=(g_Y)_{0}$. The pull back $g_Y\circ F$ is denoted by $g$. We shall use the notation in \S with $\gbigt=\nbigt[\ast D_X]$. \subsubsection{Preliminary} We have a natural isomorphism $F_{\dagger}^i\psi_{g}^{(0)}\nbigt \simeq \psi_{g_Y}^{(0)}F_{\dagger}^i\nbigt$ of mixed twistor $\nbigd$-modules. By the spectral sequence for $\psi_{g}^{(0)}(\nbigt)$ with the weight filtration, we have the complex \begin{equation} F_{\dagger}^{-1}\Gr^W_{j+1}\psi_{g}^{(0)}(\nbigt) \stackrel{\xi_j}{\lrarr} F_{\dagger}^{0}\Gr^W_j\psi_{g}^{(0)}(\nbigt) \stackrel{\eta_j}{\lrarr} F_{\dagger}^{1}\Gr^W_{j-1}\psi_{g}^{(0)}(\nbigt), \end{equation} and $\ker \eta_j/\Image\xi_j$ is naturally isomorphic to $\Gr^W_jF_{\dagger}^0\psi_{g}^{(0)}(\nbigt) \simeq \Gr^W_j\psi_{g_Y}^{(0)}F_{\dagger}^0(\nbigt)$. For $\ell\geq 0$, let $P_{\ell}\Gr^W_{w+1+j}\psi^{(0)}_{g}(\nbigt)$ denote the image of $P\Gr^W_{w+1+j}\psi^{(\ell)}_{g}(\nbigt) \lrarr \Gr^W_{w+1+j}\psi^{(0)}_{g}(\nbigt)$. Note $P_{\ell}\Gr^W_{w+1+j}\psi_{g}^{(\ell)}(\nbigt)=0$ if $j+2\ell<0$. We have the primitive decomposition \[ \Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) =\bigoplus_{\ell\geq 0} P_{\ell}\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt). \] \begin{lem} For $j\geq 0$, the morphism $F_{\dagger}^0P\Gr^W_{w+1+j}\psi^{(0)}_g(\nbigt) \lrarr F_{\dagger}^1 \Gr^W_{w+j}\psi^{(0)}_g(\nbigt)$ factors through \[ F_{\dagger}^1 P_0\Gr^W_{w+j}\psi^{(0)}_g(\nbigt) \oplus F_{\dagger}^1 P_1\Gr^W_{w+j}\psi^{(0)}_g(\nbigt). \] \end{lem} \pf Note that $\xi_j$ and $\eta_j$ are compatible with the canonical morphisms \[ \nbign^{\ell}: F_{\dagger}^i\Gr^W_m\psi_g^{(0)}\nbigt \lrarr F_{\dagger}^i\Gr^W_m\psi_g^{(-\ell)}\nbigt =F_{\dagger}^i\Gr^W_{m-2\ell}\psi_g^{(0)}\nbigt \otimes\newTate(-\ell). \] For $j\geq 0$, the morphisms $\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) \lrarr \Gr^W_{w+1+j}\psi_g^{(-j)}(\nbigt) =\Gr^W_{w+1-j}\psi_g^{(0)}(\nbigt) \otimes\newTate(-j)$ are isomorphisms. Then, the claim easily follows. \hfill\qed \vspace{.1in} The restriction of $\eta_{w+1+j}$ to $F_{\dagger}^0P\Gr^W_{w+1+j}\psi^{(0)}_g(\nbigt)$ induces the following morphisms: \[ \eta_{k,w+1+j}: F_{\dagger}^0P\Gr^W_{w+1+j}\psi^{(0)}_g(\nbigt) \lrarr F_{\dagger}^1 P_k\Gr^W_{w+j}\psi^{(0)}_g(\nbigt), \quad (k=0,1). \] Note that $\Gr^W_j\Cok\bigl(\psi^{(1)}_g(\nbigt) \lrarr \psi^{(0)}_g(\nbigt)\bigr) \simeq P\Gr^W_j\psi^{(0)}_g\nbigt$. By using the spectral sequence for the cokernel, we obtain the following complex \begin{equation} \begin{CD} F_{\dagger}^{-1}P\Gr^W_{j+1}\psi_{g}^{(0)}(\nbigt) @>{\kappa_{1j}}>> F_{\dagger}^{0}P\Gr^W_j\psi_{g}^{(0)}(\nbigt) @>{\kappa_{2j}}>> F_{\dagger}^{1}P\Gr^W_{j-1}\psi_{g}^{(0)}(\nbigt). \end{CD} \end{equation} We have $\kappa_{2,w+1+j} =\eta_{0,w+1+j}$ by construction. \begin{lem} The following diagram is commutative up to signature: \begin{equation} \begin{CD} F_{\dagger}^0P\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) @>{\eta_{1,w+1+j}}>> F_{\dagger}^1 P_1\Gr^W_{w+j}\psi^{(0)}_g(\nbigt). \\ @V{\simeq}VV @V{\simeq}VV \\ \Bigl( F_{\dagger}^0P\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) \Bigr)^{\ast} \otimes \newTate(-w-j-1) @>{\kappa_{1,w+1+j}^{\ast}}>> \Bigl( F_{\dagger}^{-1} P\Gr^W_{w+j+2}\psi^{(0)}_g(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-j-1) \end{CD} \end{equation} The vertical arrows are induced by the induced sesqui-linear dualities of $P\Gr^W_{w+k}\psi_g^{(0)}(\nbigt)$ $(k> 0)$. \end{lem} \pf Note that we have the following diagram which is commutative up to signatures. \begin{equation} \begin{CD} F_{\dagger}^0 \Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) @>{\eta_{w+1+j}}>> F_{\dagger}^1 \Gr^W_{w+j}\psi_g^{(0)}(\nbigt) \\ @V{\nbign^j}V{\simeq}V @V{\nbign^j}VV \\ F_{\dagger}^0 \Gr^W_{w+1-j}\psi_g^{(0)}(\nbigt) \otimes\newTate(-j) @>{\eta_{w+1-j}}>> F_{\dagger}^1 \Gr^W_{w-j}\psi_g^{(0)}(\nbigt) \otimes\newTate(-j) \\ @V{F_{\dagger}^0\psi_g^{(0)}\nbigs_X}V{\simeq}V @V{F_{\dagger}^1\psi_g^{(0)}\nbigs_X}VV \\ F_{\dagger}^0 \Gr^W_{-w+1-j}\psi^{(0)}_g(\nbigt^{\ast}) \otimes\newTate(-w-j) @>{\eta_{-w+1-j}'}>> F_{\dagger}^1 \Gr^W_{-w-j}\psi^{(0)}_g(\nbigt^{\ast}) \otimes\newTate(-w-j) \\ @V{a_1}V{\simeq}V @V{a_2}VV \\ \Bigl( F_{\dagger}^0 \Gr^W_{w+1+j}\psi^{(0)}_g(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-j-1) @>{\xi_{w+1+j}^{\ast}}>> \Bigl( F_{\dagger}^{-1} \Gr^W_{w+j+2}\psi^{(0)}_g(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-j-1) \end{CD} \end{equation} Here, $a_i$ are morphisms induced by the natural isomorphisms $\psi^{(0)}_g(\nbigt^{\ast}) \simeq \psi^{(1)}_g(\nbigt)^{\ast} \simeq \psi^{(0)}_g(\nbigt)^{\ast} \otimes\newTate(-1)$, and $\eta'_{-w+1-j}$ denotes $\eta_{-w+1-j}$ for $\nbigt^{\ast}$. Let $\mu:F_{\dagger}^1 \Gr^W_{w+j}\psi_g^{(0)}(\nbigt) \lrarr \Bigl( F_{\dagger}^{-1} \Gr^W_{w+j+2}\psi^{(0)}_g(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-j-1)$ denote the composite of the right vertical arrows. The restriction of $\mu$ to $F_{\dagger}^1P_0\Gr^W_{w+j}\psi^{(0)}_g(\nbigt)$ is $0$, and the restriction of $\mu$ to $F_{\dagger}^1 P_1\Gr^W_{w+j}\psi^{(0)}_g(\nbigt)$ induces an isomorphism \[ \begin{CD} F_{\dagger}^1 P_1\Gr^W_{w+j}\psi^{(0)}_g(\nbigt) @>{\simeq}>> \Bigl( F_{\dagger}^{-1} P\Gr^W_{w+j+2}\psi^{(0)}_g(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-j-1) \end{CD} \] Thus, we obtain (). \hfill\qed \vspace{.1in} For $j\geq 0$, we have the natural morphism \begin{equation} \Ker\Bigl( \begin{CD} F_{\dagger}^0P\Gr^W_{w+1+j}\psi^{(0)}_g\nbigt @>{\eta_{w+1+j}}>> F_{\dagger}^1\Gr^W_{w+j}\psi^{(0)}_g\nbigt \end{CD} \Bigr) \lrarr P\Gr^W_{w+1+j} F_{\dagger}^0\psi^{(0)}_g\nbigt. \end{equation} \begin{lem} The morphism {\rm()} is an epimorphism. \end{lem} \pf Let $\nbigt=(\nbigm_1,\nbigm_2,C)$. Let $W$ denote the filtration of $\psi_g^{(0)}\nbigm_2$ associated to the weight filtration of $\psi_g^{(0)}\nbigt$. It is enough to prove that \begin{equation} \Ker\Bigl( F_{\dagger}^0P\Gr^W_{w+1+j}\psi^{(0)}_g\nbigm_2 \lrarr F_{\dagger}^1\Gr^W_{w+j}\psi^{(0)}_g\nbigm_2 \Bigr) \lrarr P\Gr^W_{w+1+j} F_{\dagger}^0\psi^{(0)}_g\nbigm_2 \end{equation} is an epimorphism. Let $f_1$ be a section of $\Ker\Bigl( F_{\dagger}^0\Gr^W_{w+1+j}\psi^{(0)}_g\nbigm_2 \lrarr F_{\dagger}^1\Gr^W_{w+j}\psi^{(0)}_g\nbigm_2 \Bigr)$ such that $\nbign^{j+1}f_1=\del f_2$ for $f_2\in F_{\dagger}^{-1}\Gr^W_{w-j}\psi^{(0)}_g\nbigm_2 \lambda^{j+1}$. We have $f_2'\in F_{\dagger}^{-1}\Gr^W_{w+1+j+1}\psi^{(0)}_g\nbigm_2$ such that $\nbign^{j+1}f_2'=f_2$. Then, $f_1-\del f_2'$ is a section of $\Ker\Bigl( F_{\dagger}^0P\Gr^W_{w+1+j}\psi^{(0)}_g\nbigm_2 \lrarr F_{\dagger}^1\Gr^W_{w+j}\psi^{(0)}_g\nbigm_2 \Bigr)$. Hence, () is an epimorphism. \hfill\qed \begin{lem} Under the identification $\Gr^W_{w+1}(\nbigt[\ast g]) \simeq P\Gr^W_{w+1}\psi_g^{(0)}(\nbigt)$, the kernel of \begin{equation} \begin{CD} F_{\dagger}^0P\Gr^W_{w+1}\psi^{(0)}_g\nbigt @>{\eta_{w+1}}>> F_{\dagger}^1\Gr^W_w\psi^{(0)}_g\nbigt \end{CD} \end{equation} is contained in $\Ker\beta_{w+1}$. (Recall that we use the notation in {\rm\S} with $\gbigt=\nbigt[\ast g]$.) \end{lem} \pf Recall that $\phi^{(0)}_g\nbigt$ is equal to the image of $\psi^{(1)}_g\nbigt \lrarr \psi^{(0)}_g\nbigt$. Let $\nbigc$ denote the cokernel. Let $\nbigc_1\subset\psi^{(0)}_g\nbigt$ denote the inverse image of $P\Gr^W_{w+1}\psi^{(0)}_g\nbigt\subset\nbigc$ by the projection $\psi_g^{(0)}\nbigt\lrarr\nbigc$. The extension $0\lrarr \phi^{(0)}_g\nbigt \lrarr \nbigc_1\lrarr P\Gr^W_{w+1}\psi_g^{(0)}\nbigt \lrarr 0$ induces the following morphism: \begin{equation} F_{\dagger}^0P\Gr^W_{w+1}\psi^{(1)}_g\nbigt \lrarr F_{\dagger}^1\phi^{(0)}_g(\nbigt). \end{equation} Because $\phi^{(0)}_g$ is an exact functor, the kernel of $\beta_{w+1}$ is equal to the kernel of (). Moreover, the kernel of () is equal to the kernel of the following induced morphism \begin{equation} F_{\dagger}^0P\Gr^W_{w+1}\psi^{(1)}_g\nbigt \stackrel{c_1}{\lrarr} \Gr^W_{w+1}F_{\dagger}^1\phi^{(0)}_g(\nbigt). \end{equation} We have the following complex associated to $\phi^{(0)}_g(\nbigt)$ with the weight filtration, and $\Gr^W_{w+1}F_{\dagger}^1\phi^{(0)}_g(\nbigt)$ is $\Ker c_3\big/\Image c_2$: \[ F_{\dagger}^0 \Gr^W_{w+1}\phi^{(0)}_g(\nbigt) \stackrel{c_2}{\lrarr} F_{\dagger}^1\Gr^{W}_w\phi^{(0)}_g(\nbigt) \stackrel{c_3}{\lrarr} F_{\dagger}^2\Gr^{W}_{w-1}\phi^{(0)}_g(\nbigt). \] We have $\Gr^W_{w}\psi^{(0)}_g(\nbigt) =\Gr^W_w\phi^{(0)}_g(\nbigt)$. The image of $\eta_{w+1}$ is contained in $\Ker c_3$, and $c_1$ is the composite of \[ \begin{CD} F^0_{\dagger}P\Gr^W_{w+1}\psi_g^{(1)}\nbigt @>{\eta_{w+1}}>> \Ker c_3 @>>> \Ker c_3\big/\Image c_2. \end{CD} \] Then, the claim of the lemma follows. \hfill\qed \subsubsection{Proof of Theorem and Theorem } For $j>w$, we have natural isomorphisms $\Gr^W_{j}F^i_{\dagger}\nbigt[\ast D_X] \simeq \Gr^W_{j}F^i_{\dagger}\psi_g\nbigt$. For $j>w$, the morphism $\kappa_{1j}$ is identified with $\alpha_j$. For $j>w+1$, the morphism $\kappa_{2j}$ in () is identified with $\beta_j$. We also have Lemma . By Lemma , we have \[ \Ker\Bigl( F_{\dagger}^0P\Gr^W_{j}\psi^{(0)}_g\nbigt \lrarr F_{\dagger}^1\Gr^W_{j-1}\psi^{(0)}_g\nbigt \Bigr) \simeq \Ker \beta_j \cap \Ker\bigl( \alpha_j^{\ast}\circ F_{\dagger}^0\nbigs_{X,j} \bigr). \] Hence, Lemma implies Theorem . \vspace{.1in} Let us consider the morphism \[ \begin{CD} F_{\dagger}^0\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) @>{F_{\dagger}^0\psi_g^{(0)}\nbigs_X\circ(-\nbign)^j}>> \Bigl( F_{\dagger}^0\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt) \Bigr)^{\ast} \otimes\newTate(-w-j-1) \end{CD} \] It induces a sesqui-linear duality for $\Ker \eta_{w+1+j} \subset F_{\dagger}^0\Gr^W_{w+1+j}\psi_g^{(0)}(\nbigt)$: \[ \Ker \eta_{w+1+j} \lrarr \Bigl( \Ker \eta_{w+1+j} \Bigr)^{\ast}\otimes\newTate(-w-j-1) \] It is factorized as follows: \begin{multline} \Ker\eta_{w+1+j} \lrarr \Gr^W_{w+1+j} F_{\dagger}^0\psi_g^{(0)}(\nbigt) \stackrel{b}{\lrarr} \\ \Bigl( \Gr^W_{w+1+j} F_{\dagger}^0\psi_g^{(0)}(\nbigt)^{\ast} \Bigr)^{\ast}\otimes\newTate(-w-j-1) \lrarr \Bigl( \Ker\eta_{w+1+j} \Bigr)^{\ast}\otimes\newTate(-w-j-1) \end{multline} By construction, the restriction of $b$ to $P\Gr^W_{w+1+j} F_{\dagger}^0\psi_g^{(0)}(\nbigt) \simeq \Gr^W_{w+1+j} F_{\dagger}^0\nbigt[\ast D_X] =\Gr^W_{w+1+j}F_{\dagger}^0\nbigt$ is the induced sesqui-linear duality $\nbigs_{Y,w+1+j}$. Then, the claim of Theorem follows. \hfill\qed \subsection{Basic examples of induced sesqui-linear dualities} Let $X$ be a complex manifold. Set $d:=\dim X$. We have the pure twistor $\nbigd$-module $\nbigu_X(d,0)=\bigl( \nbigo_{\nbigx}\lambda^d,\nbigo_{\nbigx},C_0 \bigr)$. Here, $C_0$ is given by $C_0(s_1,\sigma^{\ast}s_2)= s_1\cdot \overline{\sigma^{\ast}(s_2)}$. The canonical polarization $\nbigu_X(d,0)\lrarr \nbigu_X(d,0)^{\ast}\otimes\newTate(-d)$ is given by $\bigl( \lambda^d,1 \bigr) \longleftrightarrow \bigl( (-1)^d\cdot 1\cdot\lambda^{d},\, (-1)^d\cdot\lambda^d\cdot \lambda^{-d} \bigr)$. Let $D=\sum k_iD_i$ be an effective divisor of $X$ such that $\bigcup D_i$ is normal crossing. We describe the induced graded polarization on $\nbigu_X(d,0)[\ast D]$. \subsubsection{The simplest case} Let us consider the case that $D=(t)_0$ for a coordinate function $t$. We will not distinguish $D$ and $|(t)_0|$. We describe the induced graded polarizations of $\nbigu_X(d,0)[\star t]$ $(\star=\ast,!)$. It is enough to describe the induced polarizations on $\psi^{(0)}_t\nbigu_X(d,0)\simeq \nbigu_X(d,0)[\ast t]\big/\nbigu_X(d,0)$ and $\psi^{(1)}_t\nbigu_X(d,0) \simeq \Ker\bigl( \nbigu_X(d,0)[!t] \lrarr \nbigu_X(d,0) \bigr)$. \begin{lem} Let $\iota:D\lrarr X$ be the inclusion as above. The natural isomorphisms \[ \psi_t^{(0)}\nbigu_X(d,0) \simeq \iota_{\dagger} \nbigu_D(d-1,0)\otimes\newTate(-1), \quad \psi_t^{(1)}\nbigu_X(d,0) \simeq \iota_{\dagger} \nbigu_D(d-1,0) \] are compatible with the polarizations. \end{lem} \pf The natural isomorphism $\psi_{t,-\vecdelta}\nbigu_X(d,0)\otimes\nbigu_D(-1,0) \simeq \nbigu_D(d-1,0)$ is compatible with the polarizations. Then, the claim follows from \cite[Proposition 4.3.2]{Mochizuki-MTM}. \hfill\qed \vspace{.1in} We give a more explicit description of the polarizations. \begin{lem} The induced polarization $\psi_t^{(1)}\nbigu_X(d,0) \lrarr \bigl( \psi_t^{(1)} \nbigu_X(d,0) \bigr)^{\ast} \otimes\newTate(-d+1)$ is given by \begin{equation} \bigl( [t^{-1}\lambda^d], [\deldel_t] \bigr) \longleftrightarrow \bigl( (-1)^d[\deldel_t]\lambda^{d-1}, (-1)^d[t^{-1}\lambda] \bigr). \end{equation} The induced polarization $\psi_t^{(0)}\nbigu_X(d,0) \lrarr \bigl( \psi_t^{(0)} \nbigu_X(d,0) \bigr)^{\ast} \otimes\newTate(-d-1)$ is given by \begin{equation} \bigl( [\deldel_t] \lambda^{d},[t^{-1}] \bigr) \longleftrightarrow \bigl( (-1)^{d+1}[t^{-1}]\lambda^{d+1}, (-1)^{d+1}[\deldel_t]\lambda^{-1} \bigr). \end{equation} \end{lem} \pf Recall that we have a natural isomorphism of $\psi^{(1)}_t\nbigu_X(d,0) \simeq \iota_{\dagger} \psi_{-\vecdelta}\nbigu_X(d,0) \otimes\nbigu_X(-1,0)$ studied in \cite[Proposition 4.3.1]{Mochizuki-MTM}. In this case, the isomorphisms of the $\nbigr$-modules are given as follows: \[ \psi^{(0)}_t(\nbigo_{\nbigx}) \simeq \iota_{\dagger}\psi_{-\vecdelta}(\nbigo_{\nbigx}) \lambda^{-1}, \quad [t^{-1}] \longleftrightarrow (dt/\lambda)^{-1}\lambda^{-1} \] \[ \psi^{(1)}_t(\nbigo_{\nbigx}) \simeq \iota_{\dagger}\psi_{-\vecdelta}(\nbigo_{\nbigx}), \quad [\deldel_t] \longleftrightarrow -(dt/\lambda)^{-1} \] Then, according to \cite[Proposition 4.3.2]{Mochizuki-MTM}, the induced polarization of $\psi^{(1)}_t\nbigu_X(d,0)$ is given by (). Similarly, the isomorphism $\psi^{(0)}_t\nbigu_X(d,0) =\nbigu_X(d,0)[\ast t]/\nbigu_X(d,0) \simeq \iota_{\dagger} \psi_{-\vecdelta}\nbigu_X(d,0)\otimes \nbigu_X(0,-1)$ is given by \[ \bigl( -[\deldel_t]\lambda^{d}, [t^{-1}] \bigr) \longleftrightarrow \bigl( (dt/\lambda)^{-1}\lambda^d,\,\, (dt/\lambda)^{-1}\lambda^{-1} \bigr). \] Then, we obtain the claim for $\psi^{(0)}_t\nbigu_X(d,0)$. \hfill\qed \vspace{.1in} We have an isomorphism of mixed twistor $\nbigd$-modules \[ \nbigu_X(d,0)[\ast t]/\nbigu_X(d,0) \simeq \Ker\bigl( \nbigu_X(d,0)[!t] \lrarr \nbigu_X(d,0) \bigr)\otimes\newTate(-1) \] given by $\bigl( [\deldel_t]\lambda^d, [t^{-1}] \bigr) \longleftrightarrow \bigl( [t^{-1}]\lambda^{d+1}, [\deldel_t]\lambda^{-1} \bigr)$. It is compatible with the polarizations () and (). \subsubsection{Normal crossing case} Let us consider the case that $X$ is equipped with a holomorphic coordinate system $(x_1,\ldots,x_d)$ such that $D=\sum_{i=1}^m k_iD_i$ for some $(k_1,\ldots,k_m)\in\seisuu_{>0}^m$, where $D_i=\{x_i=0\}$. We have the decomposition \[ \Gr^W\bigl(\nbigu_X(d,0)[\ast D]\bigr) =\bigoplus_{j\geq d} \Gr^W_j\bigl( \nbigu_X(d,0)[\ast D] \bigr). \] For $J\subset\{1,\ldots,m\}$, we set $D_J:=\bigcap_{i\in J}D_i$. Let $\iota_J:D_J\lrarr X$ denote the inclusion. We have \begin{equation} \Gr^W_j\bigl( \nbigu_X(d,0)[\ast D] \bigr) \simeq \bigoplus_{|J|=j} \iota_{J\dagger} \nbigu_{D_J}(d-j,0) \otimes\newTate(-j). \end{equation} Note that $ \iota_{J\dagger} \nbigu_{D_J}(d-j,0) \otimes\newTate(-j)$ is equipped with the natural polarization $\bigl((-1)^d,(-1)^d\bigr)$. \begin{prop} Under the natural isomorphism {\rm()}, the induced polarization on $ \iota_{J\dagger} \nbigu_{D_J}(d-j,0) \otimes\newTate(-j)$ is equal to $ \Bigl( (-1)^{d}\prod_{i\in J}k_i^{-1}, (-1)^{d}\prod_{i\in J}k_i^{-1} \Bigr)$. \end{prop} \pf It is enough to consider the induced polarization on $\Gr^W_m\nbigu_X(d,0)$. For $\veca=(a_i)\in\seisuu^{m}$, we set $\phi^{(\veca)}:= \phi^{(a_1)}_{x_1}\circ\cdots \circ\phi^{(a_m)}_{x_m}$ and $\psi^{(\veca)}:= \psi^{(a_1)}_{x_1}\circ\cdots \circ\psi^{(a_m)}_{x_m}$. Let $\veczero=(0,\ldots,0)\in\seisuu^m$ and $\vecone=(1,\ldots,1)\in\seisuu^m$. We set $f:=\prod_{i=1}^m x_i^{k_i}$. We have $\psi_f^{(0)}\nbigu(d,0) =\bigl( \psi_f^{(1)}\nbigo_{\nbigx}\lambda^d, \psi_f^{(0)}\nbigo_{\nbigx}, \psi_f^{(0)}C \bigr)$. We consider $\phi^{(\veczero)} \psi^{(0)}_f\nbigu_X(d,0)$. Recall that $\psi_f^{(a)}(\nbigo_{\nbigx})$ is isomorphic to $\Cok\Bigl( \Pi^{a,\infty}_f\nbigo_{\nbigx}[!f] \lrarr \Pi^{a,\infty}_f\nbigo_{\nbigx}[\ast f] \Bigr)$. Then, $\phi^{(\veczero)} \psi_f^{(a)}(\nbigo_{\nbigx})$ is isomorphic to \[ \Cok\Bigl( \Pi^{a,\infty}_0\psi^{(\vecone)} \nbigo_{\nbigx} \stackrel{\rho}{\lrarr} \Pi^{a,\infty}_0\psi^{(\veczero)} \nbigo_{\nbigx} \Bigr). \] The morphism $\rho$ is induced by $(\lambda s)^j \prod_{i=1}^m\deldel_{x_i} \longmapsto (\lambda s)^{j+m} \prod_{i=1}^m k_ix_i^{-1}$. Hence, we have the following isomorphism: \begin{equation} \phi^{(\veczero)} \psi_f^{(0)} (\nbigo_{\nbigx}) \simeq \bigoplus_{0\leq j\leq m-1} \psi^{(\veczero)}(\nbigo_{\nbigx}) (\lambda s)^j \end{equation} Recall that $\psi_f^{(a)}(\nbigo_{\nbigx})$ is also isomorphic to $ \Ker\Bigl( \Pi^{-\infty,a}_f\nbigo_{\nbigx}[!f] \lrarr \Pi^{-\infty,a}_f\nbigo_{\nbigx}[\ast f] \Bigr)$, and that $\phi^{(\veczero)} \psi_f^{(a)}(\nbigo_{\nbigx})$ is isomorphic to \[ \Ker\Bigl( \Pi^{-\infty,a}_0\psi^{(\vecone)} \nbigo_{\nbigx} \lrarr \Pi^{-\infty,a}_0\psi^{(\veczero)} \nbigo_{\nbigx} \Bigr). \] We obtain the following isomorphism: \[ \phi^{(\veczero)} \psi_f^{(1)}(\nbigo_{\nbigx}) \simeq \bigoplus_{-m+1\leq j\leq 0} \psi^{(\vecone)}(\nbigo_{\nbigx}) (\lambda s)^j \] Hence, $\phi^{(\veczero)}\psi_f^{(0)}\nbigu_X(d,0)$ is naturally isomorphic to $\bigoplus_{j=0}^{m-1} \psi^{(\veczero)}\nbigu_X(d,0) \otimes \newTate(j)$, where $\psi^{(\veczero)}\nbigu_X(d,0)$ is \[ \bigl( \psi^{(\vecone)}\nbigo_{\nbigx}\lambda^d, \psi^{(\veczero)}\nbigo_{\nbigx}, \psi^{(\veczero)}C \bigr) \simeq \iota_{\mbar\dagger} \nbigu_{D_{\mbar}}(d-m,0) \otimes\newTate(-m). \] The canonical morphism $\nbign^j:\phi^{(\veczero)}\psi_f^{(a)}\nbigu_X(d,0) \lrarr \phi^{(\veczero)}\psi_f^{(a-j)}\nbigu_X(d,0)$ is given as follows; the second component is given as $b(\lambda s)^i\longmapsto b(\lambda s)^i$ if $i\leq a-j+m-1$, and $b(\lambda s)^i \longmapsto 0$ otherwise. The first component is given similarly. Let us describe the morphism $\phi^{(\veczero)}\psi_f^{(0)}\nbigu_X(d,0) \lrarr \phi^{(\veczero)}\psi_f^{(0)}\nbigu_X(d,0)^{\ast} \otimes\newTate(-d-1)$ induced by the polarization of $\nbigu_X(d,0)$. We have the isomorphism $\psi_f^{(0)}(\nbigo_{\nbigx}\lambda^d) \simeq \psi_f^{(1)}(\nbigo_{\nbigx}\lambda^{d})\lambda^{-1}$ induced by $(\lambda s)^j\longmapsto -(\lambda s)^{j+1}\lambda^{-1}$. It induces an identification $\phi^{(\veczero)} \psi_f^{(0)}(\nbigo_{\nbigx}\lambda^d) \simeq \phi^{(\veczero)} \psi_f^{(1)}(\nbigo_{\nbigx}\lambda^d) \lambda^{-1}$. Note that the isomorphism \[ \Cok\bigl( \Pi^{1,\infty}_f\nbigo_{\nbigx} \lrarr \Pi^{1,\infty}_f\nbigo_{\nbigx} \bigr) \\ \simeq \Ker\bigl( \Pi^{-\infty,1}_f\nbigo_{\nbigx} \lrarr \Pi^{-\infty,1}_f\nbigo_{\nbigx} \bigr) \] induces the isomorphism \begin{multline} \bigoplus_{1\leq j\leq m} \psi^{(\veczero)}(\nbigo_{\nbigx}) (\lambda s)^j \simeq \Cok\Bigl( \Pi^{1,\infty}_0\psi^{(\vecone)}\nbigo_{\nbigx} \lrarr \Pi^{1,\infty}_0\psi^{(\veczero)}\nbigo_{\nbigx} \Bigr) \\ \simeq \Ker\Bigl( \Pi^{-\infty,1}_0\psi^{(\vecone)}\nbigo_{\nbigx} \lrarr \Pi^{-\infty,1}_0\psi^{(\veczero)}\nbigo_{\nbigx} \Bigr) \simeq \bigoplus_{-m+1\leq j\leq 0} \psi^{(\vecone)}(\nbigo_{\nbigx}) (\lambda s)^j \end{multline} which is given by $(\lambda s)^{p+m}\prod_{i=1}^m(k_ix_i^{-1}) \longmapsto (\lambda s)^p\prod_{i=1}^m\deldel_i$. In all, the induced morphism \[ \phi^{(\veczero)}\psi^{(0)}(\nbigs): \bigoplus_{0\leq j\leq m-1} \psi^{(\veczero)}(\nbigo_{\nbigx}) (\lambda s)^j \lrarr \bigoplus_{-m+1\leq j\leq 0} \psi^{(\vecone)}(\nbigo_{\nbigx}) (\lambda s)^j \lambda^{-1} \] is given by $(\lambda s)^{p+m} \prod_{i=1}^m(k_ix_i^{-1}) \longmapsto (-1)^{d+1}\lambda^{-1}(\lambda s)^{p+1} \prod_{i=1}^m\deldel_i$. Hence, the second component of the induced polarization of $\Gr^W_{d+m}\phi^{(\veczero)}\psi^{(0)}_f\nbigu(d,0)$ is given by the isomorphisms $\prod_{i=1}^m (k_ix_i^{-1}) \longmapsto (-1)^{d+m}\lambda^{-m}\prod_{i=1}^m\deldel_i$. The first component is obtained in the same way. The isomorphism $\psi^{(\veczero)}\nbigo_{\nbigx} \simeq \iota_{\dagger}\nbigo_{\nbigd_{\mbar}} \lambda^{-m}$ is given by $\prod_{i=1}^m (x_i^{-1}\lambda) \longleftrightarrow \prod_{i=1}^m(dx_i/\lambda)^{-1}$, and the isomorphism $\psi^{(\vecone)}\nbigo_{\nbigx} \simeq \iota_{\dagger}\nbigo_{\nbigd_{\mbar}} \lambda^{-m}$ is given by $ \prod_{i=1}^m(-\deldel_i) \longleftrightarrow \prod_{i=1}^m(dx_i/\lambda)^{-1}$. Then, the claim of Proposition follows. \hfill\qed \subsection{Nearby cycle functors and maximal functors} We give a relation between the induced graded sesqui-linear dualities on the nearby cycle sheaves and the maximal sheaves. Although we do not use the results directly in this paper, the argument will be useful in \S. Let $\nbigt$ be a pure twistor $\nbigd$-module of weight $w$ on $X$. Let $f$ be a holomorphic function on $X$. We have the following exact sequences of mixed twistor $\nbigd$-modules: \[ 0\lrarr \psi_f^{(a+1)}\nbigt \lrarr \Xi^{(a)}_{f}\nbigt \lrarr \nbigt^{(a)}[\ast f] \lrarr 0 \] \[ 0\lrarr \nbigt^{(a)}[!f] \lrarr \Xi^{(a)}_f\nbigt \lrarr \psi_f^{(a)}\nbigt \lrarr 0 \] Hence, we have the following isomorphisms: \[ \Gr^W_{w-2a+j}\psi_f^{(a+1)}\nbigt \stackrel{\simeq}\lrarr \Gr^W_{w-2a+j}\Xi^{(a)}_f\nbigt \quad (j<0) \] \[ \Gr^W_{w-2a+j}\Xi^{(a)}_f\nbigt \stackrel{\simeq}{\lrarr} \Gr^W_{w-2a+j}\psi^{(a)}_f\nbigt \quad (j>0) \] The following exact sequences have the unique splittings: \[ 0\lrarr \Gr^W_{w-2a} \psi^{(a+1)}_f(\nbigt) \lrarr \Gr^W_{w-2a}\Xi^{(a)}_f\nbigt \lrarr \nbigt^{(a)} \lrarr 0 \] \[ 0\lrarr \nbigt^{(a)} \lrarr \Gr^W_{w-2a}\Xi^{(a)}_f\nbigt \lrarr \Gr^W_{w-2a}\psi_f^{(a)}\nbigt \lrarr 0 \] In other words, we have $\Gr^W_{w-2a}\Xi^{(a)}_f\nbigt =\nbigt^{(a)} \oplus \Gr^W_{w-2a}\psi_f^{(a)}\nbigt$, and $\Gr^W_{w-2a}\psi^{(a)}_f\nbigt =\Gr^W_{w-2a}\psi^{(a+1)}_f\nbigt$ in $\Gr^W_{w-2a}\Xi^{(a)}_f\nbigt$. The isomorphism $\Gr^W_{w-2a}\psi^{(a+1)}_f\nbigt \simeq \Gr^W_{w-2a}\psi^{(a)}_f\nbigt$ is induced by the canonical morphism $\nbign:\psi^{(a+1)}_f\nbigt \lrarr \psi^{(a)}_f\nbigt$. \vspace{.1in} Let $\nbign:\Xi^{(0)}_f\nbigt \lrarr \Xi^{(0)}_f\nbigt\otimes\newTate(-1)$ be the canonical morphism. Let $W(\nbign)$ denote the monodromy weight filtration of $\nbign$. For $j\geq 0$, let $P\Gr^{W(\nbign)}_{j}\Xi_f^{(0)}\nbigt$ denote the primitive part, i.e., the kernel of $\nbign^{j+1}: \Gr^{W(\nbign)}_{j}(\Xi_f^{(0)}\nbigt) \lrarr \Gr^{W(\nbign)}_{-j-2}(\Xi_f^{(0)}\nbigt) \otimes \newTate(-j-1)$. \begin{lem} We have $W(\nbign)_j=W_{w+j}$ $(j\in\seisuu)$, and \begin{equation} P\Gr^{W(\nbign)}_{j}\Xi_f^{(0)}\nbigt \simeq \left\{ \begin{array}{ll} \nbigt & (j=0)\\ P\Gr^W_{w+j}\psi_f^{(0)}(\nbigt) & (j>0) \end{array} \right. \end{equation} \end{lem} \pf It is enough to consider the case $w=0$. Because the morphism $\nbign:\Xi_f^{(0)}\nbigt\lrarr \Xi_f^{(0)}\nbigt\otimes \newTate(-1)$ induces $W_j\Xi_f^{(0)}\nbigt\lrarr \bigl( W_{j-2}\Xi_f^{(0)}\nbigt \bigr)\otimes\newTate(-1)$. Let us observe that the induced morphism $\Gr^W_{j}\Xi_f^{(0)}\nbigt \lrarr \Gr^W_{-j}\Xi_f^{(0)}\nbigt$ is an isomorphism for $j>0$. Note that $\nbign:\Xi^{(0)}_f\nbigt\lrarr \Xi^{(-1)}_f\nbigt$ factors through $\Xi^{(0)}_f\nbigt \lrarr \psi^{(0)}_f\nbigt \lrarr \Xi^{(-1)}_f\nbigt$. Hence, for $j>0$, we have the following commutative diagram: \[ \begin{CD} \Gr^{W}_j\Xi^{(0)}_f\nbigt @>{\nbign^j}>> \Gr^{W}_{j}\Xi^{(-j)}_f\nbigt @>{\nbign}>> \Gr^W_{j}\Xi^{(-j-1)}_f\nbigt \\ @V{\simeq}VV @A{\simeq}AA @A{\simeq}AA\\ \Gr^W_{j}\psi^{(0)}_f\nbigt @>{\nbign^{j-1}}>{\simeq}> \Gr^{W}_{j}\psi^{(-j+1)}_f\nbigt @>{\nbign}>>\Gr^W_{j-1}\psi^{(-j)}\nbigt \end{CD} \] Hence, we have $W=W(\nbign)$ on $\Xi^{(0)}\nbigt$. We also obtain () for $j>0$. We have the following commutative diagram: \[ \begin{CD} \Gr^W_2\Xi^{(0)}_f(\nbigt) @>>> \Gr^W_0\Xi^{(0)}_f(\nbigt)\otimes\newTate(-1) =\nbigt\oplus \Gr^W_0\psi^{(0)}_f(\nbigt) @>>> \Gr^W_{-2}\Xi^{(0)}_f(\nbigt)\otimes\newTate(-2) \\ @V{\simeq}VV @VVV @A{\simeq}AA \\ \Gr^W_2\psi^{(0)}_f(\nbigt) @>{\simeq}>> \Gr^W_0\psi^{(0)}_f(\nbigt) =\Gr^W_0\psi^{(1)}_f(\nbigt) @>{\simeq}>> \Gr^W_{-2}\psi^{(1)}_f(\nbigt) \end{CD} \] Then, we obtain () for $j=0$. \hfill\qed \vspace{.1in} Let $\nbigs$ be a Hermitian sesqui-linear duality of $\nbigt$. For $j\geq 0$, let $\nbigs^{\Xi^{(0)}}_j$ denote the composite of the following morphisms: \[ \Xi^{(0)}_f(\nbigt) \stackrel{(-\nbign)^{j}}{\lrarr} \Xi^{(-j)}_f(\nbigt) =\Xi^{(0)}_f(\nbigt)\otimes\newTate(-j) \lrarr \Xi^{(0)}_f(\nbigt^{\ast})\otimes\newTate(-j-w) \\ \simeq \Xi^{(0)}_f(\nbigt)^{\ast}\otimes\newTate(-j-w) \] The middle morphism is induced by $\nbigs$. Let $\nbigs^{\psi^{(0)}}_j$ denote the composite of the following morphisms: \begin{multline} \psi_f^{(0)}(\nbigt) \stackrel{(-\nbign)^j}{\lrarr} \psi_f^{(-j)}(\nbigt) =\psi_f^{(0)}(\nbigt)\otimes\newTate(-j) \lrarr \psi_f^{(0)}(\nbigt^{\ast})\otimes\newTate(-w-j) \simeq \psi_f^{(1)}(\nbigt)^{\ast}\otimes\newTate(-w-j) \\ \simeq \Bigl( \psi_f^{(0)}(\nbigt)\otimes\newTate(1) \Bigr)^{\ast}\otimes\newTate(-w-j) \simeq \psi_f^{(0)}(\nbigt)^{\ast}\otimes\newTate(-w-j-1) \end{multline} \begin{prop} For $j\geq 0$, the following diagram is commutative: \[ \begin{CD} \Xi_f^{(0)}(\nbigt) @>{\nbigs^{\Xi_f^{(0)}}_{j+1}}>> \Xi_f^{(0)}(\nbigt)^{\ast}\otimes\newTate(-w-j-1) \\ @VVV @AAA \\ \psi_f^{(0)}(\nbigt) @>{\nbigs^{\psi_f^{(0)}}_{j}}>> \psi_f^{(0)}(\nbigt)^{\ast} \otimes\newTate(-w-j-1) \end{CD} \] Here, the vertical morphisms are natural ones. \end{prop} \pf We can check the claim by a direct computation. We remark that, in (), the isomorphism $\newTate(1)^{\ast} \simeq \newTate(-1)$ is given by $(-1,-1)$. \hfill\qed \begin{cor} Suppose that $\nbigs$ is a polarization. For $j\geq 1$, the morphism $\nbigs^{\Xi_f^{(0)}}_{j}$ induces a polarization of $P\Gr^W_{w+j}\Xi^{(0)}_f(\nbigt)$. It is equal to the polarization of $P\Gr^W_{w+j}\psi^{(0)}_f(\nbigt)$ under the natural isomorphism. In particular, $\Xi^{(0)}_f\nbigt$ is equipped with an induced graded polarization. \hfill\qed \end{cor} \subsection{A specialization} \subsubsection{Statements} Let $X$ be a complex manifold with a normal crossing hypersurface $D$. Set $d:=\dim X$. Let $f,g\in\nbigo_X(\ast D)$. We consider a meromorphic function $F:=\tau f+g$ on $X\times\proj^1_{\tau}$. \begin{assumption} We assume the following. \begin{itemize} \item $|(f)_0|\cap |(f)_{\infty}|=\emptyset$, and $|(f)_0|\subset |(g)_{\infty}|$. We also have $D=|(f)_{\infty}|\cup|(g)_{\infty}|$. \item $F$ is pure on $X\times\{\tau\neq 0\}$, and $g$ is pure on $X\setminus|(f)_{\infty}|$. \hfill\qed \end{itemize} \end{assumption} For example, the assumptions are satisfied in the cases of Lemma and Lemma . We set $D^{(1)}:=(\proj^1\times D) \cup (X\times\{\infty\})$. We have the associated mixed twistor $\nbigd$-modules $\nbigt_{\star}(F,D^{(1)})$ $(\star=\ast,!)$ which are equipped with natural real structure. Recall that $\nbigt(F)$ denotes the image of $\nbigt_{!}(F,D^{(1)}) \lrarr \nbigt_{\ast}(F,D^{(1)})$, which is a pure twistor $\nbigd$-module of weight $d+1$. We have $\nbigt_{\star}(F,D^{(1)})[\star(\tau)_0] =\nbigt(F)[\star(\tau)_0] =\nbigt(F)[\star\tau]$ for $\star=\ast,!$ by the assumption of purity. Let $\iota:X\lrarr X\times\proj^1_{\tau}$ be given by $\iota(x)=(x,0)$. According to Proposition , for the morphism $\varphi: \nbigt(F)[!\tau]\lrarr \nbigt(F)[\ast\tau]$, we have the following isomorphisms: \[ \Cok(\varphi)\simeq \iota_{\dagger}\nbigt_{\ast}(g,D) \otimes\newTate(-1), \quad \Ker(\varphi)\simeq \iota_{\dagger}\nbigt_!(g,D) \] The polarization $\nbigs_F=\bigl((-1)^{d+1},(-1)^{d+1}\bigr)$ of $\nbigt(F)$ induces graded polarizations of $\nbigt(F)[\star \tau]$, and they induce graded polarizations of $\Cok(\varphi)$ and $\Ker(\varphi)$. The induced graded polarizations are denoted by $\vecnbigs_{\Cok(\varphi)} =(\nbigs_{\Cok(\varphi),w}\,|\,w\in\seisuu)$ and $\vecnbigs_{\Ker(\varphi)} =(\nbigs_{\Ker(\varphi),w}\,|\,w\in\seisuu)$. Note that by applying $\phi_{\tau}^{(0)}$ we have the following exact sequence: \[ 0\lrarr \Ker(\varphi)\lrarr \psi^{(1)}_{\tau}\nbigt(F) \lrarr \psi^{(0)}_{\tau}\nbigt(F) \lrarr\Cok(\varphi) \lrarr 0 \] For the weight filtrations of $\Cok(\varphi)$ and $\Ker(\varphi)$, we have \[ \Gr^W_{d+1+j} \Cok(\varphi)\simeq P\Gr^W_{d+1+j} \psi_{\tau}^{(0)}(\nbigt(F))\,\,\, (j>0), \quad\quad \Gr^W_{d+1+j}\Cok(\varphi)=0\,\,\, (j\leq 0), \] \[ \Gr^W_{d+1-j}\Ker(\varphi) \simeq P'\Gr^W_{d+1-j}\psi_{\tau}^{(1)}(\nbigt(F))\,\,\, (j>0), \quad\quad \Gr^W_{d+1-j}\Ker(\varphi)=0\,\,\, (j\leq 0). \] Here, $P'\Gr^W_{d+1-j}\psi_{\tau}^{(1)}(\nbigt(F))$ is the image of $P\Gr^W_{d+1-j}\psi_{\tau}^{(1-j)}(\nbigt(F))\lrarr \Gr^W_{d+1-j}\psi_{\tau}^{(1)}(\nbigt(F))$. The isomorphisms are compatible with the induced polarizations. We also have other polarizations. Let $\nbigt(g)$ denote the image of $\nbigt_!(g,D)\lrarr\nbigt_{\ast}(g,D)$, which is a pure twistor $\nbigd$-module of weight $d$. It is equipped with the natural polarization $\nbigs_g=((-1)^{d},(-1)^{d})$. Because $\nbigt(g)[\star (f)_{\infty}] =\nbigt_{\star}(g,D)$, we have the induced graded polarizations $\iota_{\dagger}\nbigs_{g}[\ast (f)_{\infty}]$ of $\iota_{\dagger}\nbigt_{\ast}(g,D)\otimes\newTate(-1)$, and $\iota_{\dagger}\nbigs_{g}[!(f)_{\infty}]$ of $\iota_{\dagger}\nbigt_{!}(g,D)$. We shall prove the following proposition in \S--. \begin{prop} Under the isomorphisms $\Cok(\varphi) \simeq \iota_{\dagger}\nbigt_{\ast}(g,D)\otimes\newTate(-1)$ and $\Ker(\varphi)\simeq \iota_{\dagger}\nbigt_!(g,D)$, we have $\vecnbigs_{\Cok(\varphi)}= \iota_{\dagger}\nbigs_{g}[\ast (f)_{\infty}]$ and $\vecnbigs_{\Ker(\varphi)}= \iota_{\dagger}\nbigs_{g}[!(f)_{\infty}]$. \end{prop} \subsubsection{Some consequences} Before going to the proof of Proposition , we give a consequence. Let $X$, $D$, $f$ and $g$ be as in \S. Let $\rho:X\lrarr Y$ be a projective morphism of complex manifolds. We assume that $R^i\rho_{\ast}\bigl(\Omega^j_X(\ast D)\bigr)=0$ for any $i>0$ and $j=0,\ldots,\dim X$. Set $\rho_1:=\rho\times\id_{\proj^1_{\tau}}: X\times\proj^1_{\tau} \lrarr Y\times\proj^1_{\tau}$. Note $\rho_{1\dagger}^i\nbigt(F)[\star(\tau)_0]=0$ $(i\neq 0)$ by Proposition {\rm}. Let $\iota_Y:Y\lrarr Y\times\proj^1_{\tau}$ be the inclusion induced by $\{0\}\lrarr \proj^1_{\tau}$. \begin{cor} Suppose that $\rho_{\dagger}^i\nbigt_{\star}(g,D)=0$ for $\star=\ast,!$ and for $i\neq 0$. \begin{itemize} \item Condition {\bf (A)} is satisfied for the morphism $\rho$ and the mixed twistor $\nbigd$-module $\nbigt_{\star}(g,D)$ with the graded polarization $\nbigs_g[\star(f)_{\infty}]$. The induced graded sesqui-linear duality $\bigl[\rho_{\dagger}^0\nbigs_g[\star (f)_{\infty}]\bigr]$ on $\rho^0_{\dagger}\nbigt_{\star}(g,D)$ is a graded polarization. \item Let $\nbigk_Y$ and $\nbigc_Y$ denote the kernel and the cokernel of $\rho_{1\dagger}^0\nbigt(F)[!\tau] \lrarr \rho_{1\dagger}^0\nbigt(F)[\ast\tau]$. Let $\vecnbigs_{\nbigc_Y}$ and $\vecnbigs_{\nbigk_Y}$ denote the graded polarizations induced by the graded polarization $\rho_{1\dagger}^0\nbigs_F$ of $\rho_{1\dagger}^0\nbigt(F)$. Then, under the natural isomorphisms $\iota_{Y\dagger} \rho^0_{\dagger}\nbigt_{!}(g,D) \simeq \nbigk_Y$ and $\iota_{Y\dagger} \rho^0_{\dagger}\nbigt_{\ast}(g,D) \otimes\newTate(-1) \simeq \nbigc_Y$, we have $\iota_{Y\dagger} \bigl[\rho_{\dagger}^0\nbigs_g[!(f)_{\infty}]\bigr] =\vecnbigs_{\nbigc_Y}$ and $\iota_{Y\dagger} \bigl[\rho_{\dagger}^0\nbigs_g[\ast(f)_{\infty}]\bigr] =\vecnbigs_{\nbigk_Y}$. \end{itemize} \end{cor} \pf It follows from Corollary and Proposition . \hfill\qed \vspace{.1in} Let $\kappa:X'\lrarr X$ be a projective morphism of complex manifolds such that (i) $D'=\kappa^{-1}(D)$ is normal crossing, (ii) $\kappa$ induces $X'\setminus D'\simeq X\setminus D$, (iii) Assumption is also satisfied for $f'=\kappa^{\ast}(f)$ and $g'=\kappa^{\ast}(g)$. We have $\kappa_{\dagger}^i\nbigt_{\star}(g',D')=0$ $(i\neq 0)$, and we have natural isomorphisms $\kappa_{\dagger}^0\nbigt_{\star}(g',D') \simeq \nbigt_{\star}(g,D)$. We set $\rho':=\rho\circ\kappa$. We have natural isomorphisms $\rho_{\dagger}^{\prime\,0} \nbigt_{\star}(g',D') \simeq \rho_{\dagger}^0\nbigt_{\star}(g,D)$. \begin{cor}\mbox{{}} \begin{itemize} \item Condition {\bf(A)} holds for $\kappa$ and $\nbigt_{\star}(g',D')$ with $\nbigs_{g'}[\star (f')_{\infty}]$. We have $\bigl[ \kappa^0_{\dagger}\nbigs_{g'}[\star (f')_{\infty}] \bigr] =\nbigs_g[\star(f)_{\infty}]$ under the natural isomorphisms. \item Condition {\bf(A)} holds for $\rho'$ and $\nbigt_{\star}(g',D')$ with $\nbigs_{g'}[\star (f')_{\infty}]$. We have $\bigl[ \rho^{\prime 0}_{\dagger}\nbigs_{g'}[\star (f')_{\infty}] \bigr] =\bigl[ \rho^0_{\dagger}\nbigs_g[\star(f)_{\infty}] \bigr]$ under the natural isomorphisms. \hfill\qed \end{itemize} \end{cor} \subsubsection{Preliminary} Let us return to the proof of Proposition . By Lemma , it is enough to study the polarizations on $\Cok(\varphi) \simeq \iota_{\dagger} \nbigt_{\ast}(g,D)\otimes\newTate(-1)$. The comparison of the polarizations on $\Gr^W_{d+2}\Cok(\varphi) \simeq \Gr^W_{d+2}\bigl( \iota_{\dagger}\nbigt_{\ast}(g,D) \otimes\newTate(-1)\bigr)$ can be given as in \S. Hence, we shall argue the other parts. We have the identification $\Gr^W_{d+2+j}\Cok(\varphi) =P\Gr^W_{d+2+j}\psi^{(0)}_{\tau}\nbigt(F)$ for $j\geq 0$. For $j>0$, we have the isomorphism $\Gr^W_{d+2+j}\bigl(\nbigt_{\ast}(g,D)\otimes\newTate(-1)\bigr) \simeq P\Gr^W_{d+2+j}\psi_{f^{-1}}^{(0)}(\nbigt(g)\otimes\newTate(-1))$. Let $\pi:X\times\proj^1_{\tau}\lrarr X$ be the projection. To compare $\nbigs_{\Cok(\varphi),d+j+2}$ and $\iota_{\dagger}\nbigs_g[\ast(f)_{\infty}]_{d+j+2}$ for $j> 0$, it is enough to compare $\pi_{\dagger}^0\nbigs_{\Cok(\varphi),d+j+2}$ and $\nbigs_g[\ast(f)_{\infty}]_{d+j+2}$ under the isomorphisms: \begin{equation} \pi_{\dagger}^0 P\Gr^W_{d+2+j}\psi_{\tau}^{(0)}\nbigt(F) \simeq P\Gr^W_{d+2+j} \psi_{f^{-1}}^{(0)}\bigl(\nbigt(g)\otimes\newTate(-1)\bigr) \end{equation} \begin{equation} \Bigl( P\Gr^W_{d+2+j} \psi_{f^{-1}}^{(0)}\bigl(\nbigt(g)\otimes\newTate(-1)\bigr) \Bigr)^{\ast} \simeq \Bigl( \pi_{\dagger}^0 P\Gr^W_{d+2+j}\psi_{\tau}^{(0)}\nbigt(F) \Bigr)^{\ast} \end{equation} Recall that any morphisms of pure twistor $\nbigd$-modules are uniquely determined by their underlying morphisms of $\nbigd$-modules. Hence, it is enough to compare the morphisms of the underlying $\nbigd$-modules for $\pi_{\dagger}^0\nbigs_{\Cok(\varphi),d+j+2}$ and $\nbigs_g[\ast(f)_{\infty}]_{d+j+2}$. The polarization $\nbigs_{\Cok(\varphi),d+2+j}$ of $P\Gr^W_{d+2+j}\psi_{\tau}^{(0)}\nbigt(F)$ is induced by the composite of the following morphisms: \begin{multline} \psi_{\tau}^{(0)}\nbigt(F) \stackrel{(-\nbign)^j}{\lrarr} \psi_{\tau}^{(-j)}(\nbigt(F)) =\psi_{\tau}^{(0)}\nbigt(F)\otimes\newTate(-j) \\ \stackrel{\nbigs_F}{\lrarr} \psi_{\tau}^{(0)}(\nbigt(F)^{\ast}) \otimes\newTate(-j-d-1) \lrarr \psi_{\tau}^{(1)}\bigl( \nbigt(F) \bigr)^{\ast}\otimes\newTate(-d-j-1) \\ \simeq \psi_{\tau}^{(0)}(\nbigt(F))^{\ast} \otimes\newTate(-d-j-2) \end{multline} The polarization $\nbigs_g[\ast(f)_{\infty}]_{d+2+j}$ on $\Gr^W_{d+2+j}\bigl(\nbigt_{\ast}(g,D)\otimes\newTate(-1)\bigr)$ is induced by the composite of the following morphisms: \begin{multline} \psi_{f^{-1}}^{(0)}\bigl( \nbigt(g)\otimes\newTate(-1) \bigr) \stackrel{(-\nbign)^{j-1}} \lrarr \psi_{f^{-1}}^{(0)}\bigl( \nbigt(g)\otimes\newTate(-1) \bigr)\otimes\newTate(-j+1) \\ \stackrel{\nbigs_{g}} {\lrarr} \psi_{f^{-1}}^{(0)}\bigl( \nbigt(g)^{\ast} \otimes\newTate(-1) \bigr) \otimes\newTate(-j-d-1) \simeq \psi_{f^{-1}}^{(1)}\bigl( \nbigt(g) \otimes\newTate(-1) \bigr)^{\ast} \otimes\newTate(-j-d-1) \\ \simeq \psi_{f^{-1}}^{(0)}\bigl( \nbigt(g) \otimes\newTate(-1) \bigr)^{\ast} \otimes\newTate(-j-d-2) \end{multline} Let $L(F)$ denote the $\nbigd_{X\times\proj^1}$-module underlying $\nbigt(F)$. Let $L(g)$ denote the $\nbigd_X$-module underlying $\nbigt(g)$. By the assumption of purity, we have $L(F)[\star\tau]=L_{!}(F,D^{(1)})[\star\tau] =L_{\ast}(F,D^{(1)})[\star\tau]$ and $L_{\star}(g,D)=L(g)[\star (f)_{\infty}]$. By the above consideration, Proposition is reduced to the following lemma. \begin{lem} We have the following morphisms of $\nbigd$-modules \[ B_1: \pi_{\dagger}^0 \psi_{\tau}^{(0)}L(F) \lrarr \psi^{(0)}_{f^{-1}}L(g), \quad B_2: \psi^{(1)}_{f^{-1}}L(g) \lrarr \pi_{\dagger}^0 \psi_{\tau}^{(1)}L(F) \] with the following properties: \begin{description} \item[(P1)] $B_1$ and $B_2$ induce the isomorphisms of $\nbigd$-modules underlying {\rm()} and {\rm()}. \item[(P2)] The following diagrams are commutative: \begin{equation} \begin{CD} \pi_{\dagger}^0\psi_{\tau}^{(0)}(L(F)) @>{\epsilon}>> \pi_{\dagger}^0\psi_{\tau}^{(1)}L(F) \\ @V{B_1}VV @A{B_2}AA \\ \psi_{f^{-1}}^{(0)}L(g) @>{\delta}>> \psi_{f^{-1}}^{(1)}L(g) \end{CD} \end{equation} Here, $\epsilon$ gives the morphism underlying the composite of the morphisms in {\rm()} applied $\pi_{\dagger}^{0}$, and $\delta$ gives the morphism underlying the composite of the morphisms in {\rm()}. \end{description} \end{lem} \subsubsection{The construction of the morphisms $B_1$ and $B_2$} As in Proposition , we have a natural isomorphism $\Lambda^{(a)}:\pi_{+}^0\psi_{\tau}^{(a)}L_{\ast}(F) \simeq \Xi^{(a)}_{f^{-1}}L_{\ast}(g)$. Let $B_1$ denote the composite of $\Lambda^{(0)}$ and the natural morphism $\Xi^{(0)}_{f^{-1}}L_{\ast}(g) \lrarr \psi^{(0)}_{f^{-1}}L_{\ast}(g)$. Let $B_2$ be the composite of the following morphisms: \[ \begin{CD} \psi^{(1)}_{f^{-1}}L_{\ast}(g) @>{c_1}>> \Xi^{(0)}_{f^{-1}}L_{\ast}(g) @>{c_2}>{\simeq}> \Xi^{(1)}_{f^{-1}}L_{\ast}(g) @>{c_3}>{\simeq}> \pi_{+}^0\psi_{\tau}^{(1)}L_{\ast}(F) \end{CD} \] Here, $c_1$ is the canonical morphism, $c_2$ is induced by the multiplication of $-s$, and $c_3$ is $(\Lambda^{(1)})^{-1}$. \subsubsection{Proof of {\bf (P1)}} Let us check that $B_1$ induces the morphism of $\nbigd$-modules underlying (). We have the following natural morphisms: \begin{multline} \pi_+ \psi^{(0)}_{\tau}L(F) \stackrel{a_1}{\lrarr} \pi_+ \Cok\bigl( \psi^{(1)}_{\tau}L(F)\lrarr \psi^{(0)}_{\tau}L(F) \bigr) \stackrel{a_2}{\simeq} \pi_+ \Cok\bigl( L(F)[!\tau]\lrarr L(F)[\ast\tau] \bigr) \stackrel{a_3}{\simeq} L_{\ast}(g,D) \\ \lrarr \Cok\Bigl( L_!(g,D)\lrarr L_{\ast}(g,D) \Bigr) \end{multline} (See () for $a_3$.) The composite induces the morphism of $\nbigd$-modules underlying (). Under the natural isomorphism \[ \Cok\Bigl( L_!(g,D)\lrarr L_{\ast}(g,D) \Bigr) \simeq \Cok\Bigl( \psi_{f^{-1}}^{(1)}L(g) \lrarr \psi_{f^{-1}}^{(0)}L(g) \Bigr)=:\nbiga, \] the following diagram is commutative: \[ \begin{CD} \Xi^{(0)}_{f^{-1}}L(g) @>>> L_{\ast}(g,D)\\ @VVV @VVV \\ \psi^{(0)}_{f^{-1}}L(g) @>>> \nbiga \end{CD} \] Hence, it is enough to prove that $b_1:=a_3\circ a_2\circ a_1$ is equal to the composite $b_2$ of $\Lambda^{(0)}$ and $\Xi^{(0)}_{f^{-1}}L(g)\lrarr L_{\ast}(g,D)$. The kernel of both of $b_i$ are the image of $\pi_+\psi^{(1)}_{\tau}L(F) \lrarr \pi_+\psi^{(0)}_{\tau}L(F)$. The automorphisms of $L_{\ast}(g,D)$ are only the multiplication of locally constant functions. Hence, it is enough to check the claim in the case where $D$ is empty, in which both of the morphisms are given by the correspondence $\tau^{-1}d\tau\longmapsto 1$. Therefore, we obtain that $B_1$ induces the morphism of $\nbigd$-modules underlying (). Let us check that $B_2$ induces the morphism of $\nbigd$-modules underlying (). We have the following natural morphisms: \begin{multline} \Ker\Bigl( L_!(g,D)\lrarr L_{\ast}(g,D) \Bigr) \lrarr L_!(g,D) \stackrel{c_1}{\simeq} \pi_+ \Ker\bigl( L(F)[!\tau]\lrarr L(F)[\ast\tau] \bigr) \\ \stackrel{c_2}{\simeq} \pi_+ \Ker\bigl( \psi^{(1)}_{\tau}L(F)\lrarr \psi^{(0)}_{\tau}L(F) \bigr) \stackrel{c_3}{\lrarr} \pi_+ \psi^{(1)}_{\tau}L(F) \end{multline} (See () for $c_1$.) As before, it is enough to check that $d_1:=c_3\circ c_2\circ c_1$ is equal to the composite $d_2$ of the following: \[ L_!(g,D)\lrarr \Xi^{(0)}_{f^{-1}}L(g) \stackrel{-s}{\simeq} \Xi^{(1)}_{f^{-1}}L(g) \stackrel{(\Lambda^{(1)})^{-1}}{\lrarr} \pi_+^0\psi_{\tau}^{(1)}L_{\ast}(F) \] The images of $d_i$ are equal, and the automorphisms of $L_!(g,D)$ are multiplications of locally constant functions. Hence, it is enough to consider the case $D$ is empty. We have the following: \begin{equation} \psi^{(1)}_{\tau}L(F) \simeq \Cok\Bigl( \Pi^{1,N}_{\tau}L(F)[!\tau] \lrarr \Pi^{1,N}_{\tau}L(F)[\ast\tau] \Bigr) \simeq \Ker\Bigl( \Pi^{-N,1}_{\tau}L(F)[!\tau] \lrarr \Pi^{-N,1}_{\tau}L(F)[\ast\tau] \Bigr) \end{equation} Under the isomorphism (), $-\del_{\tau}(1)+f$ in $\Pi^{-N,1}_{\tau}L(F)[!\tau]$ corresponds to $-s/\tau$ in $\Pi^{1,N}_{\tau}L(F)[\ast\tau]$. By using it, we can check $d_1=d_2$. Hence, we obtain that $B_2$ induces the isomorphism of $\nbigd$-modules underlying (). \subsubsection{Proof of {\bf(P2)}} Let us look at the morphisms of the $\nbigd$-module underlying (): \begin{equation} \psi_{\tau}^{(0)}L(F) \stackrel{(-s)^j}\lrarr \psi_{\tau}^{(0)}L(F) \stackrel{(-1)^{d+1}}{\lrarr} \psi_{\tau}^{(0)}L(F) \stackrel{-s}{\lrarr} \psi_{\tau}^{(1)}L(F) \end{equation} By applying $\pi_+^0$, we obtain the following morphisms: \[ \Xi^{(0)}_{f^{-1}}\bigl( L(g) \bigr) \stackrel{(-s)^j}{\lrarr} \Xi^{(0)}_{f^{-1}}\bigl( L(g) \bigr) \stackrel{(-1)^{d+1}}{\lrarr} \Xi^{(0)}_{f^{-1}}\bigl( L(g) \bigr) \\ \stackrel{-s}{\simeq} \Xi^{(1)}_{f^{-1}}\bigl( L(g) \bigr) \] Let us look at the morphism of $\nbigd$-modules underlying (): \[ \psi^{(0)}_{f^{-1}}(L(g)) \stackrel{(-s)^{j-1}}{\lrarr} \psi^{(0)}_{f^{-1}}(L(g)) \stackrel{(-1)^{d}}{\simeq} \psi^{(0)}_{f^{-1}}(L(g)) \stackrel{-1}{\simeq} \psi^{(0)}_{f^{-1}}(L(g)) \stackrel{-s}{\simeq} \psi^{(1)}_{f^{-1}}(L(g)) \] We have the following commutative diagram: \[ \begin{CD} \Xi^{(0)}_{f^{-1}}L(g) @>{(-1)^{d+1}(-s)^j}>> \Xi^{(0)}_{f^{-1}}L(g) @>{-s}>> \Xi^{(1)}_{f^{-1}}L(g) \\ @VVV @AAA \\ \psi^{(0)}_{f^{-1}}L(g) @>{(-1)^{d+1}(-s)^{j}}>> \psi^{(1)}_{f^{-1}}L(g) \end{CD} \] It implies the commutativity of (). Thus, we finish the proof of Lemma and Proposition . \hfill\qed \subsection{A push-forward} \subsubsection{Statement} Let $X$ be a complex manifold with a normal crossing hypersurface $D$. Set $d:=\dim X$. \begin{assumption} Let $f$ be a meromorphic function on $(X,D)$ such that (i) $|(f)_{\infty}|=D$, (ii) $f$ is non-degenerate along $D$, (iii) $(f)_0$ is smooth. \hfill\qed \end{assumption} We set $Z:=|(f)_0|$. We have the meromorphic function $tf$ on $X\times\proj^1_t$. We have the associated mixed twistor $\nbigd$-modules $\nbigt(tf)[\star t]$ and $\nbigt_{\star}(tf)$. Let $\pi:X\times\proj^1_t\lrarr X$ be the projection. According to Proposition and Corollary , we have $\pi_{\dagger}^i\nbigt(tf)[\star t]=0$ for $i\neq 0$, and we have the isomorphisms \begin{equation} \Psi_{\ast}: \pi_{\dagger}^0\nbigt(tf)[\ast t] \simeq \nbigu_X(d,0)[!Z][\ast D]\otimes\newTate(-1), \quad \Psi_!: \pi_{\dagger}^0\nbigt(tf)[! t] \simeq \nbigu_X(d,0) [\ast Z][! D]. \end{equation} Recall that $\nbigt(tf)$ is the image of $\nbigt_{!}(tf)\lrarr \nbigt_{\ast}(tf)$. We have the description $\nbigt(tf) =\bigl( \nbigl(tf)\lambda^{d+1}, \nbigl(tf),C_{tf} \bigr)$ as an $\nbigr$-triple. The polarization is given by $\nbigs_{tf}=\bigl((-1)^{d+1},(-1)^{d+1}\bigr)$. It induces graded polarizations $\vecnbigs_{tf\star}$ (resp. $\nbigs_{tf}[\star t]$) on $\nbigt_{\star}(tf)$ (resp. $\nbigt(tf)[\star t]$). The polarization $\nbigs_0=((-1)^d,(-1)^d)$ of $\nbigu_X(d,0)$ induces graded polarizations $\nbigs_0[\ast (f)_0][!(f)_{\infty}]$ on $\nbigu_X(d,0)[\ast Z][!D]$, and $\nbigs_0[!(f)_0][\ast (f)_{\infty}]$ on $\nbigu_X(d,0)[! Z][\ast D]\otimes\newTate(-1)$. \begin{prop} \mbox{{}} \begin{itemize} \item We have $\pi_{\dagger}^i\Gr^W\nbigt(tf)[\star t]=0$ and $\pi_{\dagger}^i\gr^w\nbigt_{\star}(tf)=0$ unless $i\neq 0$. In particular, we have \[ \Gr^W \pi^0_{\dagger}\nbigt_{\star}(tf) \simeq \pi^0_{\dagger} \Gr^W \nbigt_{\star}(tf), \quad\quad \Gr^W \pi^0_{\dagger}\nbigt(tf)[\star t] \simeq \pi^0_{\dagger} \Gr^W \nbigt(tf)[\star t]. \] Note that it implies that Condition {\bf (A)} holds for the projection $\pi$ and the mixed twistor $\nbigd$-modules $\nbigt(tf)[\star t]$ with $\nbigs_{tf}[\star t]$, and the mixed twistor $\nbigd$-modules $\nbigt_{\star}(tf)$ with $\vecnbigs_{tf\star}$. \item The isomorphisms $\Gr^W\Psi_{\star}$ $(\star=\ast,!)$ are compatible with the graded polarizations. \end{itemize} \end{prop} Before going to the proof of Proposition , we give a consequence. Let $\rho:X\lrarr Y$ be any projective morphism of complex manifolds such that $R^j\rho_{\ast}\Omega_X^i(\ast D)=0$ for any $j>0$ and $i=0,1,\ldots,\dim X$. We also suppose $\rho_{\dagger}^i\nbigu_X(d,0)[!Z_f][\ast D]=0$ for $i\neq 0$. Note that it implies $\rho_{\dagger}^i\nbigu_X(d,0)[\ast Z_f][! D]=0$ for $i\neq 0$. \begin{cor}\mbox{{}} Condition {\bf (A)} holds for $\rho$ and $\nbigu_X(d,0)[!Z_f][\ast D]$ with $\nbigs_0[!(f)_0][\ast (f)_{\infty}]$. We have \[ \bigl[ \rho_{\dagger}^0 \nbigs_{0}[!(f)_0][\ast (f)_{\infty}] \bigr] =\bigl[ (\rho\circ\pi)_{\dagger}^0 \nbigs_{tf}[\ast (t)_0] \bigr]. \] Condition {\bf (A)} holds for $\rho$ and $\nbigu_X(d,0)[\ast Z_f][!D]$ with $\nbigs_0[\ast(f)_0][!(f)_{\infty}]$. We have \[ \bigl[ \rho_{\dagger}^0 \nbigs_{0}[\ast(f)_0][!(f)_{\infty}] \bigr] =\bigl[ (\rho\circ\pi)_{\dagger}^0 \nbigs_{tf}[!(t)_0] \bigr]. \] \end{cor} \pf Set $X^{(1)}:=X\times\proj^1_t$ and $D^{(1)}:=(X\times\{0,\infty\})\cup(D\times\proj^1)$. We have $R^j(\rho\circ\pi)_{\ast} \Omega_{X^{(1)}}^i(\ast D^{(1)})=0$ for any $j>0$ and $i=0,\ldots,\dim X^{(1)}$. By the assumption and Proposition , we have $(\rho\circ\pi)_{\dagger}^j\nbigt_{\star}(tf,D^{(1)})=0$ for $j\neq 0$. According to Corollary and Lemma , Condition {\bf(A)} holds for $\rho\circ\pi$ and $\nbigt(tf)[\star t]=\nbigt_{\star}(tf,D^{(1)})$ with $\nbigs_{tf}[\ast t]$. Then, the claim follows from Proposition . \hfill\qed \subsubsection{First claim of Proposition } Let us prove the first claim of Proposition . We argue the case $\star=\ast$. We obtain the other case by using the Hermitian duality. We have $\Gr^W_j\nbigt(tf)[\ast t]= \Gr^W_{j}\nbigt_{\ast}(tf)=0$ for $j<d+1$, and $\Gr^W_{d+1}\nbigt(tf)[\ast t]= \Gr^W_{d+1}\nbigt_{\ast}(tf)= \nbigt(tf)$. The support of $\Gr^W_{j}\nbigt(tf)[\ast t]$ and $\Gr^W_{j}\nbigt_{\ast}(tf)$ $(j>d+1)$ are contained in $X\times\{t=0\}$. Hence, it is enough to prove $\pi_{\dagger}^i\nbigt(tf)=0$ for $i\neq 0$. Let $A_1$ and $A_2$ denote the kernel of $\nbigt(tf)[!t]\lrarr \nbigt(tf)$ and the cokernel of $\nbigt(tf)\lrarr \nbigt(tf)[\ast t]$. The support of $A_i$ are contained in $\{t=0\}$. Hence, we obtain $\pi_{\dagger}^i\nbigt(tf)=0$ $(i\neq 0)$, and the following exact sequences: \[ 0\lrarr \pi_{\dagger}^0A_1 \lrarr \pi_{\dagger}^0\nbigt(tf)[!t] \lrarr \pi_{\dagger}^0\nbigt(tf) \lrarr 0 \] \[ 0\lrarr \pi_{\dagger}^0\nbigt(tf) \lrarr \pi_{\dagger}^0\nbigt(tf)[\ast t] \lrarr \pi_{\dagger}^0A_2 \lrarr 0 \] In particular, we obtain the first claim of Proposition . \subsubsection{Preliminary for the second claim} Let us study the second claim. Let $A_3$ denote the cokernel of $\nbigt_{\ast}(tf)\lrarr \nbigt(tf)[\ast t]$. Let $\iota_Z:Z\lrarr X$ denote the inclusion, and let $D_Z:=D\cap Z$. According to Proposition , we have the following isomorphisms of integrable mixed twistor $D$-modules with real structure: \[ \Psi_1: \pi_{\dagger}^0\nbigt_{\ast}(tf) \simeq \iota_{Z\dagger}\nbigu_Z(d-1,0)[\ast D_Z] \otimes\newTate(-1), \quad\quad \Psi_2: \pi_{\dagger}^0A_3\simeq \nbigu_X(d,0)\otimes\newTate(-1)[\ast D] \] Note that the strict support of $\Gr^W \Bigl( \iota_{Z\dagger}\nbigu_Z(d-1,0)[\ast D_Z]\otimes\newTate(-1) \Bigr)$ are contained in $Z$, and that the strict support of any direct summand of $\Gr^W\nbigu_X(d,0)\otimes\newTate(-1)[\ast D]$ are not contained in $Z$. Hence, we have the unique splittings \begin{equation} \Gr^W\pi^0_{\dagger}\nbigt(tf)[\ast t] =\Gr^W\pi^0_{\dagger}\nbigt_{\ast}(tf) \oplus \Gr^W\pi^0_{\dagger}A_3, \end{equation} \begin{equation} \Gr^W\Bigl( \nbigu_X(d,0)\otimes\newTate(-1)[!Z][\ast D] \Bigr) = \Gr^W\Bigl( \iota_{Z\dagger}\nbigu_Z(d-1,0)[\ast D_Z]\otimes\newTate(-1) \Bigr) \oplus \Gr^W\nbigu_X(d,0)[\ast D]\otimes\newTate(-1), \end{equation} and we have $\Gr^W\Psi_{\ast}=\Gr^W\Psi_1\oplus \Gr^W\Psi_2$. The graded polarizations are compatible with the decompositions (, ). By Proposition , $\Gr^W\Psi_{2|X\setminus Z}$ is compatible with the graded polarizations. Hence, we obtain that $\Gr^W\Psi_{2}$ is compatible with the graded polarizations. Therefore, it is enough to study that $\Gr^W\Psi_1$ is compatible with the graded polarizations. We remark that the graded polarization on $\Gr^W\Bigl( \iota_{Z\dagger}\nbigu_Z(d-1,0)[\ast D_Z]\otimes\newTate(-1) \Bigr)$ is equal to the graded polarization induced by the canonical polarization of $\nbigu_Z(d-1,0)$, which follows from Lemma . \subsubsection{The degree $(d+1)$-part} Let us study the isomorphism $\Gr^W_{d+1}\Psi_1: \pi_{\dagger}^0\nbigt(tf) \simeq \iota_{Z\dagger}\nbigu_Z(d-1,0)\otimes\newTate(-1)$. We may assume that $D$ is empty. To describe $\Gr^W_{d+1}\Psi_1$, let us look at the isomorphisms of the $\nbigr$-modules underlying $\Psi_{\ast}$ more closely, by assuming $D=\emptyset$. Set $Y:=X\times\proj^1$. We have the following: \[ \pi_{\dagger}^0\nbigl(tf)[\ast t] \simeq \Cok\Bigl( \pi_{\ast} \nbigl(tf)[\ast t] \stackrel{\varphi_1}{\lrarr} \pi_{\ast} \nbigl(tf)[\ast t] \cdot (dt/\lambda) \Bigr) \simeq \Cok\Bigl( \nbigo_{\nbigx}[t] \stackrel{\varphi_2}{\lrarr} \nbigo_{\nbigx}[t]\cdot (dt/t\lambda) \Bigr) \] Here, $\varphi_1(c)=\deldel_t(c)\cdot (dt/\lambda)$ and $\varphi_2(h\,t^j) =\bigl( jht^{j-1}+hft^j/\lambda \bigr)dt$. For any local section $c$ of $\nbigo_{\nbigx}[t]\cdot dt/t\lambda$, let $[c]_{\ast}$ denote the induced section of $\pi_{\dagger}^0\nbigl(tf)[\ast t]$. The isomorphism $\pi_{\dagger}^0\bigl( \nbigl(tf)[\ast t]\bigr)(\ast Z) \simeq \lambda^{-1}\nbigo_{\nbigx}(\ast Z)$ is given by $[dt/\lambda t]_{\ast} \longleftrightarrow \lambda^{-1}$, which we can check by using (). Similarly, we have the following: \[ \pi_{\dagger}^0\bigl( \nbigl(tf)[!t] \bigr) \simeq \Cok\Bigl( \pi_{\ast} \nbigl(tf)[!t] \stackrel{\varphi_1}{\lrarr} \pi_{\ast} \nbigl(tf)[!t]\cdot(dt/\lambda) \Bigr) \simeq \Cok\Bigl( \nbigo_{\nbigx}[t]\cdot t \stackrel{\varphi_2}{\lrarr} \nbigo_{\nbigx}[t]\cdot(dt/\lambda) \Bigr) \] Here, $\varphi_i$ are given in the same ways. For any local section $c$ of $\nbigo_{\nbigx}[t]\cdot(dt/\lambda)$, let $[c]_!$ denote the induced section of $\nbigo_{\nbigx}[t]$. The isomorphism $\pi_{\dagger}^0(\nbigl(tf)[!t])(\ast Z) \simeq \nbigo_{\nbigx}(\ast Z)$ is given by $[fdt/\lambda]_!\longleftrightarrow 1$, which we can check by using (). We have the following: \[ \pi_{\dagger}^0\nbigl(tf) \simeq \Cok\Bigl( \pi_{\ast}\nbigl(tf) \stackrel{\varphi_1}{\lrarr} \pi_{\ast}\nbigl(tf) \cdot(dt/\lambda) \Bigr) \simeq \Cok\Bigl( \nbigo_{\nbigx}[t] \stackrel{\varphi_2}{\lrarr} \nbigo_{\nbigx}[t] \cdot(dt/\lambda) \Bigr) \] For any local section of $\nbigo_{\nbigx}[t](dt/\lambda)$, let $[c]$ denote the induced local section of $\pi_{\dagger}^0\nbigl(tf)$. We remark $\lambda^2\del_{\lambda}[dt/\lambda]=0$. \begin{lem} The isomorphism $\Gr^W_{d+1}\Psi_1: \pi_{\dagger}\nbigt(tf)\simeq \iota_{Z\dagger}\nbigu_Z(d-1,0)\otimes\newTate(-1)$ is given as a pair of the following morphisms: \[ \pi_{\dagger}\nbigl(tf)\lambda^{d+1} \stackrel{a_1}{\llarr} \iota_{Z\dagger}\nbigo_{\nbigz}\lambda^{d}, \quad a_1\bigl( \iota_{Z\ast}(\lambda^d(dt/\lambda)^{-1}) \bigr) =[dt/\lambda]\cdot\lambda^{d+1} \] \[ \pi_{\dagger}\nbigl(tf) \stackrel{a_2}{\lrarr} \iota_{Z\dagger}\nbigo_{\nbigz}\lambda^{-1}, \quad a_2\bigl([dt/\lambda]\bigr) =-\iota_{Z\ast}(\lambda^{-1}(dt/\lambda)^{-1}) \] \end{lem} \pf We have the epimorphism $b_1: \pi_{\dagger}^0\nbigl(tf)[!t] \lrarr \pi_{\dagger}^0\nbigl(tf)$ and the monomorphism $b_2: \pi_{\dagger}^0\nbigl(tf) \lrarr \pi_{\dagger}^0\nbigl(tf)[\ast t]$. We have $b_1([c]_!)=[c]$ and $b_2([c])=[c]_{\ast}$. We have $[dt/\lambda]_{!}=f^{-1}[fdt/\lambda]_{!}$. For a holomorphic local coordinate $(x_1,\ldots,x_n)$ for $x_1=f$, we have $[dt/\lambda]_{\ast} =\deldel_{x_1}[dt/t\lambda]_{\ast}$. Then, we obtain the claim of the lemma by construction. \hfill\qed \vspace{.1in} We can easily check that the isomorphism $\Gr^W_{d+1}\Psi_1$ is compatible with the polarizations by the explicit description in Lemma . \subsubsection{Preliminary for the other parts} Let us study $\Gr^W_{d+2+\ell}\Psi_1$ for $\ell\geq 0$. We have only to consider the issue locally around any point of $D\cap Z$. We may assume to have a holomorphic coordinate $(x_0,\ldots,x_{d-1})$ on $X$ such that $Z=\{x_0=0\}$ and $D=\bigcup_{i=1}^{m}\{x_i=0\}$, such that $f=x_0\prod_{i=1}^mx_i^{k_i}$. We set $x^{\veck}:=\prod_{i=1}^mx_i^{k_i}$. We have the isomorphism \begin{equation} \phi^{(0)}_{x_0} \pi_{\dagger}^0\Gr^W_{d+\ell+2}(\nbigt(tf)[\ast t]) \simeq P\Gr^W_{d+\ell+2} \phi^{(0)}_{x_0}\pi_{\dagger}\psi_t^{(0)}(\nbigt(tf)) \end{equation} The polarization on $\phi^{(0)}_{x_0} \pi_{\dagger}^0\Gr^W_{d+\ell+2}(\nbigt(tf)[\ast t])$ is induced by the composite of the following: \begin{multline} \phi_{x_0}^{(0)} \pi_{\dagger} \psi_t^{(0)}(\nbigt(tf)) \stackrel{(-\nbign)^{\ell}}{\lrarr} \phi_{x_0}^{(0)} \pi_{\dagger} \psi_t^{(0)}(\nbigt(tf)) \otimes\newTate(-\ell) \\ \stackrel{a_1}{\lrarr} \phi_{x_0}^{(0)} \pi_{\dagger} \psi_t^{(0)}(\nbigt(tf)^{\ast}) \otimes\newTate(-\ell-d-1) =\phi_{x_0}^{(0)} \pi_{\dagger} \psi_t^{(1)}(\nbigt(tf))^{\ast} \otimes\newTate(-\ell-d-1) \\ \stackrel{a_2}{\lrarr} \phi_{x_0}^{(0)} \pi_{\dagger} \psi_t^{(0)}(\nbigt(tf))^{\ast} \otimes\newTate(-\ell-d-2) \end{multline} Here, $a_1$ is induced by the polarization on $\nbigt(tf)$, and $a_2$ is the canonical isomorphism. We have the isomorphism \begin{equation} \Gr^W_{d+\ell+2} \iota_{Z\dagger}\nbigu_Z(d,-1)[\ast D_Z] \simeq P\Gr^W_{d+\ell+2}\psi^{(0)}_{x^k} \iota_{Z\dagger}\nbigu_Z(d,-1) \end{equation} The polarization on $\Gr^W_{d+\ell+2} \iota_{Z\dagger}\nbigu_Z(d,-1)[\ast D_Z]$ is induced by the composite of the following morphisms: \begin{multline} \psi^{(0)}_{x^k} \iota_{Z\dagger}\nbigu_Z(d,-1) \stackrel{(-\nbign)^{\ell}}{\lrarr} \psi^{(0)}_{x^k} \iota_{Z\dagger}\nbigu_Z(d,-1) \otimes\newTate(-\ell) \\ \stackrel{b_1}{\lrarr} \psi^{(0)}_{x^k}\bigl( \iota_{Z\dagger}\nbigu_Z(d,-1)^{\ast} \bigr) \otimes\newTate(-\ell-d-1) =\psi^{(1)}_{x^k}\bigl( \iota_{Z\dagger}\nbigu_Z(d,-1) \bigr)^{\ast} \otimes\newTate(-\ell-d-1) \\ \stackrel{b_2}{\lrarr} \psi^{(0)}_{x^k}\bigl( \iota_{Z\dagger}\nbigu_Z(d,-1) \bigr)^{\ast} \otimes\newTate(-\ell-d-2) \end{multline} Here, $b_1$ is induced by the polarization of $\nbigu_Z(d,-1)$, and $b_2$ is the canonical isomorphism. We have the following isomorphism induced by (), () and $\Gr^W_{d+\ell+2}\Psi_1$: \begin{equation} \Lambda: P\Gr^W_{d+\ell+2} \phi^{(0)}_{x_0}\pi_{\dagger}\psi_t^{(0)}(\nbigt(tf)) \simeq P\Gr^W_{d+\ell+2} \psi^{(0)}_{x^k} \iota_{Z\dagger}\nbigu_Z(d,-1) \end{equation} We obtain the compatibility of $\Gr^W_{d+\ell+2}\Psi_1$ and the polarizations from the following lemma. \begin{lem} We have isomorphisms of $\nbigd$-modules \begin{equation} \Phi_0: \phi^{(0)}_{x_0}\pi_{+}\psi_t^{(0)}L(tf) \simeq \psi_{x^k}^{(0)}\iota_{Z+}\nbigo_Z \end{equation} \begin{equation} \Phi_1: \phi^{(0)}_{x_0}\pi_+\psi_t^{(1)}L(tf) \simeq \psi_{x^k}^{(1)}\iota_{Z+}\nbigo_Z \end{equation} with the following property. \begin{description} \item[(P1)] $\Phi_0$ induces the isomorphism of the $\nbigd$-modules underlying $\Lambda$, and $\Phi_1^{-1}$ induces the isomorphism of the $\nbigd$-modules underlying $\Lambda^{\ast}$. \item[(P2)] The following diagram is commutative: \begin{equation} \begin{CD} \phi^{(0)}_{x_0}\pi_{+}\psi_t^{(0)}L(tf) @>{d_1}>> \phi^{(0)}_{x_0}\pi_{+}\psi_t^{(1)}L(tf)\\ @V{\Phi_0}VV @V{\Phi_1}VV \\ \psi_{x^k}^{(0)}\iota_{Z+}\nbigo_Z @>{d_2}>> \psi_{x^k}^{(1)}\iota_{Z+}\nbigo_Z \end{CD} \end{equation} Here, $d_1$ and $d_2$ are the morphisms of $\nbigd$-modules underlying {\rm()} and {\rm()}, respectively. \end{description} \end{lem} \subsubsection{Construction of the morphisms $\Phi_i$} According to Proposition , we have the isomorphism: \begin{multline} \pi_+\psi^{(a)}_tL(tf) \simeq \Xi^{\prime(a)}_{f^{-1}}\nbigo_X \simeq \Cok\Bigl( \Pi^{a,N}_{f^{-1}}\nbigo_X[\ast Z][!D] \stackrel{-s}{\lrarr} \Pi^{a,N}_{f^{-1}}\nbigo_X[!Z][\ast D] \Bigr) \\ \simeq \Ker\Bigl( \Pi^{-N,a}_{f^{-1}}\nbigo_X[\ast Z][!D] \stackrel{-s}{\lrarr} \Pi^{-N,a}_{f^{-1}}\nbigo_X[!Z][\ast D] \Bigr) \end{multline} We have the isomorphism of $\nbigd$-modules (see Lemma for the signature): \begin{equation} \phi_{x_0}^{(0)} \pi_+\psi^{(0)}_tL(tf) \simeq \Cok\Bigl( \iota_{Z+} \Pi^{0,N}_{x^{\veck}} \nbigo_Z[!D_Z] \stackrel{-1}{\lrarr} \iota_{Z+} \Pi^{0,N}_{x^{\veck}} \nbigo_Z[\ast D_Z] \Bigr) \simeq \iota_{Z+} \psi_{x^{\veck}}^{(0)}\nbigo_Z \end{equation} Let $\Phi_0$ be the composite of the isomorphisms in (). We have the isomorphism of $\nbigd$-modules: \begin{equation} \phi_{x_0}^{(0)}\pi_+\psi_t^{(1)}L(tf) \simeq \Ker\Bigl( \iota_{Z+} \Pi^{-N,1}_{x^{\veck}}\nbigo_Z[!D_Z] \stackrel{-1}{\lrarr} \Pi^{-N,1}_{x^{\veck}}\nbigo_Z[\ast D_Z] \Bigr) \simeq \iota_{Z+} \psi^{(1)}_{x^{\veck}}\nbigo_Z \end{equation} We define $\Phi_1$ as the composite of (). \subsubsection{Property {\bf (P1)}} Let $\Cok$ denote the cokernel of $L(tf)[!t]\lrarr L(tf)[\ast t]$. We have the following commutative diagram: \begin{equation} \begin{CD} \Pi^{0,N}_tL(tf)[\ast t] @>>> L(tf)[\ast t] \\ @VVV @VVV \\ \psi_t^{(0)}L(tf) @>>> \Cok \end{CD} \end{equation} We apply $\phi_{x_0}^{(0)}\pi_+$ to (). By the construction of the isomorphism $\Phi_0$, () is transformed to the following: \[ \begin{CD} \iota_{Z+}\Pi^{0,N}_{x^{\veck}}\nbigo_Z[\ast D_Z] @>>> \iota_{Z+}\nbigo_Z[\ast D_Z] \\ @VVV @VVV \\ \iota_{Z+}\psi^{(0)}_{x^{\veck}}\nbigo_Z @>>> \iota_{Z+}\bigl( \nbigo_Z[\ast D_Z]/\nbigo_Z \bigr) \end{CD} \] It means that $\Phi_0$ is the morphism of $\nbigd$-modules underlying $\Lambda$. Similarly, let $\Ker_1$ denote the kernel of $L(tf)[!t]\lrarr L(tf)[\ast t]$. Then, we have the following commutative diagram: \begin{equation} \begin{CD} \Ker @>>> \psi_t^{(1)} L(tf) \\ @AAA @AAA\\ L(tf)[!t] @>>> \Pi^{-N,1}L(tf)[!t] \end{CD} \end{equation} We apply $\phi_{x_0}^{(0)}\pi_+$ to (). Let $\Ker_2$ denote the kernel of $\iota_+\nbigo_Z[! D_Z]\lrarr \nbigo_Z$. Then, $()$ is transformed to the following: \[ \begin{CD} \Ker_2 @>>> \iota_{Z+}\psi_{x^{\veck}}^{(1)}\nbigo_Z \\ @AAA @AAA \\ \iota_{Z+}\nbigo_Z[!D_Z] @>>> \iota_{Z+} \Pi^{-N,1}_{x^{\veck}}\nbigo_Z[!D_Z] \end{CD} \] It implies that $\Phi_1$ underlies $\Lambda^{\ast}$. Thus, we obtain that {\bf(P1)} is satisfied. \subsubsection{Property {\bf (P2)}} The morphism $d_1$ is induced by the following morphisms: \begin{equation} \psi_{t}^{(0)}L(tf) \stackrel{(-s)^{\ell}}{\lrarr} \psi_t^{(0)}L(tf) \stackrel{(-1)^{d+1}}{\lrarr} \psi_t^{(0)}L(tf) \stackrel{-s}\simeq \psi_t^{(1)}L(tf) \end{equation} The morphism $d_2$ is induced by the following: \begin{equation} \psi^{(0)}_{x_k}\iota_{Z+}\nbigo_Z \stackrel{(-s)^{\ell}}{\lrarr} \psi^{(0)}_{x_k}\iota_{Z+}\nbigo_Z \stackrel{(-1)^{d}}{\lrarr} \psi^{(0)}_{x_k}\iota_{Z+}\nbigo_Z \stackrel{-s}{\lrarr} \psi^{(1)}_{x_k}\iota_{Z+}\nbigo_Z \end{equation} We remark that the following diagram is commutative: \[ \begin{CD} \phi_{x_0}^{(0)} \pi_+\psi^{(0)}_{t}L(tf) @>{s}>> \phi_{x_0}^{(0)} \pi_+\psi^{(1)}_{t}L(tf) \\ @V{\Phi_0}VV @V{\Phi_1}VV \\ \psi^{(0)}_{x^{\veck}} \iota_{Z+}\nbigo_Z @>{-s}>> \psi^{(1)}_{x^{\veck}} \iota_{Z+}\nbigo_Z \end{CD} \] The signature comes from $-1$ in () and (). Hence, we obtain the desired commutativity of (). Thus, we finish the proof of Lemma and Proposition . \hfill\qed \subsection{Better behaved GKZ-hypergeometric systems} \subsubsection{The associated toric varieties} We consider a finite subset $\nbiga:=\{\veca_1,\ldots,\veca_m\}\subset \seisuu^n\setminus\{(0,\ldots,0)\}$, where $\veca_i=(a_{ij}\,|\,j=1,\ldots,n)$. We assume that $\nbiga$ generates $\seisuu^n$. Formally, we set $\veca_0:=(0,\ldots,0)$. We set $T^n:=\Spec\cnum[t_1^{\pm 1},\ldots,t_n^{\pm 1}]$. The associated complex manifold is also denoted by $T^n$ if there is no risk of confusion. We consider a morphism $\psi_{\nbiga}:T^n\lrarr \proj^{m}$ given by $\psi_{\nbiga}(t_1,\ldots,t_n) =[1:t^{\veca_1}:\ldots:t^{\veca_m}]$, where $t^{\veca_i}:=\prod_{j=1}^nt_j^{a_{ij}}$. Let $X_{\nbiga}$ denote the closure of $\psi_{\nbiga}(T^n)$ in $\proj^{m}$. For any subset $\nbigb\subset\real^n$, let $\Conv(\nbigb)$ denote the convex hull of $\nbigb\subset\real^n$. Note that $0$ is an interior point of $\Conv(\nbiga\cup\{0\})$ if and only if $0$ is an interior point of $\Conv(\nbiga)$. For any face $\sigma$ of $\Conv(\nbiga\cup\{0\})$, we put $I_{\sigma}:= \sigma\cap\bigl(\nbiga\cup\{0\}\bigr)$ and $J_{\sigma}:=(\nbiga\cup\{0\})\setminus I_{\sigma}$. We have the subspace \[ \proj_{\sigma}:=\bigl\{ [z_0:z_1:\cdots:z_m]\,\big|\, z_{j}=0\,\,(a_j\in J_{\sigma}) \bigr\}. \] We set $\proj_{\sigma}^{\ast}:= \bigl\{ [z_0:\ldots:z_m]\in \proj_{\sigma}\,\big|\, z_i\neq 0\,\, (i\in I_{\sigma}), z_i=0\,\,(i\in J_{\sigma}) \bigr\}$. Recall the following. \begin{prop}[Proposition 1.9 ] For any face $\sigma$ of $\Conv(\nbiga\cup\{0\})$, we have the non-empty intersection $X_{\nbiga}^{\sigma}=X_{\nbiga}\cap \proj_{\sigma}^{\ast}$, and it is a $T^n$-orbit of $[z_0^{\sigma}:\cdots:z_m^{\sigma}]$, where $z_i^{\sigma}=1$ $(i\in I_{\sigma})$ or $z_i^{\sigma}=0$ $(i\in J_{\sigma})$. We have the decomposition into the orbits $X_{\nbiga} =\coprod_{\sigma} X_{\nbiga}^{\sigma}$ where $\sigma$ runs through the set of faces of $\Conv(\nbiga\cup\{0\})$. \hfill\qed \end{prop} Recall that we have a desingularization of $X_{\nbiga}$ in the category of toric varieties. (See for example.) Namely, there exist a smooth fan $\Sigma$ and a toric birational projective morphism $\varphi_{\Sigma}: X_{\Sigma}\lrarr X_{\nbiga}$, where $X_{\Sigma}$ denotes the smooth toric variety associated to $\Sigma$. In that situation, we set $D_{\Sigma}:=\bigcup_{\tau\in\Sigma(1)} D_{\tau}$, where $\Sigma(1)$ denotes the set of the $1$-dimensional cones of $\Sigma$, and $D_{\tau}\subset X_{\Sigma}$ denotes the hypersurface corresponding to $\tau$. We have $T^n=X_{\Sigma}\setminus D_{\Sigma}$. \subsubsection{$\nbigd$-modules associated to families of Laurent polynomials} Suppose that we are given an algebraic map $\gamma: S\lrarr H^0(\proj^m,\nbigo(1))$. It determines a family of Laurent polynomials $(\psi_{\nbiga}\times\id_S)^{\ast}(s/z_0)$ on $T^n\times S$ denoted by $F_{\gamma}$. We obtain the algebraic $\nbigd$-module $L_{T^n\times S}(F_{\gamma})$ on $T^n\times S$ given as the line bundle $\nbigo_{T^n\times S}\cdot e$ with $\nabla e=e\,dF_{\gamma}$, where $e$ denotes a global frame. We obtain the following algebraic $\nbigd$-modules on $\proj^m\times S$: \[ L_{\star}(\nbiga,S,\gamma):= (\psi_{\nbiga}\times\id_S)_{\star}L_{T^n\times S}(F_{\gamma}) \quad (\star=!,\ast) \] The image of $L_{!}(\nbiga,S,\gamma) \lrarr L_{\star}(\nbiga,S,\gamma)$ is denoted by $L_{\min}(\nbiga,S,\gamma)$, which is the minimal extension of $L_{T^n\times S}(F_{\gamma})$ on $\proj^m\times S$ via $\psi_{\nbiga}\times\id_S$. Let $\pi:\proj^m\times S\lrarr S$ denote the projection. We obtain the algebraic $\nbigd$-modules $\pi^i_{+}L_{\star}(\nbiga,S,\gamma)$ $(\star=\ast,!)$ and $\pi^i_{+}L_{\min}(\nbiga,S,\gamma)$. \vspace{.1in} It is convenient for us to take a toric desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$ when we work in the complex analytic setting or when we are interested in the twistor property of the $\nbigd$-modules. We set $(X_{\Sigma,S},D_{\Sigma,S}) :=(X_{\Sigma},D_{\Sigma})\times S$. Let $\varphibar_{\Sigma,S}$ denote $X_{\Sigma,S}\lrarr \proj^m\times S$ which is the composite of $\varphi_{\Sigma}\times\id_S$ and the inclusion $X_{\nbiga}\times S\lrarr \proj^m\times S$. We have the meromorphic function $F_{\gamma,\Sigma}:= \varphibar_{\Sigma,S}^{\ast}(s)/\varphibar_{\Sigma,S}^{\ast}(z_0)$ on $(X_{\Sigma},D_{\Sigma})\times S$. We obtain the $\nbigd_{X_{\Sigma,S}}$-modules $L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$ $(\star=\ast,!)$ on $X_{\Sigma,S}$. Then, we have $\varphibar_{\Sigma,S+}^i L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})=0$ for $i\neq 0$. In the algebraic case, we have $\varphibar_{\Sigma,S+}^0 L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq L_{\star}(\nbiga,S,\gamma)$. In the complex analytic setting, we adopt it as the definition, i.e., $L_{\star}(\nbiga,S,\gamma):= \varphibar_{\Sigma,S+}^0 L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$. Let $\pi_{\Sigma}:X_{\Sigma}\times S\lrarr S$ denote the projection. Then, we have the following natural isomorphisms for $\star=\ast,!$: \begin{equation} \pi_{\Sigma+} L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \pi_{+} L_{\star}(\nbiga,S,\gamma) \end{equation} If we are given another desingularization $\varphi_{\Sigma'}:X_{\Sigma'}\lrarr X_{\nbiga}$, we can find a smooth fan $\Sigma''$ and toric morphisms $\psi_1:X_{\Sigma''}\lrarr X_{\Sigma}$ and $\psi_2:X_{\Sigma''}\lrarr X_{\Sigma'}$ such that $\varphi_{\Sigma}\circ\psi_1 =\varphi_{\Sigma'}\circ\psi_2=: \varphi_{\Sigma''}$. We have natural isomorphisms $(\psi_1\times\id_S)_{+} L_{\star}(F_{\gamma,\Sigma''},D_{\Sigma'',S}) \simeq (\psi_1\times\id_S)^0_{+} L_{\star}(F_{\gamma,\Sigma''},D_{\Sigma'',S}) \simeq L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$. We have similar isomorphisms for $L_{\star}(F_{\gamma,\Sigma'},D_{\Sigma',S})$. Hence, we have the following natural isomorphisms on $\proj^m\times S$: \[ \varphibar^0_{\Sigma+} L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \varphibar^0_{\Sigma''+} L_{\star}(F_{\gamma,\Sigma''},D_{\Sigma'',S}) \simeq \varphibar^0_{\Sigma'+} L_{\star}(F_{\gamma,\Sigma'},D_{\Sigma',S}) \] In this sense, the $\nbigd$-modules $L_{\star}(\nbiga,S,\gamma)$ are independent of the choice of a resolution. We also have the following natural isomorphisms: \begin{equation} \pi^i_{\Sigma+}L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \pi^i_{\Sigma''+}L_{\star}(F_{\gamma,\Sigma''},D_{\Sigma'',S}) \simeq \pi^i_{\Sigma'+}L_{\star}(F_{\gamma,\Sigma'},D_{\Sigma',S}) \simeq \pi^i_{+}L_{\star}(\nbiga,S,\gamma) \end{equation} \begin{rem} Although it is convenient for us to take a toric desingularization $X_{\Sigma}$ of $X_{\nbiga}$, we can also use any toric completion of $T^n$ for the study of $\pi^i_{+}L_{\star}(\nbiga,S,\gamma)$. Let $X_{\Sigma_1}$ be any $n$-dimensional smooth toric variety. We fix an inclusion $T^n\subset X_{\Sigma_1}$. The family of Laurent polynomials $F_{\gamma}$ gives a meromorphic function $F_{\gamma,\Sigma_1}$ on $(X_{\Sigma_1,S},D_{\Sigma_1,S})$. Then, we naturally have $\pi_{\Sigma_1+}L_{\star}(F_{\gamma,\Sigma_1},D_{\Sigma_1}) \simeq \pi_{\Sigma+}L_{\star}(F_{\gamma,\Sigma},D_{\Sigma}) \simeq \pi_+L_{\star}(\nbiga,S,\gamma)$. \hfill\qed \end{rem} Let $\gamma_i:S\lrarr H^0(\proj^m,\nbigo(1))$ $(i=1,2)$ be holomorphic maps. We give some conditions under which we have $\pi^i_{+}L_{\star}(\nbiga,S,\gamma_1) \simeq \pi^i_{+}L_{\star}(\nbiga,S,\gamma_2)$. We have the $T^n$-action on $\proj^m$ given by $(t_1,\ldots,t_n)[z_0:\cdots:z_m] =[t^{\veca_0}z_0:\cdots:t^{\veca_m}z_m]$. For $b\in T^n$, let $m_b$ denote the induced action on $\proj^m$, and let $m_b^{\ast}$ denote the induced action on $H^0\bigl(\proj^m,\nbigo(1)\bigr)$. \begin{lem} Suppose that there exists a holomorphic map $\kappa:S\lrarr T^n$ such that $\gamma_1=m_{\kappa}^{\ast}\gamma_2$, i.e., $\gamma_1(x)=m_{\kappa(x)}^{\ast}\gamma_2(x)$ for any $x\in S$. Then, we have the natural isomorphisms $\pi_{+}L_{\star}(\nbiga,S,\gamma_1) \simeq L_{\star}(\nbiga,S,\gamma_2)$ for $\star=\ast,!$. If $\gamma_i$ and $\kappa$ are algebraic, then we have the isomorphisms of algebraic $\nbigd$-modules. \end{lem} \pf We have the isomorphism $\widetilde{m}_{\kappa}: \proj^m\times S\lrarr \proj^m\times S$ induced by $\kappa$ and the identity of $S$. We have an isomorphism $ \widetilde{m}_{\kappa+}L_{\star}(\nbiga,S,\gamma_1) \simeq L_{\star}(\nbiga,S,\gamma_2)$. Applying $\pi_{+}$, we obtain the desired isomorphisms. \hfill\qed \vspace{.1in} The following lemma is clear by construction. \begin{lem} Let $\rho:S\lrarr \cnum$ be a holomorphic map such that $\gamma_1=\gamma_2+\rho\cdot z_0$. Then, we have the isomorphisms of the analytic $\nbigd$-modules $L_{\star}(\nbiga,S,\gamma_1) \simeq L_{\star}(\nbiga,S,\gamma_2)$. In particular, we have the isomorphisms of the analytic $\nbigd$-modules $\pi_{+}^iL_{\star}(\nbiga,S,\gamma_1) \simeq \pi_{+}^iL_{\star}(\nbiga,S,\gamma_2)$. \hfill\qed \end{lem} \subsubsection{Non-degenerate families of sections} Let $p_1:\proj^m\times S\lrarr S$ be the projection. Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo_{\proj^m}(1)\bigr)$ be any holomorphic map. We have the expression $\gamma(x)=\sum_{i=0}^m \gamma_i(x) z_i$. It determines a family of Laurent polynomials $F_{\gamma}(x,t)=\sum_{i=0}^m\gamma_i(x)t^{\veca_i}$. For any face $\sigma$ of $\Conv(\nbiga\cup\{0\})$, we consider the associated family of Laurent polynomials $F_{\gamma,\sigma}(x,t)= \sum_{\veca_i\in\sigma}\gamma_i(x)t^{\veca_i}$. We can regard it as a function on $T^n\times S$. Let $dF_{\gamma,\sigma}$ denote the exterior derivative of $F_{\gamma,\sigma}$, which is a section of $\Omega^1_{T^n\times S}$. We use the following non-degeneracy condition which is a minor generalization of the classical non-degeneracy condition (see also , ) in the sense that we also consider the derivatives in the $S$-direction. \begin{df} $\gamma$ is called non-degenerate at $\infty$ for $X_{\nbiga}$ if we have $(F_{\gamma,\sigma})^{-1}(0)\cap (dF_{\gamma,\sigma})^{-1}(0)=\emptyset$ for any face $\sigma$ of $\Conv(\nbiga\cup\{0\})$ such that $0\not\in\sigma$. \hfill\qed \end{df} The condition is equivalent to that the zero-divisors of $F_{\gamma,\sigma}$ are smooth and reduced for any face $\sigma$ of $\Conv(\nbiga\cup\{0\})$ such that $0\not\in\sigma$. We remark the following. \begin{lem} If $\sigma\subsetneq\Conv(\nbiga\cup\{0\})$ and $0\not\in\sigma$, we have $(F_{\gamma,\sigma})^{-1}(0)\cap (dF_{\gamma,\sigma})^{-1}(0)=\emptyset$ if and only if $(dF_{\gamma,\sigma})^{-1}(0)=\emptyset$. \end{lem} \pf Under the assumption, there exists $\beta=(\beta_1,\ldots,\beta_n)\in(\rnum^n)^{\lor}$ such that $\langle \beta,\veca_i \rangle=1$ for any $\veca_i\in I_{\sigma}$. Because $\Bigl( \sum_{j=1}^n\beta_jt_j\del_{t_j} \Bigr)t^{\veca_i} =\langle \beta,\veca_i \rangle\cdot t^{\veca} =t^{\veca}$, we have $F_{\gamma,\sigma}^{-1}(0) \supset \bigl( dF_{\gamma,\sigma} \bigr)^{-1}(0)$. Hence, the condition $F_{\gamma,\sigma}^{-1}(0)\cap (dF_{\gamma,\sigma})^{-1}(0)=\emptyset$ is equivalent to that $\bigl(dF_{\gamma,\sigma}\bigr)^{-1}(0)=\emptyset$. \hfill\qed \vspace{.1in} Let us reword the condition. We have the section $s_{\gamma}$ of the line bundle $p_1^{\ast}\nbigo_{\proj^m}(1)$ such that $s_{\gamma|\proj^m\times \{x\}} =\gamma(x)\in H^0(\proj^m,\nbigo_{\proj^m}(1))$ for any $x\in S$. We set $H_{\gamma}:=s_{\gamma}^{-1}(0) \subset \proj^m\times S$. We also put $H_{\infty}:=\{z_0=0\}\subset\proj^m$. \begin{lem} Let $\gamma:S\lrarr H^0(\proj^m,\nbigo(1))$ be a morphism. For any face $\sigma$ of $\Conv(\nbiga\cup\{0\})$, the following conditions are equivalent. \begin{itemize} \item $(F_{\gamma,\sigma})^{-1}(0) \cap (dF_{\gamma,\sigma})^{-1}(0)=\emptyset$. \item There exists an open neighbourhood $\nbigu_1$ of $X_{\nbiga}^{\sigma}\times S$ in $\proj^m\times S$ such that (i) $H_{\gamma}\cap \nbigu_1$ is smooth (ii) $H_{\gamma}$ is transversal with $X_{\nbiga}^{\sigma}\times S$. \end{itemize} \end{lem} \pf Let $Q\in X_{\nbiga}^{\sigma}$ be the point $[z^{\sigma}_0:\cdots:z^{\sigma}_m]$ as in \S. By using the $T^n$-action on $X_{\nbiga}^{\sigma}$, we consider the map $\psi:T^n\times S\lrarr X_{\nbiga}^{\sigma}\times S$ given by $\psi(t,x)=(t\cdot Q,x)$. Take any $i_0\in I_{\sigma}$. We have \[ \psi^{\ast}(s_{\gamma})/\psi^{\ast}(z_{i_0}) =\sum_{\veca_i\in I_{\sigma}} \gamma_i(x) t^{\veca_i-\veca_{i_0}} =:G_{\gamma,\sigma}(t,x). \] Then, the second condition holds if and only if (i) $G_{\gamma,\sigma}$ is not constantly $0$, (ii) the divisor $(G_{\gamma,\sigma})=(F_{\gamma,\sigma})$ is smooth. It is equivalent to that $(dF_{\gamma,\sigma})^{-1}(0)\cap F_{\gamma,\sigma}^{-1}(0)=\emptyset$. \hfill\qed \begin{cor} $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$ if and only if the following holds: \begin{itemize} \item There exists a neighbourhood $\nbigu$ of $H_{\infty}\times S$ in $\proj^m\times S$ such that $H_{\gamma}\cap \nbigu$ is a smooth hypersurface in $\nbigu$. \item If $\sigma$ is a face of $\Conv(\nbiga\cup\{0\})$ such that $0\not\in\sigma$, then $H_{\gamma}$ is transversal with $X_{\nbiga}^{\sigma}\times S$. \hfill\qed \end{itemize} \end{cor} The first condition in this corollary is trivial if the image of $\gamma$ does not contain $0\in H^0(\proj^m,\nbigo(1))$. The conditions particularly imply that $\gamma(U)\not\subset \cnum\cdot z_0$ for any open subset $U\subset S$. \vspace{.1in} Let $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$ be any toric desingularization. Let $\varphibar_{\Sigma}:X_{\Sigma}\lrarr \proj^m$ be the composite of $\varphi_{\Sigma}$ and the inclusion $X_{\nbiga}\lrarr \proj^m$. \begin{lem} Let $\gamma:S\lrarr H^0(\proj^m,\nbigo(1))$ be a holomorphic map which is non-degenerate at $\infty$ for $X_{\nbiga}$. Then, $F_{\gamma,\Sigma}$ is non-degenerate along $D_{\Sigma}\times S$. We also have $|(F_{\gamma,\Sigma})_{\infty}|= \varphibar_{\Sigma}^{-1}(H_{\infty})\times S$. \end{lem} \pf Let $T_{\rho}$ be a $T^n$-orbit in $X_{\Sigma}$ contained in $\varphi_{\Sigma}^{-1}(H_{\infty})$. Note that $\varphi_{\Sigma}(T_{\rho})$ is contained in $X_{\Sigma}\cap H_{\infty}$, and the morphism $T_{\rho}\lrarr \varphi_{\Sigma}(T_{\rho})$ is smooth. Let $\varphibar_{\Sigma,S}:=\varphibar_{\Sigma}\times\id_S$. The restriction of $\varphibar_{\Sigma,S}^{\ast}(s_{\gamma})$ to $T_{\rho}\times S$ is not $0$, and the $0$-divisor of $\varphibar_{\Sigma,S}^{\ast}(s_{\gamma})_{|T_{\rho}\times S}$ is smooth. Then, the claims of the lemma are easy to see. \hfill\qed \vspace{.1in} Any element $s\in H^0(\proj^m,\nbigo(1))$ determines a morphism $\gamma_s$ from the one-point set to $H^0(\proj^m,\nbigo(1))$. \begin{df} We say that $s$ is non-degenerate at $\infty$ for $X_{\nbiga}$ if $\gamma_s$ is non-degenerate at $\infty$ for $X_{\nbiga}$ in the sense of Definition {\rm}. \hfill\qed \end{df} It is standard that there exist non-empty Zariski open subsets $U\subset H^0(\proj^m,\nbigo_{\proj^m}(1))$ such that any $s\in U$ is non-degenerate at $\infty$ for $X_{\nbiga}$. Let $U_{\nbiga}^{\reg}$ be the union of such open subsets. \subsubsection{Basic examples of non-degenerate family of sections} The following lemma is clear. \begin{lem} A holomorphic map $\gamma:S\lrarr U_{\nbiga}^{\reg}$ is non-degenerate at $\infty$ for $X_{\nbiga}$. \hfill\qed \end{lem} Note that even if $\gamma:S\lrarr H^0(\proj^m,\nbigo(1))$ is non-degenerate at $\infty$ for $X_{\nbiga}$, the image of $\gamma$ is not necessarily contained in $U_{\nbiga}^{\reg}$. The following lemma is clear. \begin{lem} Let $\gamma:S\lrarr H^0(\proj^m,\nbigo(1))$ be a holomorphic map. Let $\Phi_{\gamma}:H_{\gamma}\lrarr \proj^m$ be the induced morphism, i.e., the composite of the inclusion $H_{\gamma}\lrarr \proj^m\times S$ and the projection $\proj^m\times S\lrarr \proj^m$. Suppose that $H_{\gamma}$ is a smooth hypersurface, and that any critical value of $\Phi_{\gamma}$ is not contained in $X_{\nbiga}\cap H_{\infty}$. Then, $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$. \hfill\qed \end{lem} \begin{example} We set $W:=\bigl\{ \sum_{i=1}^m\alpha_iz_i \bigr\}\subset H^0(\proj^m,\nbigo(1))$. Then, the inclusion $\iota:W\lrarr H^0(\proj^m,\nbigo(1))$ is non-degenerate at $\infty$ for $X_{\nbiga}$ for any $\nbiga$. Indeed, we set $\Htilde_{\iota}\subset \bigl( \cnum^{m+1}\setminus\{(0,\ldots,0)\} \bigr)\times W:= \Bigl\{(z_0,\ldots,z_m;\alpha_1,\ldots,\alpha_m)\,\Big|\, \sum_{i=1}^m\alpha_iz_i=0 \Bigr\}$. Then, $\Htilde_{\iota}$ is clearly smooth. Because $H_{\iota}$ is the quotient of $\Htilde_{\iota}$ by the natural free $\cnum^{\ast}$-action, $H_{\iota}$ is also smooth. Let $\Htilde_{\iota}\lrarr \cnum^{m+1}\setminus\{(0,\ldots,0)\}$ be the projection. Then, the critical values are $\{(z_0,0,\ldots,0)\,|\,z_0\in\cnum^{\ast}\}$. Hence, the set of the critical values of $\Phi_{\iota}:H_{\iota}\lrarr \proj^m$ is $\{[1:0:\cdots:0]\}$. Then, by Lemma {\rm}, we obtain that $\iota$ is non-degenerate. \hfill\qed \end{example} Set $W^{\ast}:=\bigl\{ \sum_{i=1}^{m}\alpha_iz_i\,\big|\, \alpha_i\in\cnum^{\ast} \bigr\}$. We have the action of $T^n$ on $W^{\ast}$ by $t\cdot (\alpha_1,\ldots,\alpha_m) =(t^{\veca_1}\alpha_1,\ldots,t^{\veca_m}\alpha_m)$. The action is free because $\nbiga$ generates $\seisuu^n$. Let $q:W^{\ast}\lrarr W^{\ast}/T^n$ be the projection. \begin{lem} Let $\gamma:S\lrarr W^{\ast}$ be any holomorphic map such that $q\circ\gamma$ is submersive. Then, $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$. \end{lem} \pf Let $\sigma$ be any face of $\Conv(\nbiga\cup\{0\})$ such that $0\not\in\sigma$. Take any $(P,x)\in H_{\gamma}\subset \proj^m\times S$ such that $P\in X_{\nbiga}^{\sigma}$. We have $P\in H_{\gamma(x)}$. Let $\Delta_{\epsilon}:=\bigl\{z\in\cnum\,\big|\,|z|<\epsilon\bigr\}$ for $\epsilon>0$. For any holomorphic map $g:\Delta_{\epsilon}\lrarr W^{\ast}$, we have the family of hyperplanes $H_g\subset\proj^m\times \Delta_{\epsilon}$, and we obtain the induced map $\Phi_g:H_{g}\lrarr \proj^m$. Because $P\neq [1:0:\cdots:0]$, we can find a holomorphic map $g_0:\Delta\lrarr W^{\ast}$ such that (i) $g_0(0)=\gamma(x)$, (ii) the tangent map of $\Phi_{g_0}$ at $(P,0)$ is an isomorphism. For some $0<\epsilon'<\epsilon$, we can find holomorphic maps $g_1:\Delta_{\epsilon'}\lrarr S$ and $h_1:\Delta_{\epsilon'}\lrarr T^n$ such that (i) $g_1(0)=x$, (ii) $h_1(0)=1$, (iii) $\gamma\circ g_1(x)=h_1(x)\cdot g_0(x)$ for $x\in \Delta_{\epsilon'}$. Then, it is easy to see that the tangent map of $\Phi_{\gamma}$ at $(P,x)$ is surjective. \hfill\qed \begin{example} Let $r:W^{\ast}/T^n\lrarr W^{\ast}$ be any section of $q$. Then, the induced morphism $r:W^{\ast}/T^n\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ is non-degenerate. \hfill\qed \end{example} \subsubsection{$\nbigd$-modules associated to non-degenerate families of sections} Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ be any holomorphic map. Suppose that $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$. We have the $\nbigd$-modules $L_{\star}(\nbiga,S,\gamma)$ on $\proj^m\times S$. We set $H_{\infty,S}:=H_{\infty}\times S$. \begin{lem} We naturally have $L_{\star}(\nbiga,S,\gamma)(\ast H_{\infty,S}) \simeq L_{\star}(\nbiga,S,\gamma)$ $(\star=\ast,!)$. \end{lem} \pf For a desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$, we obtain the $\nbigd$-modules $L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$. Because $F_{\gamma,\Sigma}$ is non-degenerate along $D_{\Sigma,S}$, we have \[ L_{\star}(\nbiga,S,\gamma)(\ast H_{\infty,S}) \simeq \varphibar^0_{\Sigma,S+}\Bigl( L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \bigl(\ast (F_{\gamma,\Sigma})_{\infty}\bigr) \Bigr) \simeq \varphibar^0_{\Sigma,S+}\bigl( L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \bigr) \simeq L_{\star}(\nbiga,S,\gamma). \] Thus, we are done. \hfill\qed \vspace{.1in} Let $K_{\nbiga,S,\gamma}$ and $C_{\nbiga,S,\gamma}$ denote the kernel and the cokernel of $L_{!}(\nbiga,S,\gamma) \lrarr L_{\ast}(\nbiga,S,\gamma)$. \begin{cor} We have $M(\ast H_{\infty,S}) \simeq M$ for the $\nbigd$-modules $M=K_{\nbiga,S,\gamma},C_{\nbiga,S,\gamma}, L_{\min}(\nbiga,S,\gamma)$. \hfill\qed \end{cor} \begin{prop} We have $\pi_+^iM=0$ $(i\neq 0)$ for the $\nbigd$-modules \[ M=L_{\star}(\nbiga,S,\gamma)\,\,(\star=\ast,!), L_{\min}(\nbiga,S,\gamma), K_{\nbiga,S,\gamma}, C_{\nbiga,S,\gamma}. \] \end{prop} \pf Let us prove $\pi_+^iL_{\star}(\nbiga,S,\gamma)=0$ $(i\neq 0)$. By using the duality, it is enough to prove that $\pi_{+}^iL_{\star}(\nbiga,S,\gamma)=0$ for $i>0$ and for $\ast,!$. By Lemma , we have $R^j\pi_{\ast}\bigl( L_{\star}(\nbiga,S,\gamma) \otimes\Omega^i_{\proj^m} \bigr)=0$ for any $j>0$ and any $i$. By using the expression \[ \pi^j_{+}L_{\star}(\nbiga,S,\gamma) \simeq \hyperr^j\pi_{\Sigma\ast}\bigl( \Omega^{\bullet+m}_{\proj^m\times S/S} \otimes L_{\star}(\nbiga,S,\gamma) \bigr), \] we obtain the claim. Let us prove the vanishings for $K_{\nbiga,S,\gamma}$, $C_{\nbiga,S,\gamma}$ and $L_{\min}(\nbiga,S,\gamma)$. By Corollary , we obtain $\pi^i_{+}K_{\nbiga,S,\gamma}=0$, $\pi^i_{+}C_{\nbiga,S,\gamma}=0$ and $\pi_{+}^iL_{\min}(\nbiga,S,\gamma)=0$ for $i>0$. Let $-\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ denote the composite of $\gamma$ and the multiplication of $-1$ on $H^0\bigl(\proj^m,\nbigo(1)\bigr)$. We naturally have $\DDD K_{\nbiga,S,\gamma}\simeq C_{\nbiga,S,-\gamma}$, $\DDD C_{\nbiga,S,\gamma}\simeq K_{\nbiga,S,-\gamma}$, and $\DDD L_{\min}(\nbiga,S,\gamma) \simeq L_{\min}(\nbiga,S,-\gamma)$. Then, we obtain $\pi^i_{+}K_{\nbiga,S,\gamma}=0$, $\pi^i_{+}C_{\nbiga,S,\gamma}=0$ and $\pi_{+}^iL_{\min}(\nbiga,S,\gamma)=0$ for $i<0$. \hfill\qed \begin{cor} The following are exact sequences: \[ 0\lrarr \pi_{+}^0K_{\nbiga,S,\gamma} \lrarr \pi_+^0L_!(\nbiga,S,\gamma) \lrarr \pi_+^0L_{\min}(\nbiga,S,\gamma) \lrarr 0 \] \[ 0\lrarr \pi_+^0L_{\min}(\nbiga,S,\gamma) \lrarr \pi_+^0L_{\ast}(\nbiga,S,\gamma) \lrarr \pi_{+}^0C_{\nbiga,S,\gamma} \lrarr 0 \] In particular, $\pi_+^0L_{\min}(\nbiga,S,\gamma)$ is isomorphic to the image of $\pi_+^0L_!(\nbiga,S,\gamma) \lrarr \pi_+^0L_{\ast}(\nbiga,S,\gamma)$. \hfill\qed \end{cor} \begin{cor} For any desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$, we have $\pi_{\Sigma+}^i L_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) =0$ for $i\neq 0$. \hfill\qed \end{cor} \begin{cor} Suppose that $\gamma$ satisfies the stronger condition that $\gamma(S)\subset U_{\nbiga}^{\reg}$. Then, the $\nbigd$-modules $\pi_+^0L_{\star}(\nbiga,S,\gamma)$, $\pi_+^0L_{\min}(\nbiga,S,\gamma)$, $\pi_+^0K_{\nbiga,S,\gamma}$ and $\pi_+^0C_{\nbiga,S,\gamma}$ are locally free $\nbigo_S$-modules with a flat connection. \end{cor} \pf We obtain the claims for $\pi_+^0L_{\star}(\nbiga,S,\gamma)$ from Corollary and the isomorphisms (). Then, the claims for the others follows. \hfill\qed \vspace{.1in} Suppose moreover that $0$ is an interior point of $\Conv(\nbiga)$. In this case, we have $L_{!}(\nbiga,S,\gamma) \simeq L_{\min}(\nbiga,S,\gamma) \simeq L_{\ast}(\nbiga,S,\gamma)$, which is denoted by $L(\nbiga,S,\gamma)$. We have $\pi_+^iL(\nbiga,S,\gamma)=0$ for $i\neq 0$. For a desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$, we have the $\nbigd$-module $L(F_{\Sigma,\gamma}) =L_{!}(F_{\gamma,\Sigma},D_{\Sigma,S}) =L_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S})$. We have the following corollary as a special case. \begin{cor} If $0$ is an interior point of $\Conv(\nbiga)$, we have $\pi^i_{\Sigma+}L(F_{\gamma,\Sigma}) =0$ for $i\neq 0$. We naturally have $\pi^0_{\Sigma+}L(F_{\gamma,\Sigma}) \simeq \pi^0_{+}L(\nbiga,S,\gamma)$. \hfill\qed \end{cor} \begin{example} The $\nbigd$-modules associated to the non-degenerate section in Example {\rm} are called the GKZ-hypergeometric systems. The $\nbigd$-modules associated to the non-degenerate sections in Example {\rm} are called the reduced GKZ-hypergeometric systems. Note that it is independent of the choice of a section, according to Lemma {\rm }. \hfill\qed \end{example} \subsubsection{De Rham complexes} Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ be a holomorphic map which is non-degenerate at $\infty$ for $X_{\nbiga}$. Take a toric desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$. \begin{prop} If $0$ is contained in the interior part of $\Conv(\nbiga)$, we have the following natural isomorphism for $L(F_{\gamma,\Sigma})= L_{!}(F_{\gamma,\Sigma},D_{\Sigma,S}) =L_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S})$: \begin{equation} \pi^0_{\Sigma+}\bigl( L(F_{\gamma,\Sigma}) \bigr) \simeq \hyperr^n\pi_{\Sigma\ast}\bigl( \Omega^{\bullet}_{X_{\Sigma,S}/S}(\ast D_{\Sigma,S}), d+dF_{\gamma,\Sigma} \bigr) \end{equation} If $\gamma$ is algebraic, it means $\pi_{+}^0\bigl( L(\nbiga,S,\gamma) \bigr) \simeq \hyperr^n\pi_{\ast}\bigl( \Omega^{\bullet}_{T^n\times S/S}, d+dF_{\gamma} \bigr)$. If $0$ is a boundary point of $\Conv(\nbiga\cup\{0\})$, we have the following isomorphisms: \begin{equation} \pi^0_{\Sigma+} L_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \hyperr^n\pi_{\Sigma\ast}\Bigl( \Omega_{X_{\Sigma,S}/S}^{\bullet} (\log D_{\Sigma,S})(\ast (F_{\gamma,\Sigma})_{\infty}), d+dF_{\gamma,\Sigma} \Bigr) \end{equation} \begin{equation} \pi^0_{\Sigma+} L_!(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \hyperr^n\pi_{\Sigma\ast}\Bigl( \Omega_{X_{\Sigma,S}/S}^{\bullet} (\log D_{\Sigma,S}) (-D_{\Sigma,S}) (\ast (F_{\gamma,\Sigma})_{\infty}), d+dF_{\gamma,\Sigma} \Bigr) \end{equation} \end{prop} \pf We immediately () from $L(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \bigl( \nbigo_{X_{\Sigma,S}}(\ast D_{\Sigma,S}), d+dF_{\gamma,\Sigma} \bigr)$. By applying the arguments in \S and \S, we obtain () and (). \hfill\qed \vspace{.1in} Let $\check{\varphi}_{\nbiga}: \check{X}_{\nbiga}\lrarr X_{\nbiga}$ denote the normalization of $X_{\nbiga}$, which is the normal toric variety. Let us rewrite () and () in terms of $\check{X}_{\nbiga}$. We put $\check{D}_{\nbiga}:= \check{X}_{\nbiga}\setminus T^n$. We also set $\check{D}_{\nbiga,\infty}:= \check{\varphi}_{\nbiga}^{-1}(X_{\nbiga}\cap H_{\infty})$. Let $\Omega^i_{\check{X}_{\nbiga}}(\log \check{D}_{\nbiga})$ denote the sheaf of logarithmic $i$-forms $(\check{X}_{\nbiga},\check{D}_{\nbiga})$ as in . Let $\Omega^i_{\check{X}_{\nbiga},\check{D}_{\nbiga}}$ denote the sheaf of $i$-forms on $X_{\nbiga}$ whose restrictions to each stratum of $\check{D}_{\nbiga}$ is $0$ as in . Let $\Omega_{\check{X}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}_{\nbiga,S})$ and $\Omega^{\bullet}_{(\check{X}_{\nbiga},\check{D}_{\nbiga})\times S/S}$ denote the pull back of $\Omega_{\check{X}_{\nbiga}}^{\bullet} (\log \check{D}_{\nbiga})$ and $\Omega^{\bullet}_{(\check{X}_{\nbiga},\check{D}_{\nbiga})}$ by the projection $\check{X}_{\nbiga}\times S\lrarr \check{X}_{\nbiga}$, respectively. \begin{prop} We have the following isomorphisms: \begin{equation} \pi^0_{+}L_{\ast}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omega_{\check{X}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}_{\nbiga,S})(\ast \check{D}_{\nbiga,\infty,S}), d+dF_{\gamma} \Bigr) \end{equation} \begin{equation} \pi^0_{+}L_{!}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omega_{(\check{X}_{\nbiga},\check{D}_{\nbiga})\times S/S}^{\bullet} (\ast \check{D}_{\nbiga,\infty,S}), d+dF_{\gamma} \Bigr) \end{equation} \end{prop} \pf According to \cite[Lemma 6.1]{Batyrev-VMHS}, we have \begin{equation} R\varphi_{\Sigma\ast}\Omega^i_{X_{\Sigma}}(\log D_{\Sigma}) \simeq \Omega^i_{\check{X}_{\nbiga}}(\log \check{D}_{\nbiga}). \end{equation} For a smooth toric variety $X_{\Sigma}$, we have $\Omega^i_{X_{\Sigma},D_{\Sigma}}\simeq \Omega^i_{X_{\Sigma}}(\log D_{\Sigma})(-D_{\Sigma})$. According to \cite[Proposition 1.8]{Danilov-de-Rham}, we have \begin{equation} R\varphi_{\Sigma\ast}\Omega^i_{X_{\Sigma}}(\log D_{\Sigma})(-D_{\Sigma}) \simeq \Omega^i_{\check{X}_{\nbiga},\check{D}_{\nbiga}}. \end{equation} Then, we obtain () and () from () and (). \hfill\qed \vspace{.1in} In the algebraic setting, we set $X^{\aff}_{\nbiga}:=X_{\nbiga}\setminus H_{\infty}$. Let $\check{\varphi}^{\aff}:\check{X}^{\aff}\lrarr X^{\aff}$ denote the normalization. Let $\check{D}^{\aff}_{\nbiga}:= \check{X}^{\aff}_{\nbiga}\setminus T^n$. We obtain the following similarly. \begin{prop} We have the following isomorphisms: \begin{equation} \pi^0_{+}L_{\ast}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omega_{\check{X}^{\aff}_{\nbiga,S}/S}^{\bullet} (\log \check{D}^{\aff}_{\nbiga,S}), d+dF_{\gamma} \Bigr) \end{equation} \begin{equation} \pi^0_{+}L_{!}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omega_{(\check{X}^{\aff}_{\nbiga,S}, \check{D}^{\aff}_{\nbiga,S})/S}^{\bullet}, d+dF_{\gamma} \Bigr) \end{equation} \hfill\qed \end{prop} \begin{rem} In the case of Example {\rm}, by using Proposition {\rm}, we can describe {\rm()} and {\rm()} as the $\nbigd$-modules associated to better behaved GKZ-systems in {\rm}. \end{rem} \subsection{Mixed twistor $\nbigd$-modules} \subsubsection{Mixed twistor $\nbigd$-modules associated to families of Laurent polynomials} We use the notation in \S. Let $\gamma:S\lrarr H^0(\proj^m,\nbigo(1))$ be a holomorphic map. We take a toric desingularization $\varphi_{\Sigma}: X_{\Sigma} \lrarr X_{\nbiga}$. Let $d_S:=\dim S$. We have the integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$ $(\star=\ast,!)$ with real structure on $X_{\Sigma}\times S$: \[ \nbigt_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S}) =\Bigl( \lambda^{n+d_S}\nbigl_{!}(F_{\gamma,\Sigma},D_{\Sigma,S}), \nbigl_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S}), \nbigc_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S}) \Bigr). \] \[ \nbigt_{!}(F_{\gamma,\Sigma},D_{\Sigma,S}) =\Bigl( \lambda^{n+d_S} \nbigl_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S}), \nbigl_{!}(F_{\gamma,\Sigma},D_{\Sigma,S}), \nbigc_{!}(F_{\gamma,\Sigma},D_{\Sigma,S}) \Bigr). \] The weight filtration is denoted by $W$. The image of $\nbigt_!(\nbiga,S,\gamma) \lrarr \nbigt_{\ast}(\nbiga,S,\gamma)$ is denoted by $\nbigt_{\min}(\nbiga,S,\gamma)$. On $\proj^m\times S$, we obtain the $\nbigrtilde_{\proj^m\times S}$-modules $\nbigl_{\star}(\nbiga,S,\gamma):= \varphi^0_{\Sigma,S+} \nbigl_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$ and the integrable mixed twistor $\nbigd$-modules with induced real structure $\nbigt_{\star}(\nbiga,S,\gamma):= \varphibar^0_{\Sigma,S+} \nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$: \[ \nbigt_{\ast}(\nbiga,S,\gamma) =\Bigl( \lambda^{n+d_S}\nbigl_{!}(\nbiga,S,\gamma), \nbigl_{\ast}(\nbiga,S,\gamma), \nbigc_{\ast}(\nbiga,S,\gamma) \Bigr) \] \[ \nbigt_{!}(\nbiga,S,\gamma) =\Bigl( \lambda^{n+d_S}\nbigl_{\ast}(F_{\gamma},D_{\nbiga,S}), \nbigl_{!}(F_{\gamma},D_{\nbiga,S}), \nbigc_!(\nbiga,S,\gamma) \Bigr) \] Here, $\nbigc_{\star}(\nbiga,S,\gamma)$ $(\star=\ast,!)$ are obtained as the push-forward of $\nbigc_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$. They are twistor enhancement of $L_{\star}(\nbiga,S,\gamma)$. As in the case of $\nbigd$-modules (\S), they are independent of the choice of a toric desingularization. We also obtain integrable mixed twistor $\nbigd$-modules with real structure $\pi^i_{\dagger}\bigl( \nbigt_{\star}(\nbiga,S,\gamma) \bigr)$ $(\star=\ast,!)$ on $S$ which are naturally isomorphic to $\pi^i_{\Sigma\dagger}\bigl( \nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \bigr)$. \begin{rem} Let $X_{\Sigma_1}$ be any $n$-dimensional smooth toric variety with a fixed inclusion $T^n\subset X_{\Sigma_1}$. We have the meromorphic function $F_{\gamma,\Sigma_1}$ on $(X_{\Sigma_1,S},D_{\Sigma_1,S})$ associated to the family of Laurent polynomials $F_{\gamma}$. We have the associated integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(F_{\gamma,\Sigma_1},D_{\Sigma_1,S})$. As in the case of $\nbigd$-modules, we naturally have $\pi^i_{\Sigma_1\dagger} \nbigt_{\star}(F_{\gamma,\Sigma_1},D_{\Sigma_1,S}) \simeq \pi^i_{\dagger} \nbigt_{\star}(\nbiga,S,\gamma)$. \hfill\qed \end{rem} \subsubsection{Non-degenerate family of sections} Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo_{\proj^m}(1)\bigr)$ be a holomorphic map which is non-degenerate at $\infty$ for $X_{\nbiga}$. \begin{prop} We naturally have $\nbigl_{\star}(\nbiga,S,\gamma) (\ast H_{\infty,S}) \simeq \nbigl_{\star}(\nbiga,S,\gamma)$. \end{prop} \pf We obtain the claim by using the argument in Lemma together with Corollary and Proposition . \hfill\qed \vspace{.1in} Let $\nbigk_{\nbiga,S,\gamma}$ and $\nbigc_{\nbiga,S,\gamma}$ denote the kernel and the cokernel of $\nbigt_!(\nbiga,S,\gamma) \lrarr \nbigt_{\ast}(\nbiga,S,\gamma)$. \begin{cor} Let $\nbigm$ be the $\nbigrtilde_{\proj^m\times S}$-modules underlying $\nbigk_{\nbiga,S,\gamma}$, $\nbigc_{\nbiga,S,\gamma}$ or $\nbigt_{\min}(\nbiga,S,\gamma)$. Then, we have $\nbigm(\ast H_{\infty,S})\simeq\nbigm$. \hfill\qed \end{cor} We obtain the following from Proposition . \begin{prop} We have $\pi^i_+\nbigt=0$ $(i\neq 0)$ for the integrable mixed twistor $\nbigd$-modules for $\nbigt=\nbigt_{\star}(\nbiga,S,\gamma)$ $(\star=\ast,!)$ $\nbigt_{\min}(\nbiga,S,\gamma)$, $\nbigk_{\nbiga,S,\gamma}$, and $\nbigc_{\nbiga,S,\gamma}$. \hfill\qed \end{prop} We obtain the following corollary. \begin{cor} We have the following exact sequences of integrable mixed twistor $\nbigd$-modules on $S$: \[ 0\lrarr \pi_+^0\nbigk_{\nbiga,S,\gamma} \lrarr \pi_+^0\nbigt_!(\nbiga,S,\gamma) \lrarr \pi_+^0\nbigt_{\min}(\nbiga,S,\gamma) \lrarr 0 \] \[ 0\lrarr \pi_+^0\nbigt_{\min}(\nbiga,S,\gamma) \lrarr \pi_+^0\nbigt_{\ast}(\nbiga,S,\gamma) \lrarr \pi_+^0\nbigc_{\nbiga,S,\gamma} \lrarr 0 \] As a result, the image of the morphism $\pi_+^0\nbigt_!(\nbiga,S,\gamma) \lrarr \pi_+^0\nbigt_{\ast}(\nbiga,S,\gamma)$ is naturally isomorphic to the integrable pure twistor $\nbigd$-module $\pi_+^0\nbigt_{\min}(\nbiga,S,\gamma)$. Moreover, we have \[ \Gr^W_{n+d_S} \pi_{\dagger}^0\nbigt_{!}(\nbiga,S,\gamma) \simeq \pi_{\dagger}^0\nbigt_{\min}(\nbiga,S,\gamma) \simeq \Gr^W_{n+d_S} \pi_{\dagger}^0\nbigt_{\ast}(\nbiga,S,\gamma). \] \hfill\qed \end{cor} \begin{cor} For any desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$, we have $\pi^i_{\Sigma\dagger}\bigl( \nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \bigr)=0$ $(i\neq 0)$. \hfill\qed \end{cor} \subsubsection{Descriptions in terms of the de Rham complexes} Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo_{\proj^m}(1)\bigr)$ be a morphism which is non-degenerate at $\infty$ for $X_{\nbiga}$. We give descriptions of the $\nbigrtilde$-modules $\pi_{\dagger}^0\nbigl_{\star}(\nbiga,S,\gamma)$. We use the notation in \S and \S. \begin{prop} Take a toric desingularization $\gamma:X_{\Sigma}\lrarr X_{\nbiga}$. If $0$ is an interior point of $\Conv(\nbiga)$, we have \[ \pi^0_{\Sigma\dagger}\bigl( \nbigl(F_{\gamma,\Sigma}) \bigr) \simeq \hyperr^n\pi_{\Sigma\ast}\bigl( \Omegabar^{\bullet}_{X_{\Sigma,S}/S}(\ast D_{\Sigma,S}), d+\lambda^{-1}dF_{\gamma,\Sigma} \bigr). \] If moreover $\gamma$ is algebraic, it means $\pi_{\dagger}^0\bigl( \nbigl(F_{\gamma}) \bigr) \simeq \hyperr^n\pi_{\ast}\bigl( \Omegabar^{\bullet}_{T^n\times S/S}, d+\lambda^{-1}dF_{\gamma} \bigr)$. If $0$ is a boundary point of $\Conv(\nbiga\cup\{0\})$, we have the following natural isomorphisms: \begin{equation} \pi^0_{\Sigma\dagger} \nbigl_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \hyperr^n\pi_{\Sigma\ast}\Bigl( \Omegabar_{X_{\Sigma,S}/S}^{\bullet} (\log D_{\Sigma,S})(\ast (F_{\gamma,\Sigma})_{\infty}), d+\lambda^{-1}dF_{\gamma,\Sigma} \Bigr) \end{equation} \begin{equation} \pi^0_{\Sigma\dagger} \nbigl_!(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \hyperr^n\pi_{\Sigma\ast}\Bigl( \Omegabar_{X_{\Sigma,S}/S}^{\bullet} (\log D_{\Sigma,S}) (-D_{\Sigma,S}) (\ast (F_{\gamma,\Sigma})_{\infty}), d+\lambda^{-1}dF_{\gamma,\Sigma} \Bigr) \end{equation} \end{prop} \pf They are the special cases of the isomorphisms given in \S. \hfill\qed \vspace{.1in} Let $q:\cnum_{\lambda}\times \check{X}_{\nbiga}\times S \lrarr \check{X}_{\nbiga}$ be the projection. Let $\Omegabar^i_{\check{X}_{\nbiga,S}/S}(\log \check{D}_{\nbiga,S})$ denote $\lambda^{-i}q^{\ast} \Omega^i_{\check{X}_{\nbiga}}(\log \check{D}_{\nbiga})$. Let $\Omegabar^i_{\check{X}_{\nbiga},\check{D}_{\nbiga}}$ denote $\lambda^{-i}q^{\ast} \Omega^i_{\check{X}_{\nbiga},\check{D}_{\nbiga}}$. \begin{prop} We have the following natural isomorphisms: \begin{equation} \pi^0_{\dagger}\nbigl_{\ast}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{\check{X}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}_{\nbiga,S})(\ast \check{D}_{\nbiga,\infty,S}), d+\lambda^{-1}dF_{\gamma} \Bigr) \end{equation} \begin{equation} \pi^0_{\dagger}\nbigl_{!}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{(\check{X}_{\nbiga},\check{D}_{\nbiga})\times S/S}^{\bullet} (\ast \check{D}_{\nbiga,\infty,S}), d+\lambda^{-1}dF_{\gamma} \Bigr) \end{equation} In the algebraic setting, we have the following isomorphisms: \begin{equation} \pi^0_{\dagger}\nbigl_{\ast}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{\check{X}^{\aff}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}^{\aff}_{\nbiga,S}), d+\lambda^{-1}dF_{\gamma} \Bigr) \end{equation} \begin{equation} \pi^0_{\dagger}\nbigl_{!}(\nbiga,S,\gamma) \simeq \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{(\check{X}^{\aff}_{\nbiga}, \check{D}^{\aff}_{\nbiga})\times S/S}^{\bullet}, d+\lambda^{-1}dF_{\gamma} \Bigr) \end{equation} \end{prop} \pf We obtain the isomorphisms from Proposition with the isomorphisms () and (). \hfill\qed \begin{cor} The image of the natural morphism \begin{equation} \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{(\check{X}^{\aff}_{\nbiga}, \check{D}^{\aff}_{\nbiga})\times S/S}^{\bullet}, d+\lambda^{-1}dF_{\gamma} \Bigr) \lrarr \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{\check{X}^{\aff}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}^{\aff}_{\nbiga,S}), d+\lambda^{-1}dF_{\gamma} \Bigr) \end{equation} underlies a pure twistor $\nbigd$-module. \hfill\qed \end{cor} Let $U_{\nbiga}^{\reg}\subset H^0(\proj^m,\nbigo(1))$ be the open subset as in \S. Suppose that the image of $\gamma$ is contained in $U_{\nbiga}^{\reg}$. In this case, $\pi^0_{\dagger}\bigl( \nbigl_{\star}(\nbiga,S,\gamma) \bigr)$ are locally free $\nbigo_{\cnum_{\lambda}\times S}$-modules. Let $\pi^0_{\dagger}\bigl( \nbigl_{\star}(\nbiga,S,\gamma) \bigr)^{\lor}$ be the dual as $\nbigo_{\cnum_{\lambda}\times S}$-modules, which are naturally equipped with the meromorphic connection. They are naturally isomorphic to $\lambda^{-d_S}\DDD\pi^0_{\dagger}\bigl( \nbigl_{\star}(\nbiga,S,\gamma) \bigr)$ up to signatures. The real structure of $\pi_{\dagger}\nbigt_{!}(\nbiga,S,\gamma)$ gives \[ \pi_{\dagger}^0 \nbigl_{!}(\nbiga,S,\gamma) \simeq j^{\ast}\DDD\pi^0_{\dagger} \bigl( \lambda^{n+d_S} \nbigl_{\ast}(\nbiga,S,\gamma) \bigr) \simeq \lambda^{-n-d_S} j^{\ast}\DDD\pi^0_{\dagger} \nbigl_{\ast}(\nbiga,S,\gamma) \simeq \lambda^{-n} \Bigl( j^{\ast}\pi_{\dagger}^0 \nbigl_{\ast}(\nbiga,S,\gamma) \Bigr)^{\lor} \] \begin{cor} $\pi_{\dagger}^0 \nbigl_{!}(\nbiga,S,\gamma)$ is naturally identified with $\lambda^{-n}j^{\ast}\hyperr^n\pi_{\ast}\Bigl( \Omegabar_{\check{X}^{\aff}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}^{\aff}_{\nbiga,S}), d+\lambda^{-1}dF_{\gamma} \Bigr)^{\lor}$. The image of the natural morphism \[ \lambda^{-n} j^{\ast}\hyperr^n\pi_{\ast}\Bigl( \Omegabar_{\check{X}^{\aff}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}^{\aff}_{\nbiga,S}), d+\lambda^{-1}dF_{\gamma} \Bigr)^{\lor} \lrarr \hyperr^n\pi_{\ast}\Bigl( \Omegabar_{\check{X}^{\aff}_{\nbiga}\times S/S}^{\bullet} (\log \check{D}^{\aff}_{\nbiga,S}), d+\lambda^{-1}dF_{\gamma} \Bigr) \] underlies an integrable pure twistor $\nbigd$-module. \hfill\qed \end{cor} \begin{rem} If $S$ and $\gamma$ are as in Example {\rm}, we can describe {\rm()} and {\rm()} in terms of systems of differential equations as in {\rm\S}. \hfill\qed \end{rem} \subsubsection{Graded polarizations} Let $\nbiga=\{\veca_1,\ldots,\veca_m\}$ and $\veca_0:=(0,\ldots,0)$. Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ be a morphism which is non-degenerate at $\infty$ for $X_{\nbiga}$. We obtain the family of Laurent polynomials \[ F_{\gamma}(x,t) =\sum_{i=0}^m\gamma_i(x)t^{\veca_i} \] Take $\vecb\in\seisuu^{n}$ such that $0$ is contained in the interior part of $\Conv(\nbiga\cup\{\vecb\})$. It gives a monomial $t^{\vecb}$. We shall observe that the mixed twistor $\nbigd$-modules $\pi_{\dagger}^0\nbigt(\nbiga,S,\gamma)$ are equipped with graded sesqui-linear dualities depending on the choice of $\vecb$. Let $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$ be a toric desingularization. We have the meromorphic function $t^{\vecb}_{\Sigma}$ on $(X_{\Sigma},D_{\Sigma})$ induced by $t^{\vecb}$. We also have the meromorphic functions $F_{\gamma,\Sigma}$ and $t_{\Sigma,S}^{\vecb}$ on $(X_{\Sigma,S},D_{\Sigma,S})$ induced by $F_{\gamma}$ and $t^{\vecb}$. \begin{lem} Suppose that $|(t_{\Sigma}^{\vecb})_0| \cap |(t_{\Sigma}^{\vecb})_{\infty}| =\emptyset$. Then, the other conditions in Lemma {\rm} are satisfied for $g=F_{\gamma,\Sigma}$ and $f=t^{\vecb}_{\Sigma,S}$. \end{lem} \pf By the non-degeneracy assumption on $\gamma$, $F_{\gamma,\Sigma}$ is non-degenerate along $D_{\Sigma,S}$. We clearly have $|(t_{\Sigma,S}^{\vecb})_0|\cup |(t_{\Sigma,S}^{\vecb})_{\infty}| \subset D_{\Sigma,S}$. It is enough to prove that $D_{\Sigma,S}=\bigl|(t_{\Sigma,S}^{\vecb})_{\infty}\bigr| \cup\bigl|(F_{\gamma,\Sigma})_{\infty}\bigr|$. We set $\veca_{m+1}:=\vecb$ and $\nbigatilde:=\nbiga\cup\{\veca_{m+1}\}$. We set $\Stilde:=S\times\cnum_{\tau}$. We can naturally regard $H^0\bigl(\proj^{m+1},\nbigo(1)\bigr) =H^0\bigl(\proj^m,\nbigo(1)\bigr) \times \bigl( \cnum\cdot z_{m+1} \bigr)$. We consider the morphism $\gammatilde:=\gamma\times\id: \Stilde\lrarr H^0\bigl(\proj^{m+1},\nbigo(1)\bigr)$. It is enough to prove that $F_{\gammatilde}(x,\tau,t)= F_{\gamma}(x,t)+\tau t_{\Sigma,S}^{\vecb}$ is non-degenerate at $\infty$ for $X_{\nbigatilde}$. Let $\sigma$ be any face of $\Conv(\nbigatilde)$. If $\vecb\not\in\sigma$, then $\sigma$ is a face of $\Conv(\nbiga\cup\{0\})$ such that $0\not\in\sigma$. Hence, we have $(dF_{\gammatilde,\sigma})^{-1}(0) =(dF_{\gamma,\sigma})^{-1}(0) =\emptyset$ by applying Lemma to $\nbiga$. If $\vecb\in\sigma$, then $\del_{\tau}F_{\gammatilde,\sigma} =t_{\Sigma,S}^{\vecb}$ is nowhere vanishing. Thus, we obtain the claim of the lemma. \hfill\qed \vspace{.1in} We have the canonical polarization $\bigl( (-1)^{n+d_S},(-1)^{n+d_S} \bigr)$ on the pure twistor $\nbigd$-module $\nbigt(F_{\gamma,\Sigma},D_{\Sigma,S})$ of weight $n+d_S$ on $X_{\Sigma,S}$, which we denote by $\nbigs_{\gamma,\Sigma}$. By Lemma , we have \[ \nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) =\nbigt(F_{\gamma,\Sigma},D_{\Sigma,S}) [\star (t_{\Sigma,S}^{\vecb})_{\infty}]. \] As explained in \S, we have the induced graded polarization $\nbigs_{\gamma,\Sigma}[\star(t^{\vecb}_{\Sigma,S})_{\infty}]$ on $\nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$, which we denote by $\nbigs_{\gamma,\Sigma,\star,\vecb}$. By Corollary , Condition {\bf(A)} holds for the projection $\pi_{\Sigma}$ and the mixed twistor $\nbigd$-modules $\nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S})$ with $\nbigs_{\gamma,\Sigma,\star,\vecb}$, and we obtain the graded polarization $\bigl[ \pi_{\Sigma\dagger}^0 \nbigs_{\gamma,\Sigma,\star,\vecb} \bigr]$ on $\pi_{\Sigma\dagger}^0 \nbigt_{\star}(F_{\gamma,\Sigma},D_{\Sigma,S}) \simeq \pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$. We obtain the following lemma from Corollary . \begin{lem} The graded polarizations $\bigl[ \pi_{\Sigma\dagger}^0 \nbigs_{\gamma,\Sigma,\star,\vecb} \bigr]$ on $\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$ are independent of the choice of a toric desingularization $\varphi_{\Sigma}$. We denote them by $\vecnbigs_{\star}(\nbiga,S,\gamma,\vecb)$. \hfill\qed \end{lem} If $0$ is contained in the interior part of $\Conv(\nbiga)$, then $\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$ is pure, and $\vecnbigs_{\star}(\nbiga,S,\gamma,\vecb)$ is equal to $\pi_{\dagger}^0\nbigs_{\gamma,\Sigma}$ for any toric desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$. \subsubsection{The smooth part and the induced graded pairings} Recall that we have the open subset $U_{\nbiga}^{\reg}$ in $H^0\bigl(\proj^m,\nbigo_{\proj^m}(1)\bigr)$. (See {\rm\S}.) For any $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo_{\proj^m}(1)\bigr)$, we set $S^{\reg}:=\gamma^{-1}(U_{\nbiga}^{\reg})$. \begin{prop} Let $\nbigm$ be the $\nbigr_{S}$-modules underlying $\pi_{\dagger}^0\nbigt$ for $\nbigt= \nbigt_{\star}(\nbiga,S,\gamma)$ $(\star=\ast,!,\min)$, $\nbigk_{\nbiga,S,\gamma}$ and $\nbigc_{\nbiga,S,\gamma}$. Then, $\nbigm_{|S^{\reg}}$ is a locally free $\nbigo_{\cnum_{\lambda}\times S^{\reg}}$-module. In particular, the mixed twistor $\nbigd$-module $\pi_{\dagger}^0\nbigt_{|S^{\reg}}$ comes from a graded polarizable variation of mixed twistor structure on $S^{\reg}$. It is admissible along $S\setminus S^{\reg}$. \end{prop} \pf It follows from Corollary and the general property of mixed twistor $\nbigd$-modules. \hfill\qed \begin{cor} The restriction of {\rm()} and {\rm()} to $S^{\reg}$ are locally free $\nbigo_{\cnum_{\lambda}\times S^{\reg}}$-modules. Equivalently, the restriction of {\rm()}, {\rm()}, {\rm()} and {\rm()} to $S^{\reg}$ are locally free $\nbigo_{\cnum_{\lambda}\times S^{\reg}}$-modules. The restriction of the image of {\rm()} to $S^{\reg}$ is also an $\nbigo_{\cnum_{\lambda}\times S^{\reg}}$-module. \hfill\qed \end{cor} Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo_{\proj^m}(1)\bigr)$ be a morphism which is non-degenerate at $\infty$ for $X_{\nbiga}$. Suppose that (i) $S^{\reg}\neq\emptyset$, (ii) we are given a hypersurface $Y$ which contains $S\setminus S^{\reg}$. Take $\vecb\in\seisuu^n$ such that $0$ is an interior point of $\Conv(\nbiga\cup\{\vecb\})$. We have the mixed twistor $\nbigd$-modules $\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$ with the induced graded polarization $\vecnbigs_{\star}(\nbiga,S,\gamma,\vecb)$ on $S$. They are also equipped with the natural real structure. We set \[ \nbigv_{\star}(\nbiga,S,\gamma):= \pi_{\dagger}^0\nbigl_{\star}(\nbiga,S,\gamma)(\ast Y). \] We obtain the filtration $\Wtilde$ and the graded pairing $P_{\star}(\nbiga,S,\gamma,\vecb)=\bigl( P_{\star}(\nbiga,S,\gamma,\vecb)_k\,\big|\, k\in\seisuu \bigr)$ by the procedure in \S. We explain it in this situation. The $\nbigr$-modules $\pi_{\dagger}^0\nbigl_{\star}(\nbiga,S,\gamma)$ are equipped with the filtration $W$ underlying the weight filtration of $\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$. We set \[ \Wtilde_{k}\nbigv_{\star}(\nbiga,S,\gamma):= \Bigl( W_{k+d_S} \pi_{\dagger}^0\nbigl_{\star}(\nbiga,S,\gamma) \Bigr)(\ast Y) \] The polarization and the real structure of $\Gr^W_{k+d_S}\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$ induce a pairing $P_{\star}(\nbiga,S,\gamma)_k$ of weight $k$ on $\Gr^{\Wtilde}_k\nbigv_{\star}(\nbiga,S,\gamma)$. Thus, we obtain a graded pairing $P_{\star}(\nbiga,S,\gamma)= \bigl(P_{\star}(\nbiga,S,\gamma)_k\,\big|\,k\in\seisuu\bigr)$ on $\bigl( \nbigv_{\star}(\nbiga,S,\gamma),\Wtilde \bigr)$. In this way, we obtain mixed TEP-structures $\bigl( \nbigv_{\star}(\nbiga,S,\gamma), \Wtilde, P_{\star}(\nbiga,S,\gamma) \bigr)$. \subsubsection{Algebraicity} Let $Z$ be any smooth complex quasi-projective variety. An integrable mixed twistor $\nbigd$-module $(\nbigt,W)$ on $Z$ is called algebraic if the following holds: \begin{itemize} \item Let $\Zbar$ be a smooth projective manifold with an open immersion $Z\lrarr \Zbar$. Then, there exists an integrable mixed twistor $\nbigd$-module $\nbigt'$ on $\Zbar$ such that $\nbigt'_{|Z}=\nbigt$. \end{itemize} We also have the notion of the algebraicity of the underlying $\nbigrtilde$-modules of integrable mixed twistor $\nbigd$-modules. Let $\nbigd^a_{\cnum_{\lambda}\times Z}$ denote the sheaf of algebraic linear differential operators on $\cnum_{\lambda}\times Z$. Let $\Theta_Z^a$ denote the algebraic tangent sheaf of $Z$. Let $p:\cnum_{\lambda}\times Z\lrarr Z$ denote the projection. Let $\nbigrtilde^{a}_{Z}\subset \nbigd^a_{\cnum_{\lambda}\times Z}$ denote the sheaf of subalgebras generated by $\lambda\cdot p^{\ast}\Theta_Z$ and $\lambda^2\del_{\lambda}$ over $\nbigo_{\cnum_{\lambda}\times Z}$. We say that an $\nbigrtilde_Z$-module $\nbigm$ is algebraic if there exists an $\nbigrtilde^a_{Z}$-module $\nbigm^a$ such that $\nbigm$ is isomorphic to the analytification of $\nbigm^a$. \vspace{.1in} We naturally regard $H^0\bigl(\proj^m,\nbigo(1)\bigr)$ as a quasi-projective variety. Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ be an algebraic morphism of smooth quasi-projective varieties. \begin{prop} \mbox{{}} \begin{itemize} \item The integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(\nbiga,S,\gamma)$ and $\pi_{\dagger}^i\nbigt_{\star}(\nbiga,S,\gamma)$ are algebraic. The underlying $\nbigrtilde$-modules of $\nbigt_{\star}(\nbiga,S,\gamma)$ and $\pi_{\dagger}^i\nbigt_{\star}(\nbiga,S,\gamma)$ are algebraic. \item For any toric desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$, the integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(F_{\gamma,S},D_{\Sigma,S})$ are algebraic. The underlying $\nbigrtilde$-modules of $\nbigt_{\star}(F_{\gamma,S},D_{\Sigma,S})$ are algebraic. \end{itemize} \end{prop} \pf The claims for $\nbigt_{\star}(F_{\gamma,S},D_{\Sigma,S})$ is clear. Then, we obtain the claims for the others by construction. \hfill\qed \subsection{Description as a specialization} \subsubsection{$\nbigd$-modules} Let $\nbiga=\{\veca_1,\ldots,\veca_m\}\subset\seisuu^{n}$ be a finite subset which generates $\seisuu^n$. We take $\veca_{m+1}\in\seisuu^{n}$ such that $0$ is contained in $\Conv(\nbigatilde)$, where $\nbigatilde:=\nbiga\cup\{\veca_{m+1}\}$. Let us compare the $\nbigd$-modules associated to $\nbiga$ and $\nbigatilde$. We identify $\proj^{m}$ with the subspace of $\proj^{m+1}$ determined by $z_{m+1}=0$. We naturally regard $H^0\bigl( \proj^{m+1},\nbigo(1) \bigr) =H^0\bigl( \proj^{m},\nbigo(1) \bigr) \times \cnum\cdot z_{m+1}$. Let $\gamma:S\lrarr H^0\bigl(\proj^m,\nbigo(1)\bigr)$ be a morphism which is non-degenerate at $\infty$ for $X_{\nbiga}$. We set $\Stilde:=S\times\cnum_{\tau}$. We have the induced map $\gammatilde=\gamma\times\id: \Stilde\lrarr H^0\bigl( \proj^{m+1},\nbigo(1) \bigr)$ given by $\gammatilde(x,\tau) =\gamma(x)+\tau z_{m+1}$. Let $\pi:\proj^m\times S\lrarr S$ and $\pitilde:\proj^{m+1}\times \Stilde\lrarr \Stilde$ denote the projections. We have the following proposition on the associated $\nbigd$-modules. \begin{prop} \mbox{{}} \begin{itemize} \item We have $L_{\ast}(\nbigatilde,\Stilde,\gammatilde) \simeq L_{!}(\nbigatilde,\Stilde,\gammatilde) \simeq L_{\min}(\nbigatilde,\Stilde,\gammatilde)$. We also have $\pitilde^i_+L_{\star}(\nbigatilde,\Stilde,\gammatilde)=0$ for $i\neq 0$. \item Let $\iota:S\lrarr \Stilde$ be the inclusion induced by $\{0\}\lrarr \cnum$. The kernel and the cokernel of the morphism \[ \pitilde^0_+L_{\min}(\nbigatilde,\Stilde,\gammatilde)(!\tau) \lrarr \pitilde^0_+L_{\min}(\nbigatilde,\Stilde,\gammatilde)(\ast\tau) \] are naturally isomorphic to $\iota_{+} \pi_{+}^0L_{!}\bigl( \nbiga,S,\gamma \bigr)$ and $\iota_{+} \pi_{+}^0L_{\ast}\bigl( \nbiga,S,\gamma \bigr)$, respectively. \end{itemize} \end{prop} \pf Let $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbigatilde}$ be a toric desingularization. We may assume to have a toric morphism $\varphi_{\Sigma,\nbiga}:X_{\Sigma}\lrarr X_{\nbiga}$ which is a desingularization. For the meromorphic function $f=t^{\veca_{m+1}}$ on $X_{\Sigma}$, we may assume $|(f)_0|\cap|(f)_{\infty}|=\emptyset$. We have $F_{\Stilde,\gammatilde} =F_{S,\gamma}+\tau t^{\veca_{m+1}}$. We can check that the assumption in Lemma is satisfied for $F_{S,\gamma}$ and $t^{\veca_{m+1}}$. (See \S.) Hence, we have the purity of $F_{\Stilde,\gammatilde}$ on $(X_{\Sigma,\Stilde},D_{\Sigma,\Stilde})$. Let $\pitilde_{\Sigma}:X_{\Sigma,\Stilde}\lrarr \Stilde$ be the projection. By using Proposition , we obtain $\pitilde^i_{\Sigma+} L(F_{\Stilde,\gammatilde},D_{\Sigma,\Stilde})=0$ for $i\neq 0$. Thus, we obtain the first claim of Proposition . Let $\iota_{1}:X_{\Sigma,S}\lrarr X_{\Sigma,\Stilde}$ denote the inclusion induced by $\{0\}\lrarr \cnum$. According to Proposition , the kernel and the cokernel of the morphism $L(F_{\Stilde,\gammatilde},D_{\Sigma,\Stilde})(!\tau) \lrarr L(F_{\Stilde,\gammatilde},D_{\Sigma,\Stilde})(\ast\tau)$ are naturally isomorphic to $\iota_{1+} L_{!}\bigl( F_{S,\gamma},D_{\Sigma,S}\bigr)$ and $\iota_{1+} L_{\ast}\bigl( F_{S,\gamma},D_{\Sigma,S}\bigr)$, respectively. Let $\pi_{\Sigma}:X_{\Sigma}\times S\lrarr S$ be the projection. By Proposition , we have $\pi_{\Sigma}^i L_{\star}(F_{S,\gamma},D_{\Sigma,S})=0$ for $i\neq 0$. We also have the vanishing in the first claim. Then, we obtain the second claim of Proposition . \hfill\qed \vspace{.1in} We have the canonical nilpotent map $N$ on $\psi_{\tau}\pitilde^0_{+}L_{\ast}(\nbigatilde,\Stilde,\gammatilde) \simeq \psi_{\tau}\pitilde_{+}^0L_!(\nbigatilde,\Stilde,\gammatilde)$ induced by $t\del_t$. \begin{cor} We have the following commutative diagram: \[ \begin{CD} \Ker N @>>> \Cok N\\ @V{\simeq}VV @V{\simeq}VV \\ \pi^0_{+}L_{!}(\nbiga,S,\gamma) @>>> \pi^0_{+}L_{\ast}(\nbiga,S,\gamma). \end{CD} \] The lower horizontal arrow is the natural morphism. \end{cor} \pf It follows from Proposition . \hfill\qed \subsubsection{Graded polarized mixed twistor $\nbigd$-modules} We continue to use the setting in \S, and we consider the associated mixed twistor $\nbigd$-modules. Let us show the relations between the mixed twistor $\nbigd$-modules associated $(\nbiga,S,\gamma)$ and $(\nbigatilde,\Stilde,\gammatilde)$, although we have already used it in \S implicitly. \begin{prop} \mbox{{}} \begin{itemize} \item We have $\nbigt_{!}(\nbigatilde,\Stilde,\gammatilde) \simeq \nbigt_{\ast}(\nbigatilde,\Stilde,\gammatilde) \simeq \nbigt_{\min}(\nbigatilde,\Stilde,\gammatilde)$, and $\pitilde_{\dagger}^i\nbigt_{\min}(\nbigatilde,\Stilde,\gammatilde)=0$ for $i\neq 0$. \item Let $\nbigk(\nbigatilde,\Stilde,\gammatilde)$ and $\nbigc(\nbigatilde,\Stilde,\gammatilde)$ denote the kernel and the cokernel of the following morphism of the mixed twistor $\nbigd$-modules: \[ \pitilde^0_{\dagger}\nbigt_{\ast}(\nbigatilde,\Stilde,\gammatilde)[!\tau] \lrarr \pitilde^0_{\dagger}\nbigt_{\ast}(\nbigatilde,\Stilde,\gammatilde)[\ast \tau]. \] Then, we have the following natural isomorphisms: \[ \nbigk(\nbigatilde,\Stilde,\gammatilde) \simeq \iota_{\dagger}\pi_{\dagger}^0 \nbigt_{!}(\nbiga,S,\gamma), \quad \nbigc(\nbigatilde,\Stilde,\gammatilde) \simeq \iota_{\dagger}\pi_{\dagger}^0 \nbigt_{\ast}(\nbiga,S,\gamma) \otimes\newTate(-1). \] The isomorphisms are compatible with the real structures. \end{itemize} \end{prop} \pf We obtain the first claim from Proposition . As in the proof of Proposition , by using the vanishing of the cohomology in Proposition and the first claim of this proposition, we obtain the isomorphisms from Proposition . We can easily compare the real structures by using Proposition . \hfill\qed \vspace{.1in} On one hand, we have the graded polarizations $\vecnbigs_{\star}(\nbiga,S,\gamma,\veca_{m+1})$ on $\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$ as explained in \S. They induce graded sesqui-linear dualities on $\iota_{\dagger}\pi_{\dagger}^0 \nbigt_!(\nbiga,S,\gamma)$ and $\iota_{\dagger}\pi_{\dagger}^0 \nbigt_{\ast}(\nbiga,S,\gamma) \otimes\newTate(-1)$. On the other hand, as explained in \S, the pure twistor $\nbigd$-module $\pi_{\dagger}^0\nbigt_{\min}(\nbigatilde,\Stilde,\gammatilde)$ is equipped with the induced polarization. It induces the graded polarizations $\vecnbigs_{\nbigk}$ and $\vecnbigs_{\nbigc}$ of $\nbigk(\nbigatilde,\Stilde,\gammatilde)$ and $\nbigc(\nbigatilde,\Stilde,\gammatilde)$, as explained in \S. \begin{prop} The isomorphisms in Proposition {\rm} are compatible with the graded polarizations. \end{prop} \pf We use the notation in the proof of Proposition . Let $\nbigk_1$ and $\nbigc_1$ be the kernel and the cokernel of the morphism of the mixed twistor $\nbigd$-modules $\nbigt_!(F_{\gammatilde,\Sigma},D_{\Sigma,\Stilde}) \lrarr \nbigt_{\ast}(F_{\gammatilde,\Sigma},D_{\Sigma,\Stilde})$. According to Proposition , the isomorphisms $\nbigk_1\simeq \iota_{1\dagger} \nbigt_{!}(F_{\gamma,\Sigma},D_{\Sigma,S})$ and $\nbigc_1\simeq \iota_{1\dagger} \nbigt_{\ast}(F_{\gamma,\Sigma},D_{\Sigma,S})$ are compatible with the graded sesqui-linear dualities. Then, the claims of the propositions follow from Corollary and the construction of the graded polarizations. \hfill\qed \vspace{.1in} Let $\nbign: \psitilde_{\tau,-\vecdelta} \pitilde_{\dagger}^0\nbigt(\nbigatilde,\Stilde,\gammatilde) \otimes\nbigu(-1,0) \lrarr \psitilde_{\tau,-\vecdelta} \pitilde_{\dagger}^0\nbigt(\nbigatilde,\Stilde,\gammatilde) \otimes\nbigu(0,-1)$ be the morphism given by the pair $(-t\del_t,-t\del_t)$. \begin{cor} We have the following commutative diagram: \begin{equation} \begin{CD} \Ker(\nbign) @>>> \Cok(\nbign)\otimes\newTate(1)\\ @V{\simeq}VV @V{\simeq}VV \\ \pi_{\dagger}^0\nbigt_{!}(\nbiga,S,\gamma) @>>> \pi_{\dagger}^0\nbigt_{\ast}(\nbiga,S,\gamma) \end{CD} \end{equation} The horizontal arrows are the natural morphisms. \end{cor} \pf Because $\iota_{\dagger}\Cok(\nbign)$ and $\iota_{\dagger}\Ker(\nbign)$ are naturally isomorphic to $\nbigc(\nbigatilde,\Stilde,\gammatilde)$ and $\nbigk(\nbigatilde,\Stilde,\gammatilde)$, we obtain the vertical isomorphisms in () from Proposition . The commutativity of the diagram follows from Corollary . \hfill\qed \subsubsection{The induced mixed TEP-structures} We set $\nbigv:= \pitilde_{\dagger}^0\nbigl(\nbigatilde,\Stilde,\gammatilde)(\ast\tau)$. Let $\lambda N:\psi_{\tau,-\vecdelta}(\nbigv) \lrarr \psi_{\tau,-\vecdelta}(\nbigv)$ be the morphism induced by $\lambda\tau\del_{\tau}$. We restate the result in \S in this situation. \begin{assumption} $\nbigv$ is regular singular along $\tau=0$. There exists an open subset $\Btilde$ in $\Stilde$ such that $\nbigv_{|\Btilde}$ is a locally free $\nbigo_{\cnum_{\lambda}\times \Btilde}(\ast \tau)$-module. \hfill\qed \end{assumption} We set $B:=\iota^{-1}(\Btilde)$ which is an open subset in $S$. Under the assumption $\Cok(\lambda N)_{|B}$ is a locally free $\nbigo_{\cnum_{\lambda}\times B}$-module. We also have the pairing $P_{\nbigatilde}$ of weight $n$ on $\nbigv_{|\Btilde}$ induced by the real structure and the polarization of $\pitilde_{\dagger}^0\nbigt(\nbigatilde,\Stilde,\gammatilde)$. By the procedure in \S, we have the weight filtration $\Wtilde^{(1)}$ and the graded pairing $(P^{(1)}_{\nbigatilde,k}\,|\,k\in\seisuu)$ on $\Cok(\lambda N)_{|B}$ and $\Ker(\lambda N)_{|B}$. \vspace{.1in} Note that $\Cok(\lambda N)$ is the underlying $\nbigr$-module of $\Cok(\nbign)\otimes\newTate(1) \simeq \pi_{\dagger}^0\nbigt_{\ast}(\nbiga,S,\gamma)$, i.e., $\Cok(\lambda N)$ is isomorphic to $\pi_{\dagger}^0\nbigl_{\ast}(\nbiga,S,\gamma)$. As explained in \S, we obtain the weight filtration $\Wtilde^{(2)}$ and the graded pairing $(P^{(2)}_{\nbigatilde,k}\,|\,k\in\seisuu)$ on $\Cok(\lambda N)_{|B}$ from the real structure and the graded polarization of $\pi_{\dagger}^0\nbigt(\nbiga,S,\gamma)$. Similarly, $\Ker(\lambda N)$ is the underlying $\nbigr$-module of $\Ker(\nbign)\simeq \pi_{\dagger}^0\nbigt_!(\nbiga,S,\gamma)$, i.e., $\Ker(\lambda N)$ is isomorphic to $\pi_{\dagger}^0\nbigl_{!}(\nbiga,S,\gamma)$. As explained in \S, we have the filtration $\Wtilde^{(2)}$ and the graded pairing $(P^{(2)}_{\nbigatilde,k}\,|\,k\in\seisuu)$ on $\Ker(\lambda N)$. \begin{prop} On $\Cok(\lambda N)$ and $\Ker(\lambda N)$, we have $\Wtilde^{(1)}=\Wtilde^{(2)}$ and $P^{(1)}_{\nbigatilde,k} =P^{(2)}_{\nbigatilde,k}$ for any $k\in\seisuu$. \end{prop} \pf It follows from Proposition and the isomorphisms $\Cok(\nbign)\otimes\newTate(1)\simeq \pi_{\dagger}^0\nbigt_{\ast}(\nbiga,S,\gamma)$ and $\Ker(\nbign)\simeq \pi_{\dagger}^0\nbigt_{!}(\nbiga,S,\gamma)$ in Corollary . \hfill\qed \subsection{A regular singular case} \subsubsection{The $\nbigd$-modules} Let $\nbiga=\{\veca_1,\ldots,\veca_m\}$ be as in \S. Suppose that there exists $\alpha\in (\seisuu^n)^{\lor}$ such that $\alpha(\veca_i)=1$ for any $\veca_i\in\nbiga$. In , and , the $\nbigd$-modules are described in terms of the Gauss-Manin connection for the relative cohomology groups of some families. We review it in the language of $\nbigd$-modules, which fits to the theory of mixed twistor $\nbigd$-modules. With an appropriate choice of a frame of $\seisuu^n$, we may assume that $\veca_i=(\vecb_i,1)\in\seisuu^{n-1}\times\seisuu$ for each $\veca_i\in\nbiga$. Let $\nbigb\subset\seisuu^{n-1}$ denote the image of $\nbiga$ via the projection of $\seisuu^{n-1}\times\seisuu\lrarr\seisuu^{n-1}$. In other words, we have $\nbiga=\{(\vecb,1)\,|\,\vecb\in\nbigb\}$. For simplicity, we impose the following in this subsection. \begin{assumption} We assume that $\nbigb$ generates $\seisuu^{n-1}$, and $0$ is contained in the interior part of $\Conv(\nbigb)$. \hfill\qed \end{assumption} Let $X_{\nbigb}$ denote the closure of the image of the morphism $\psi_{\nbigb}:T^{n-1}\lrarr \proj^{m}$ as in \S. We set $W:=\Bigl\{\sum_{i=1}^m \alpha_iz_i\,\Big|\, \alpha_i\in\cnum \Bigr\} \subset H^0\bigl(\proj^m,\nbigo(1)\bigr) =\Bigl\{ \sum_{i=0}^m\alpha_iz_i \Bigr\}$. We use the notation in \S. \begin{lem} Let $\gamma:S\lrarr W$ be a morphism of complex manifolds. The following conditions are equivalent. \begin{itemize} \item $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$. \item If $\gamma=\sum_{i=1}^m \gamma_iz_i$, the family of Laurent polynomials $F_{\gamma,\nbigb} =\sum_{i=1}^m \gamma_it^{\vecb_i}$ is $\Conv(\nbigb)$-regular in the sense that $F_{\gamma,\nbigb,\sigma}^{-1}(0) \cap (dF_{\gamma,\nbigb,\sigma})^{-1}(0) =\emptyset$ for any face $\sigma$ of $\Conv(\nbigb)$. \end{itemize} \end{lem} \pf We have the family of Laurent polynomials $F_{\gamma}=\sum \gamma_i t^{\veca_i}$ associated to $\nbiga$ and $\gamma$. For any face $\sigma$ of $\nbigb$, we have the face $\sigma(\nbiga)$ of $\nbiga$ given by $\sigma(\nbiga):=\bigl\{ (1,\vecc)\,\big|\,\vecc\in\sigma \bigr\}$. We have $F_{\gamma,\sigma(\nbiga)} =t_nF_{\gamma,\nbigb,\sigma}$ and $dF_{\gamma,\sigma(\nbiga)}= t_ndF_{\gamma,\nbigb,\sigma} +dt_n F_{\gamma,\nbigb,\sigma}$. We have $F_{\gamma,\nbigb,\sigma}^{-1}(0) =F_{\gamma,\sigma(\nbiga)}^{-1}(0)$ and $(dF_{\gamma,\sigma(\nbiga)})^{-1}(0) = F_{\gamma,\nbigb,\sigma}^{-1}(0) \cap (dF_{\gamma,\nbigb,\sigma})^{-1}(0)$. Then, the claim is clear. \hfill\qed \vspace{.1in} Let $\gamma:S\lrarr W$ be any holomorphic map. Let us study the $\nbigd$-modules $L_{\star}(\nbiga,S,\gamma)$ and $\pi^i_{+}\bigl( L_{\star}(\nbiga,S,\gamma) \bigr)$. It is convenient to consider a toric compactification which is different from that in \S. We set $X'_{\nbigb}:=X_{\nbigb}\times\proj^1$ which is naturally a toric variety. We take a toric desingularization $\varphi_{\Sigma_0}: X_{\Sigma_0}\lrarr X_{\nbigb}$. We naturally have a fan $\Sigma_0'$ such that $X_{\Sigma_0'}=X_{\Sigma_0}\times\proj^1$. We set $\varphi_{\Sigma_0'}:= \varphi_{\Sigma_0}\times\id: X_{\Sigma'_0}\lrarr X'_{\nbigb}$. Let $\varphibar_{\Sigma_0',S}$ denote the composite of $\varphi_{\Sigma_0'}\times\id_S$ and the inclusion $X_{\nbigb,S}'\lrarr \proj^m\times\proj^1\times S$. We have the meromorphic function $F_{\gamma,\Sigma_0'}$ on $(X_{\Sigma_0',S},D_{\Sigma_0',S})$ determined by the family of Laurent polynomials $F_{\gamma}$. (See \S for $F_{\gamma}$.) For $\star=\ast,!$, we have the $\nbigd$-modules $L_{\star}(F_{\gamma,\Sigma'_0},D_{\Sigma'_0,S})$ on $X_{\Sigma_0',S}$. We set \[ L'_{\star}(\nbigb,S,\gamma):= \varphibar_{\Sigma_0',S+} L_{\star}(F_{\gamma,\Sigma'_0},D_{\Sigma'_0,S}) \] on $\proj^m\times\proj^1\times S$. It is easy to see that they are independent of the choice of a desingularization $\varphi_{\Sigma_0}$ up to natural isomorphisms. Let $\pi_{\nbigb}:\proj^m\times\proj^1\times S \lrarr S$ and $\pi_{\Sigma_0'}:X_{\Sigma_0',S}\lrarr S$ denote the projections. We have the natural isomorphisms $\pi^i_{\nbigb+} L'_{\star}(\nbigb,S,\gamma) \simeq \pi^i_{\Sigma'_0+} L_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ for $\star=\ast,!$. \begin{lem} We have natural isomorphisms $\pi^i_{\nbigb+}L'_{\star}(\nbigb,S,\gamma) \simeq \pi^i_+L_{\star}(\nbiga,S,\gamma)$. \end{lem} \pf We take a toric desingularization $\varphi_{\Sigma_1}:X_{\Sigma_1}\lrarr X_{\nbiga}$. We may assume that we have a toric morphism $\rho:X_{\Sigma_1}\lrarr X_{\Sigma_0'}$ which is birational. We have natural isomorphisms $\rho_+L_{\star}(F_{\gamma,\Sigma_1},D_{\Sigma_1,S}) \simeq L_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$. They induce the desired isomorphisms. \hfill\qed \vspace{.1in} The family of Laurent polynomials $F_{\gamma}$ is described as \[ F_{\gamma}=t_nf_{\gamma}, \quad\quad f_{\gamma}:=\sum_{i=1}^m \gamma_it^{\vecb_i}. \] Let $Z_{f_{\gamma}}\subset X_{\Sigma_0,S}$ denote the zero set of $f_{\gamma}$. Suppose that $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$. As remarked in Lemma , the family of Laurent polynomials $f_{\gamma}$ is $\Conv(\nbigb)$-regular. According to Proposition , we have \begin{equation} \pi^0_{+}L_{\ast}(\nbiga,S,\gamma) \simeq \pi^0_{\Sigma'_0+} L_{\ast}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S}) \simeq \pi^0_{\Sigma_0+} \bigl( \nbigo_{X_{\Sigma_0,S}} (!Z_{f_{\gamma}}) (\ast D_{\Sigma_0,S}) \bigr). \end{equation} \begin{equation} \pi_{+}^0L_{!}(\nbiga,S,\gamma) \simeq \pi_{\Sigma_0'+}^0 L_{!}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S}) \simeq \pi^0_{\Sigma_0+} \bigl( \nbigo_{X_{\Sigma_0,S}} (\ast Z_{f_{\gamma}}) (!D_{\Sigma_0,S}) \bigr). \end{equation} We have $\pi^i_{\Sigma_0+} \bigl( \nbigo_{X_{\Sigma_0,S}} (!Z_{f_{\gamma}})(\ast D_{\Sigma_0,S}) \bigr)=0$ and $\pi^i_{\Sigma_0+} \bigl( \nbigo_{X_{\Sigma_0,S}} (\ast Z_{f_{\gamma}})(!D_{\Sigma_0,S}) \bigr)=0$ for $i\neq 0$, which follow from $\pi^i_{\Sigma+} L_{\star}(\nbiga,S,\gamma)=0$ for $i\neq 0$. If moreover the image of $\gamma$ is contained in $W\cap U_{\nbiga}^{\reg}$, the $\nbigd_S$-modules () and () are flat bundles on $S$. The fiber of () over $s\in S$ is the relative cohomology group $H^{n-1}(T^{n-1},Z^{\circ}_{f_{\gamma(s)}})$ with $\cnum$-coefficient, where $Z^{\circ}_{f_{\gamma(s)}}:=Z_{f_{\gamma(s)}}\cap T^{n-1}$. The fiber of () over $s\in S$ is $H^{n-1}(X_{\Sigma_0}\setminus Z_{f_{\gamma(s)}}, D_{\Sigma_0}\setminus Z_{f_{\gamma(s)}})$. \begin{rem} Under the identification $X_{\Sigma_0'}=\proj^1_s\times X_{\Sigma_0}$, as remarked in {\rm \S}, {\rm()} is also naturally isomorphic to $\pi^0_{\Sigma_0'+} \bigl( \nbigo_{X_{\Sigma_0}',S} (!Z_{f_{\gamma}+s}) (\ast D_{\Sigma_0'}) \bigr)$. \hfill\qed \end{rem} \subsubsection{The mixed twistor $\nbigd$-modules} We continue to use the notation in \S. Suppose that $\gamma$ is non-degenerate at $\infty$ for $X_{\nbiga}$. We have the mixed twistor $\nbigd$-modules $\nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ $(\star=\ast,!)$ on $X_{\Sigma_0',S}$. We obtain the mixed twistor $\nbigd$-modules on $\proj^m\times\proj^1\times S$: \[ \nbigt'_{\star}(\nbigb,S,\gamma) :=\varphibar_{\Sigma_0',S\dagger} \nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S}) \] They are independent of the choice of a toric desingularization $\varphi_{\Sigma_0}:X_{\Sigma_0}\lrarr X_{\nbigb}$. We have the natural isomorphisms $\pi^i_{\nbigb\dagger} \nbigt'_{\star}(\nbigb,S,\gamma) \simeq \pi^i_{\Sigma'_0\dagger} \nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ $(\star=\ast,!)$. As in the case of $\nbigd$-modules, we have the following. \begin{lem} We have natural isomorphisms $\pi^i_{\nbigb\dagger} \nbigt'_{\star}(\nbigb,S,\gamma) \simeq \pi^i_{\dagger} \nbigt_{\star}(\nbiga,S,\gamma)$ for $\star=\ast,!$. \hfill\qed \end{lem} Let $\nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ $(\star=\ast,!)$ be the $\nbigr$-modules underlying $\nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ on $X_{\Sigma_0',S}$. Set $\nbigl_{\star}'(\nbigb,S,\gamma):= \varphibar^0_{\Sigma_0',S\dagger} \nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ on $\proj^m\times\proj^1\times S$, which underlie $\pi_{\dagger}^0\nbigt'_{\star}(\nbigb,S,\gamma)$. We consider the standard $\cnum^{\ast}$-action on $\proj^1$ given by $a\cdot [y_0:y_1]=[ay_0:y_1]$. It induces $\cnum^{\ast}$-actions on $X_{\Sigma_0',S}$ and $\proj^m\times \proj^1\times S$. \begin{lem} The $\nbigr$-modules $\nbigl'_{\star}(\nbigb,S,\gamma)$ and $\nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ are $\cnum^{\ast}$-homogeneous. (See {\rm\S} for the notion of homogeneity.) \end{lem} \pf We consider the $\cnum^{\ast}$-action on $T^n$ given by $a\cdot (t_1,\ldots,t_{n-1},t_n)=(t_1,\ldots,t_{n-1},at_n)$. It induces a $\cnum^{\ast}$-action on $X_{\Sigma_0',S}$. We have $a^{\ast}F_{\gamma,\Sigma_0'}=a\cdot F_{\gamma,\Sigma_0'}$. By Proposition , the $\nbigr$-modules $\nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ are $\cnum^{\ast}$-homogeneous. Then, we obtain that $\nbigl'_{\star}(\nbigb,S,\gamma)$ are also $\cnum^{\ast}$-homogeneous. \hfill\qed \vspace{.1in} Let $q_{\nbigb}:\proj^m\times\proj^1\times S\lrarr \proj^m\times S$ and $q_{\Sigma_0}:X_{\Sigma_0',S}\lrarr X_{\Sigma_0,S}$ be the projections. \begin{cor} The $\nbigr$-modules $q_{\nbigb\dagger}\nbigl'_{\star}(\nbigb,S,\gamma)$ and $q_{\Sigma_0\dagger} \nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ are $\cnum^{\ast}$-homogeneous for the trivial $\cnum^{\ast}$-actions on $\proj^m\times S$ and $X_{\Sigma_0}\times S$. Hence, $q_{\nbigb\dagger}\nbigt'_{\star}(\nbigb,S,\gamma)$ and $q_{\Sigma_0\dagger} \nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ come from mixed Hodge modules. In particular, the underlying $\nbigd$-modules are equipped with the good filtrations such that the analytification of the Rees modules are isomorphic to $q_{\nbigb\dagger}\nbigl'_{\star}(\nbigb,S,\gamma)$ and $q_{\Sigma_0\dagger} \nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$. \end{cor} \pf Because $q_{\nbigb}$ and $q_{\Sigma_0}$ are $\cnum^{\ast}$-equivariant, the first claim follows from Lemma . The second follows from Theorem . \hfill\qed \begin{cor} The $\nbigr$-modules $\pi^0_{\nbigb\dagger}\nbigl'_{\star}(\nbigb,S,\gamma) \simeq \pi^0_{\Sigma'_0\dagger} \nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ are $\cnum^{\ast}$-homogeneous for the trivial $\cnum^{\ast}$-actions on $S$. Hence, $\pi_{\nbigb\dagger}^0\nbigt'_{\star}(\nbigb,S,\gamma)$ come from mixed Hodge modules. In particular, the underlying $\nbigd$-modules are equipped with the good filtrations such that the analytification of the Rees modules are isomorphic to $\pi^0_{\nbigb\dagger}\nbigl'_{\star}(\nbigb,S,\gamma)$. \hfill\qed \end{cor} \begin{prop} Suppose that $\gamma:S\lrarr W$ is non-degenerate at $\infty$ for $X_{\nbiga}$. Then, we have the following natural isomorphisms of the integrable mixed twistor $\nbigd$-modules compatible with the real structure: \begin{equation} \pi^0_{\dagger}\nbigt_{\ast}(\nbiga,S,\gamma) \simeq \pi^0_{\Sigma_0\dagger}\Bigl( \bigl( \nbigu_{X_{\Sigma_0,S}}(n+d_S,-1) \bigr) [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr). \end{equation} \begin{equation} \pi^0_{\dagger}\nbigt_!(\nbiga,S,\gamma) \simeq \pi^0_{\Sigma_0\dagger}\Bigl( \bigl( \nbigu_{X_{\Sigma_0,S}}(n+d_S-1,0) \bigr) [\ast Z_{f_{\gamma}}][!D_{\Sigma_0,S}] \Bigr). \end{equation} \end{prop} \pf We give an argument for (). The other case is similar. According to Proposition , we have the isomorphism of the integrable mixed twistor $\nbigd$-modules with real structure: \begin{equation} q^0_{\Sigma_0\dagger}\nbigt_{\ast} (F_{\gamma,\Sigma_0'},D_{\Sigma_0',S}) \simeq \bigl( \nbigu_{X_{\Sigma_0,S}}(n-1+d_S,0) \otimes\newTate(-1) \bigr) [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}]. \end{equation} Thus, we obtain the following isomorphisms: \begin{equation} \pi^0_{\dagger}\nbigt_{\ast}(\nbiga,S,\gamma) \simeq \pi^0_{\Sigma_0'\dagger}\nbigt_{\ast} (F_{\gamma,\Sigma_0'},D_{\Sigma_0',S}) \simeq \pi^0_{\Sigma_0\dagger}\Bigl( \bigl( \nbigu_{X_{\Sigma_0,S}}(n-1+d_S,0) \otimes\newTate(-1) \bigr) [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr). \end{equation} Thus, we obtain the claim of the proposition. \hfill\qed \vspace{.1in} Let $M$ be the pure Hodge module on $X_{\Sigma_0,S}$ of weight $\dim S+n-1$, corresponding to the constant sheaf. Let $M(-1)$ be the $(-1)$-th Tate twist of $M$. We obtain the mixed Hodge module $\pi_{\Sigma_0\dagger}^0M(-1)[!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}]$ which induces the mixed twistor $\nbigd$-module in the right hand side of (). We also have the mixed Hodge module $\pi_{\Sigma_0\dagger}^0M[\ast Z_{f_{\gamma}}][!D_{\Sigma_0,S}]$. \begin{cor} The Hodge filtrations of $\pi_{\Sigma_0\dagger}^0M(-1)[!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}]$ and $\pi_{\Sigma_0\dagger}^0M[\ast Z_{f_{\gamma}}][!D_{\Sigma_0,S}]$ are equal to the good filtrations in Corollary {\rm}. \hfill\qed \end{cor} Let $\iota:Z_{f_{\gamma}}\lrarr X_{\Sigma_0,S}$ be the inclusion of the complex manifolds. We have the following natural morphisms of integrable mixed twistor $\nbigd$-modules on $S$: \begin{multline} \pi_{\Sigma_0\dagger}^0 \nbigu_{X_{\Sigma_0,S}}(n+d_S-1,0)[\ast Z_{f_{\gamma}}][!D_{\Sigma_0,S}] \stackrel{a_1}{\lrarr} \pi_{\Sigma_0\dagger}^0 \iota_{\dagger}\nbigu_{Z_{f_{\gamma}}}(n+d_S-1,-1) \\ \stackrel{a_2}{\lrarr} \pi_{\Sigma_0\dagger}^0 \nbigu_{X_{\Sigma_0,S}}(n+d_S,-1)[!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \end{multline} \begin{prop} We have $\pi_{\Sigma_0\dagger}^0 \iota_{\dagger}\nbigu_{Z_{f_{\gamma}}}(n+d_S-1,-1) =\Image a_1\oplus\Ker a_2$. We also have \begin{multline} \Gr^W_{n+d_S} \Bigl( \pi_{\Sigma_0\dagger}^0 \nbigu_{X_{\Sigma_0,S}}(n+d_S-1,0) [\ast Z_{f_{\gamma}}][!D_{\Sigma_0,S}] \Bigr) \simeq \Image a_1 \\ \simeq \Gr^{W}_{n+d_S}\Bigl( \pi_{\Sigma_0\dagger}^0 \nbigu_{X_{\Sigma_0,S}}(n+d_S,-1) [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr) \end{multline} \end{prop} \pf According to Corollary , the morphism $a_2\circ a_1$ induces an isomorphism of the weight $(n+d_S)$-parts. Hence, we obtain $\Image (a_2)=\Image(a_2\circ a_2)$, which implies the claims of the proposition. \hfill\qed \subsubsection{Graded polarizations and induced pairings} Let $\vecb=(0,-1)\in\seisuu^{n-1}\times\seisuu$. Note that $0$ is contained in the interior part of $\Conv(\nbiga\cup\{\vecb\})$. Let $\gamma:S\lrarr W$ be a morphism which is non-degenerate at $\infty$ for $X_{\nbiga}$. As explained in \S, we obtain graded polarizations $\vecnbigs_{\star}(\nbiga,S,\gamma,\vecb)$ of $\pi_{\dagger}^0\nbigt_{\star}(\nbiga,S,\gamma)$. Suppose moreover that the image of $\gamma$ is contained in $W\cap U_{\nbiga}^{\reg}$. Then, $\nbigv'_{\star}(\nbigb,S,\gamma):= \pi^0_{\Sigma_0'\dagger} \nbigl_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ are locally free $\nbigo_{\cnum_{\lambda}\times S}$-modules. Let $W$ denote the filtration of $\nbigv'_{\star}(\nbigb,S,\gamma)$ underlying the weight filtration of $\pi_{\Sigma_0'\dagger}^0 \nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$. We set $\Wtilde_k \nbigv'_{\star}(\nbigb,S,\gamma):= W_{d_S+k} \nbigv'_{\star}(\nbigb,S,\gamma)$. Then, the polarization and the real structure of $\Gr^W_{d_S+k} \pi_{\Sigma_0',\dagger}^0 \nbigt_{\star}(F_{\gamma,\Sigma_0'},D_{\Sigma_0',S})$ induce a pairing $\Ptilde_k$ of weight $k$ on $\Gr^{\Wtilde}_k\nbigv'_{\star}(\nbigb,S,\gamma)$, as explained in \S. Let us give a description of the pairings $\Ptilde_{k}$ in the case $\star=\ast$. \vspace{.1in} We have the following isomorphism: \begin{equation} \nbigv'_{\ast}(\nbigb,S,\gamma) \simeq \pi^0_{\Sigma_0\dagger} \nbigo_{\cnum_{\lambda}\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \end{equation} Note that $\nbigo_{\cnum_{\lambda}\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}]$ is equipped with the filtration $W$ which underlies the weight filtration of $\nbigu_{X_{\Sigma_0,S}}(n+d_S,-1)[!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}]$. By using the spectral sequence, we have the following complex \begin{multline} \pi^{-1}_{\Sigma_0\dagger} \Gr^W_{d_S+n+k+1}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr) \stackrel{\alpha_{k}}{\lrarr} \\ \pi^0_{\Sigma_0\dagger} \Gr^W_{d_S+n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr) \stackrel{\beta_{k}}{\lrarr} \\ \pi^1_{\Sigma_0\dagger} \Gr^W_{d_S+n+k-1}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr), \end{multline} and the cohomology is isomorphic to $\Gr^{\Wtilde}_{n+k}\nbigv'_{\ast}(\nbigb,S,\gamma)$. The real structure and the polarization of \[ \pi^0_{\Sigma_0\dagger}\Gr^{W}_{d_S+n+k} \nbigu_{X_{\Sigma_0,S}}(n+d_S,-1) [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \] induces a pairing $\nbigptilde_{n+k}$ of weight $n+k$ on $\pi^0_{\Sigma_0\dagger} \Gr^W_{d_S+n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr)$. The pairing $\nbigptilde_{n+k}$ induces the following isomorphism: \[ \Phi_{\nbigptilde_{n+k}}: \pi^0_{\Sigma_0\dagger} \Gr^W_{d_S+n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr) \simeq \lambda^{-n-k} \pi^0_{\Sigma_0\dagger} \Gr^W_{d_S+n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr)^{\lor} \] As the dual of (), we have the following: \begin{multline} \pi^1_{\Sigma_0\dagger} \Gr^W_{d_S+n+k-1}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr)^{\lor} \stackrel{\beta^{\lor}_{k}}{\lrarr} \\ \pi^0_{\Sigma_0\dagger} \Gr^W_{d_S+n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr)^{\lor} \stackrel{\alpha^{\lor}_{k}}{\lrarr} \\ \pi^{-1}_{\Sigma_0\dagger} \Gr^W_{d_S+n+k+1}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr)^{\lor} \end{multline} We obtain the subsheaf $\Ker\beta_{k}\cap \Ker\bigl( \alpha_{k}^{\lor}\circ\Phi_{\nbigptilde_{n+k}} \bigr)$ in $\pi^0_{\Sigma_0\dagger} \Gr^W_{d_S+n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0,S}} [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}] \Bigr)$. Condition {\bf(A)} for $\pi_{\Sigma_0}$ and $\nbigu_{X_{\Sigma_0,S}}(n+d_S,-1) [!Z_{f_{\gamma}}][\ast D_{\Sigma_0,S}]$ with the graded polarization implies that the induced morphism \begin{equation} \Ker\beta_{k}\cap \Ker\bigl( \alpha_{k}^{\lor}\circ\Phi_{\nbigptilde_{n+k}} \bigr) \lrarr \Gr^{\Wtilde}_{n+k} \nbigv'_{\ast}(\nbigb,S,\gamma) \end{equation} is an epimorphism. The following proposition holds by the construction of $\Ptilde_{n+k}$. \begin{prop} The pull back of $\Ptilde_{n+k}$ by {\rm()} is equal to the restriction of $\nbigptilde_{n+k}$ to $\Ker\beta_{k}\cap \Ker\bigl( \alpha_{k}^{\lor}\circ\Phi_{\nbigptilde_{n+k}} \bigr)$. \hfill\qed \end{prop} In the sense of Proposition , a description of the complex () and the pairing $\nbigptilde_{n+k}$ gives a description of $\Ptilde_{n+k}$. \paragraph{A Hodge theoretic description of the pairings} We shall give a Hodge theoretic description of () and $\nbigptilde_{n+k}$ in the case where $S$ is a point. According to Proposition , it is enough to consider the case where $S$ is a point. To simplify the description, we denote $f_{\gamma}$ by $f$, and the zero-set is denoted by $Z_f$. We have the irreducible decomposition $D_{\Sigma_0}=\bigcup_{i\in \Lambda} D_{\Sigma_0,i}$. For $I\subset \Lambda$, we set $D_{\Sigma_0,I}:=\bigcap_{i\in I}D_{\Sigma_0,i}$. Formally, $D_{\Sigma_0,\emptyset}:=X_{\Sigma_0}$. Let $\iota_{I}:D_{\Sigma_0,I}\lrarr X_{\Sigma_0}$ be the inclusion. Let $Z_{f,I}:=Z_{f}\cap D_{\Sigma_0,I}$ which are also smooth. Let $\iota_{Z_{f},I}$ denote the inclusion $Z_{f,I}\lrarr X_{\Sigma_0}$. We set $\Lambda(k):= \bigl\{I\subset\Lambda\,\big|\,|I|=\Lambda\bigr\}$. Let $W$ be the filtration on $\lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0}} [!Z_f][\ast D_{\Sigma_0,I}]$ underlying the weight filtration of the mixed Hodge module $\nbigu_{X_{\Sigma_0}}(n,-1)[!Z_{f}][\ast D_{\Sigma_0}]$. We have the following natural isomorphism: \[ \Gr^W_{n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0}} [!Z_{f}][\ast D_{\Sigma_0}] \Bigr) \simeq \bigoplus_{I\in\Lambda(k-1)} \iota_{I\dagger} \bigl( \lambda^{-k} \nbigo_{\cnum\times D_{\Sigma_0,I}} \bigr) \oplus \bigoplus_{I\in\Lambda(k)} \iota_{Z_{f},I\dagger} \bigl( \lambda^{-k-1} \nbigo_{\cnum\times Z_{f,I}} \bigr) \] Let $F^{\bullet}$ be the Hodge filtration on $H^{\ast}(D_{\Sigma_0,I})$ which is a decreasing filtration. We set $F_{j}H^m(D_{\Sigma_0,I}):= F^{-j}H^m(D_{\Sigma_0,I})$. Let $R_FH^m(D_{\Sigma_0,I})$ be the Rees module: \[ R_FH^m(D_{\Sigma_0,I}) =\sum F_jH^m(D_{\Sigma_0,I})\lambda^{j} =\sum F^jH^m(D_{\Sigma_0,I})\lambda^{-j} \] The associated $\nbigo_{\cnum_{\lambda}}$-module is also denoted by the same notation. We have the following natural isomorphism: \begin{multline} \pi^i_{\Sigma_0\dagger} \Gr^W_{n+k}\Bigl( \lambda^{-1} \nbigo_{\cnum\times X_{\Sigma_0}} [!Z_{f}][\ast D_{\Sigma_0}] \Bigr) \simeq \\ \bigoplus_{I\in\Lambda(k-1)} \lambda^{-k} R_F H^{n-k+i}(D_{\Sigma_0,I}) \oplus \bigoplus_{I\in\Lambda(k)} \lambda^{-k-1} R_FH^{n-2-k+i}(Z_{f,I}) \end{multline} We obtain the complex () by applying the Rees construction to the following exact sequence of pure Hodge structures: \begin{multline} \bigoplus_{I\in\Lambda(k)} H^{n-k-2}(D_{\Sigma_0,I})\otimes\rnum(-k-1) \oplus \bigoplus_{I\in\Lambda(k+1)} H^{n-4-k}(Z_{f,I})\otimes\rnum(-k-2) \lrarr \\ \bigoplus_{I\in\Lambda(k-1)} H^{n-k}(D_{\Sigma_0,I})\otimes\rnum(-k) \oplus \bigoplus_{I\in\Lambda(k)} H^{n-2-k}(Z_{f,I})\otimes\rnum(-k-1) \lrarr \\ \bigoplus_{I\in\Lambda(k-2)} H^{n-k+2}(D_{\Sigma_0,I})\otimes\rnum(-k+1) \oplus \bigoplus_{I\in\Lambda(k-1)} H^{n-k}(Z_{f,I})\otimes\rnum(-k) \end{multline} Here, $\rnum(i)$ denote the $i$-th Tate object in the category of Hodge structures. We can also obtain the complex () by using the theory of mixed Hodge modules , or more explicitly by using the theory of mixed Hodge complexes (see ). \vspace{.1in} We have the following pairings on $\nbigp_I$ on $H^{n-1-|I|}(D_{\Sigma_0,I})$ and $\nbigp_{Z_f,I}$ on $H^{n-|I|-2}(Z_{f,I})$: \[ \nbigp_I(a,b)= \left(\frac{1}{2\pi\sqrt{-1}}\right)^{n-1-|I|} \int_{D_{\Sigma_0,I}}ab, \quad\quad \nbigp_{Z_f,I}(a,b)= \left(\frac{1}{2\pi\sqrt{-1}}\right)^{n-2-|I|} \int_{Z_{f,I}}ab. \] For $i\in\Lambda$, let $p_i(f)$ be the pole order of $f$ along $D_i$. For any integer $\ell$, let $\epsilon(\ell):=(-1)^{\ell(\ell-1)/2}$. We set \[ \nbigptilde_I:= \epsilon(n-|I|-1)\cdot \prod_{i\in I}p_i(f)^{-1} \cdot \nbigp_I, \quad\quad \nbigptilde_{Z_f,I}:= \epsilon(n-2-|I|)\cdot \prod_{i\in I}p_i(f)^{-1}\cdot \nbigp_{Z_f,I}. \] We have the pairing of weight $n-|I|-1$ on $R_FH^{n-|I|-1}(D_{\Sigma_0,I})$ induced by $\nbigptilde_{I}$, and the pairing of weight $n-|I|-2$ on $R_FH^{n-|I|-2}(Z_{f,I})$ induced by $\nbigptilde_{Z_f,I}$. The induced pairings are also denoted by $\nbigptilde_I$ and $\nbigptilde_{Z_f,I}$ respectively. \begin{prop} We have $\nbigptilde_{n+k} =\bigoplus_{I\in\Lambda(k-1)} \nbigptilde_I \oplus \bigoplus_{I\in\Lambda(k)}\nbigptilde_{Z_f,I}$. \end{prop} \pf It follows from Proposition and Proposition . \hfill\qed \subsection{Mixed TEP-structures on reduced GKZ-hypergeometric systems} \subsubsection{Preliminary} As in \S, we consider $\nbiga=\{\veca_1,\ldots,\veca_m\}\subset\seisuu^n$ which generates $\seisuu^n$. We set $N_{\nbiga}:=\seisuu^n$ and $M_{\nbiga}:=\seisuu^m$. Let $e_1,\ldots,e_m$ be the standard basis of $M_{\nbiga}$. We have the surjective morphism $\Xi_{\nbiga}:M_{\nbiga}\lrarr N_{\nbiga}$. Let $L_{\nbiga}:=\Ker\Xi_{\nbiga}$. We obtain the exact sequence \[ \begin{CD} 0 @>>> L_{\nbiga} @>{\Theta_{\nbiga}}>> M_{\nbiga} @>{\Xi_{\nbiga}}>> N_{\nbiga} @>>> 0. \end{CD} \] By taking the dual, we obtain the exact sequence $0\lrarr N_{\nbiga}^{\lor} \stackrel{\Xi_{\nbiga}^{\lor}}\lrarr M_{\nbiga}^{\lor} \stackrel{\Theta_{\nbiga}^{\lor}}{\lrarr} L_{\nbiga}^{\lor}\lrarr 0$. For a finitely generated free abelian group $A$, we set $A_{\cnum^{\ast}}:= \cnum^{\ast}\otimes_{\seisuu} A$. We can naturally regard it as a complex algebraic variety or a complex manifold. Particularly, we set $S_{\nbiga}:=L^{\lor}_{\nbiga,\cnum^{\ast}}$. We have the natural surjection $\Theta^{\lor}_{\nbiga}: M^{\lor}_{\nbiga,\cnum^{\ast}} \lrarr S_{\nbiga}$. We have the action $\rho_1$ of $N^{\lor}_{\nbiga,\cnum^{\ast}}$ on $M^{\lor}_{\nbiga,\cnum^{\ast}}$ induced by $-\varphi$. We also have the natural action $\rho_0$ of $N^{\lor}_{\nbiga,\cnum^{\ast}}$ on itself by the multiplication. We can naturally regard $\Theta^{\lor}_{\nbiga}: M^{\lor}_{\nbiga,\cnum^{\ast}} \lrarr S_{\nbiga}$ as the quotient of the projection $N^{\lor}_{\nbiga,\cnum^{\ast}}\times M^{\lor}_{\nbiga,\cnum^{\ast}} \lrarr M^{\lor}_{\nbiga,\cnum^{\ast}}$ via the above actions of $N^{\lor}_{\nbiga,\cnum^{\ast}}$. The identifications $M_{\nbiga}=\seisuu^m$ and $N_{\nbiga}=\seisuu^n$ induce the coordinate systems $(w_1,\ldots,w_m)$ on $M_{\nbiga,\cnum^{\ast}}^{\lor}$ and $(t_1,\ldots,t_n)$ on $N^{\lor}_{\nbiga,\cnum^{\ast}}$. We set $G_{\nbiga}:=\sum_{i=1}^m w_i$ on $M^{\lor}_{\nbiga,\cnum^{\ast}}$. The algebraic function $F_{\nbiga}=\sum_{i=1}^m w_i t^{\veca_i}$ on $N^{\lor}_{\nbiga,\cnum^{\ast}} \times M^{\lor}_{\nbiga,\cnum^{\ast}}$ is $N^{\lor}_{\nbiga,\cnum^{\ast}}$-invariant, and $G_{\nbiga}$ is the descent of $F_{\nbiga}$. Any splitting $\gamma_{\nbiga}: L_{\nbiga}^{\lor}\lrarr M_{\nbiga}^{\lor}$ of $\Theta_{\nbiga}^{\lor}$ induces an algebraic morphism $\gamma_{\nbiga}: L_{\nbiga,\cnum^{\ast}}^{\lor} \lrarr M^{\lor}_{\nbiga,\cnum^{\ast}}$, and we obtain $\id\times\gamma_{\nbiga}: N^{\lor}_{\nbiga,\cnum^{\ast}} \times L^{\lor}_{\nbiga,\cnum^{\ast}} \lrarr N^{\lor}_{\nbiga,\cnum^{\ast}} \times M^{\lor}_{\nbiga,\cnum^{\ast}}$. The splitting $\gamma_{\nbiga}$ also gives an isomorphism $M^{\lor}_{\nbiga} \simeq N^{\lor}_{\nbiga}\times L^{\lor}_{\nbiga}$, and hence $M^{\lor}_{\nbiga,\cnum^{\ast}} \simeq N^{\lor}_{\nbiga,\cnum^{\ast}} \times L^{\lor}_{\nbiga,\cnum^{\ast}}$. The pull back of $F_{\nbiga}$ by $\id\times\gamma_{\nbiga}$ is equal to $G_{\nbiga}$ under the identification. \subsubsection{Mixed twistor $\nbigd$-modules and the induced mixed TEP-structures} We naturally regard $M_{\nbiga,\cnum^{\ast}}^{\lor} =(\cnum^{\ast})^m =\bigl\{ \sum_{i=1}^m\alpha_iz_i\,\big|\, \alpha_i\neq 0 \bigr\}$, and $N^{\lor}_{\nbiga,\cnum^{\ast}}=T^n$ in \S. Let $\gamma_{\nbiga}: L_{\nbiga,\cnum^{\ast}}^{\lor} \lrarr M_{\nbiga,\cnum^{\ast}}^{\lor}$ be the splitting as above. Note that it is non-degenerate at $\infty$ for $X_{\nbiga}$, as remarked in Example . We obtain the following integrable mixed twistor $\nbigd$-modules with real structure on $S_{\nbiga}$: \[ \nbigt_{\nbiga,\star}:= \pi_{\dagger}^0 \nbigt_{\star}\bigl(\nbiga, S_{\nbiga},\gamma_{\nbiga}\bigr) \quad (\star=\ast,!). \] The underlying $\nbigd$-modules are the reduced GKZ-hypergeometric systems. As in Lemma , $\nbigt_{\nbiga,\star}$ are independent of the choice of a splitting $\gamma_{\nbiga}$. We set $\nbigt_{\nbiga,\min} :=\pi_{\dagger}^0 \nbigt_{\min}(\nbiga,S_{\nbiga},\gamma_{\nbiga})$. By Corollary , it is isomorphic to the image of the natural morphism $\nbigt_{\nbiga,!} \lrarr \nbigt_{\nbiga,\ast}$, and $\Gr^W_{m}\nbigt_{\nbiga,\ast} =\Gr^W_m\nbigt_{\nbiga,!} =\Gr^W_m\nbigt_{\nbiga,\min} =\nbigt_{\nbiga,\min}$. If $0$ is contained in the interior part of $\Conv(\nbiga)$, we have $\nbigt_{\nbiga!} =\nbigt_{\nbiga,\ast} =\nbigt_{\nbiga,\min}$. We set $\nbigl_{\nbiga,\star}:=\nbigl_{\star}\bigl(\nbiga, S_{\nbiga},\gamma_{\nbiga}\bigr)$ and $\nbigv_{\nbiga,\star}:= \pi_{\dagger}^0\nbigl_{\nbiga,\star}$. The $\nbigrtilde$-modules $\nbigv_{\nbiga,\star}$ underlie $\nbigt_{\nbiga,\star}$. More precisely, $\nbigt_{\nbiga,\star}$ is expressed as a pair of $\nbigr$-modules $\lambda^{m}\nbigv_{\nbiga,!}$ and $\nbigv_{\nbiga,\ast}$ with the induced sesqui-linear pairing, and $\nbigt_{\nbiga,!}$ is expressed as a pair of $\nbigr$-modules $\lambda^{m}\nbigv_{\nbiga,\ast}$ and $\nbigv_{\nbiga,!}$ with the induced sesqui-linear pairing. Let $W$ denote the filtration of $\nbigv_{\nbiga,\star}$ underlying the weight filtration of $\nbigt_{\nbiga,\star}$. We set $\Wtilde_{k}\nbigv_{\nbiga,\star}:= W_{k+n-m}\nbigv_{\nbiga,\star}$. Note $\dim S_{\nbiga}=m-n$. Let $\nbigl_{\nbiga,\min}$ be the image of $\nbigl_{\nbiga,!}\lrarr\nbigl_{\nbiga,\ast}$, and set $\nbigv_{\nbiga,\min}:= \pi_{\dagger}^0\nbigl_{\nbiga,\min}$. It is naturally isomorphic to the image of $\nbigv_{\nbiga,!}\lrarr\nbigv_{\nbiga,\ast}$ by Corollary . The $\nbigrtilde$-module $\nbigv_{\nbiga,\min}$ underlies $\nbigt_{\nbiga\min}$. We have $\Gr^{\Wtilde}_n\nbigv_{\nbiga,!} \simeq \Gr^{\Wtilde}_n\nbigv_{\nbiga,\ast} \simeq \nbigv_{\nbiga,\min}$. If $0$ is an interior point of $\Conv(\nbiga)$, then $\nbigt_{\nbiga\ast}=\nbigt_{\nbiga,\min}=\nbigt_{\nbiga,!}$ is pure of weight $m$, and it is equipped with the canonical polarization. As explained in \S, even if $0$ is not an interior point of $\Conv(\nbiga)$, we have the graded polarizations $\vecnbigs_{\nbiga,\star,\vecb}$ of $\nbigt_{\nbiga,\star}$ depending on the choice of $\vecb\in\seisuu^n$ such that $0$ is an interior point of $\Conv(\nbiga\cup\{\vecb\})$. The weight $m$-part of $\vecnbigs_{\nbiga,\star,\vecb}$ are independent of $\vecb$. Let $S_{\nbiga}^{\reg}:=\gamma_{\nbiga}^{-1}(U_{\nbiga}^{\reg})$. The restriction $\nbigv_{\nbiga,\star|S_{\nbiga}^{\reg}}$ are locally free $\nbigo_{\cnum_{\lambda}\times S_{\nbiga}^{\reg}}$-modules. Take a hypersurface $Y$ such that $Y\supset S_{\nbiga}\setminus S_{\nbiga}^{\reg}$. The real structure and the graded polarization of $\nbigt_{\nbiga,\star}$ induce graded pairing $P_{\nbiga,\star,\vecb}$ of $(\nbigv_{\nbiga,\star}(\ast Y),\Wtilde)$. In this way, we obtain mixed TEP-structures $(\nbigv_{\nbiga,\star}(\ast Y),\Wtilde, P_{\nbiga,\star,\vecb})$. \begin{rem} Let $\cnum[M_{\nbiga}]$ denote the group ring of $M_{\nbiga}$ over $\cnum$. We may naturally regard $M^{\lor}_{\nbiga,\cnum^{\ast}}$ as the algebraic variety $\Spec \cnum[M_{\nbiga}]$. Then, $\nbigt_{\nbiga,\star}$ is algebraic in the sense of {\rm\S}. \hfill\qed \end{rem} \begin{rem} If $0$ is contained in the interior part of $\Conv(\nbiga)$, we shall often omit the subscripts $\ast,!$ because $\nbigt_{\nbiga!}=\nbigt_{\nbiga\ast}=\nbigt_{\nbiga,\min}$. \hfill\qed \end{rem} \subsubsection{Homogeneity} Let $e_1^{\lor},\ldots,e_m^{\lor}$ denote the dual basis of $M_{\nbiga}^{\lor}$. We have the morphism $\seisuu\lrarr M_{\nbiga}^{\lor}$ given by $1\longmapsto \gminiv:=\sum_{i=1}^m e_i^{\lor}$. It induces a $\cnum^{\ast}$-action on $M^{\lor}_{\nbiga,\cnum^{\ast}}$. By the composition $\seisuu\lrarr M^{\lor}_{\nbiga} \lrarr L^{\lor}_{\nbiga}$, we obtain a $\cnum^{\ast}$-action on $S_{\nbiga}$. The map $\Theta_{\nbiga}^{\lor}: M^{\lor}_{\nbiga,\cnum^{\ast}}\lrarr S_{\nbiga}$ is $\cnum^{\ast}$-equivariant. For the $\cnum^{\ast}$-action, we have $a^{\ast}G_{\nbiga}=a\cdot G_{\nbiga}$ for any $a\in \cnum^{\ast}$. We consider the action of $\cnum^{\ast}$ on $\cnum_{\lambda}$ given by the multiplication. For the diagonal $\cnum^{\ast}$-action on $\cnum_{\lambda}\times M^{\lor}_{\nbiga,\cnum^{\ast}}$, we have $a^{\ast}(\lambda^{-1}G_{\nbiga}) =\lambda^{-1}G_{\nbiga}$ for any $a\in\cnum^{\ast}$. \begin{lem} The $\nbigrtilde$-modules $\nbigv_{\nbiga,\star}$ on $S_{\nbiga}$ are $\cnum^{\ast}$-homogeneous in the sense of {\rm\S}. \end{lem} \pf Take any toric desingularization $\varphi_{\Sigma}:X_{\Sigma}\lrarr X_{\nbiga}$. We have the $\cnum^{\ast}$-action on $X_{\Sigma}\times S_{\nbiga}$ which is the extension of $T^n\times S_{\nbiga}\simeq M^{\lor}_{\nbiga,\cnum^{\ast}}$. For the action, we have $a^{\ast}(\lambda^{-1}F_{\gamma_{\nbiga},\Sigma}) =\lambda^{-1}F_{\gamma_{\nbiga},\Sigma}$. We set $Y:=X_{\Sigma}\times S_{\nbiga}$ and $D_Y:=D_{\Sigma}\times S_{\nbiga}$. By Proposition , the $\nbigrtilde$-modules $\nbigl_{\star}(F_{\gamma_{\nbiga},\Sigma},D_{Y})$ are $\cnum^{\ast}$-homogeneous. Then, $\nbigv_{\nbiga,\star} \simeq \pi^0_{\Sigma\dagger} \nbigl_{\star}(F_{\gamma_{\nbiga},\Sigma},D_Y)$ are also $\cnum^{\ast}$-homogeneous. It is easy to check that the $\cnum^{\ast}$-actions are independent of the choice of $\varphi_{\Sigma}$. \hfill\qed \subsubsection{Variation of Hodge structure} Let us consider the special case $\Theta_{\nbiga}^{\lor}(\gminiv)=0$, i.e., the natural $\cnum^{\ast}$-action on $S_{\nbiga}$ is trivial. By the general result in Theorem , we have the following. \begin{prop} If $\Theta_{\nbiga}^{\lor}(\gminiv)=0$, then the integrable mixed twistor $\nbigd$-modules $\nbigt_{\nbiga,\star}$ come from mixed $\real$-Hodge modules. \hfill\qed \end{prop} \begin{rem} In particular, the underlying $\nbigd$-modules of $\nbigt_{\nbiga,\star}$ are regular singular. Note that the condition $\Theta^{\lor}_{\nbiga}(\gminiv)=0$ is equivalent to the standard criterion for the GKZ-system to be regular singular. Indeed, we have $\Theta_{\nbiga}^{\lor}(\gminiv)=0$ if and only if there exists $\alpha\in N_{\nbiga}^{\lor}$ such that $\alpha(\Xi_{\nbiga}(e_i))=1$ $(i=1,\ldots,m)$. If $\Theta_{\nbiga}^{\lor}(\gminiv)=0$, we have $\alpha\in N_{\nbiga}^{\lor}$ such that $\Xi_{\nbiga}^{\lor}(\alpha)(\gminiv)$, and hence we have $\langle \alpha,\Xi_{\nbiga}(e_i) \rangle =\langle \Xi_{\nbiga}^{\lor}(\alpha),e_i \rangle =\langle\gminiv,e_i\rangle =1$. Conversely, suppose that there exists $\alpha\in N_{\nbiga}^{\lor}$ such that $\alpha(\Xi_{\nbiga}(e_i))=1$. Because $\langle \Xi_{\nbiga}^{\lor}(\alpha),e_i \rangle=1$, we have $\Xi_{\nbiga}^{\lor}(\alpha)=\gminiv$. \hfill\qed \end{rem} \begin{cor} $\nbigt_{\nbiga,\star|S^{\reg}_{\nbiga}}$ comes from a graded polarizable variation of mixed Hodge structure on $S^{\reg}_{\nbiga}$. For any algebraic embedding $S^{\reg}_{\nbiga}\subset Z$, the graded polarizable variation of mixed Hodge structure is admissible along $Z\setminus S^{\reg}_{\nbiga}$. \end{cor} \pf It follows from the algebraicity in Proposition and a general property of mixed twistor $\nbigd$-modules. \hfill\qed \subsubsection{Appendix: Comparison with the construction of Reichelt-Sevenheck} As in \S, we set $V:=H^0(\proj^m,\nbigo(1))$. Let $Z\subset\proj^m\times V$ be the $0$-set of the universal section of $\nbigo_{\proj^m}(1)\boxtimes\nbigo_V$. We decompose $V=V_1\times V_2$, where $V_1:=\bigl\{\alpha_0z_0\,\big|\,\alpha_0\in\cnum\bigr\}$ and $V_2:=\bigl\{\sum_{i=1}^m\alpha_iz_i\,\big|\,\alpha_i\in\cnum\bigr\}$. We identify $V_2=M_{\nbiga}^{\lor}\otimes\cnum$. We have the splitting $\gamma_{\nbiga}: S_{\nbiga}=L_{\nbiga}^{\lor}\otimes\cnum^{\ast} \lrarr M_{\nbiga}^{\lor}\otimes\cnum^{\ast} \subset V_2$. We obtain $\id\times\gamma_{\nbiga}: V_1\times S_{\nbiga} \lrarr V_1\times V_2$. Let $Z_{\nbiga}$ be the fiber product of $Z$ and $V_1\times S_{\nbiga}$ over $V_1\times V_2$. We have the naturally induced morphisms $q_1:Z_{\nbiga}\lrarr \proj^m$ and $q_2:Z_{\nbiga}\lrarr V_1\times S_{\nbiga}$. Recall that we have the morphism $\psi_{\nbiga}:T^n\lrarr \proj^m$ induced by $\nbiga$. Let $U:=\psi_{\nbiga}(T^n)$. We set $Z_{\nbiga,U}:=Z_{\nbiga}\times_{\proj^m}U$. Let $\iota_{Z_{\nbiga,U}}: Z_{\nbiga,U}\lrarr \proj^m\times (V_1\times S_{\nbiga})$ be the inclusion. Let us consider the pure Hodge module $\bigl( \nbigo_{Z_{\nbiga,U}},F \bigr)$ and the mixed Hodge modules $\iota_{Z_{\nbiga,U}\star}\bigl( \nbigo_{Z_{\nbiga,U}},F \bigr)$ $(\star=\ast,!)$ as in \S. Let $\nbigm^{IC}(Z_{\nbiga,U})$ denote the image of the morphism $\iota_{Z_{\nbiga,U}!}\bigl( \nbigo_{Z_{\nbiga,U}},F \bigr) \lrarr \iota_{Z_{\nbiga,U}\ast}\bigl( \nbigo_{Z_{\nbiga,U}},F \bigr)$. Let $\pi_{V_1\times S_{\nbiga}}: \proj^m\times (V_1\times S_{\nbiga}) \lrarr V_1\times S_{\nbiga}$ be the projection. Then, we obtain the mixed Hodge modules $\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}\star}\bigl( \nbigo_{Z_{\nbiga,U}},F \bigr)$ $(\star=!,\ast)$ and $\pi^0_{V_1\times S_{\nbiga}\dagger} \nbigm^{IC}(Z_{\nbiga,U})$. By applying the procedure in \S, we obtain the following $\nbigrtilde_{S_{\nbiga}}$-modules: \[ G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}\star} (\nbigo_{Z_{\nbiga,U}},F)\bigr), \quad\quad G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \nbigm^{IC}(Z_{\nbiga,U})\bigr). \] \begin{prop} We have isomorphisms of $\nbigrtilde_X$-modules: \[ \lambda \cdot \nbigv_{\nbiga\star} \simeq G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}\star} (\nbigo_{Z_{\nbiga,U}},F) \bigr) \] We also have the following commutative diagram: \[ \begin{CD} \lambda\cdot\nbigv_{\nbiga!} @>{\simeq}>> G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}!} (\nbigo_{Z_{\nbiga,U}},F)\bigr) \\ @VVV @VVV \\ \lambda\cdot\nbigv_{\nbiga\ast} @>{\simeq}>> G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}\ast} (\nbigo_{Z_{\nbiga,U}},F)\bigr) \end{CD} \] \end{prop} \pf We can obtain the claim from Proposition by using the non-characteristic pull back. We can also prove it directly by the argument in the proof of Proposition . \hfill\qed \begin{cor} We naturally have $\lambda\cdot\nbigv_{\nbiga,\min} \simeq G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \nbigm^{IC}(Z_{\nbiga,U})\bigr)$. \end{cor} \pf Reichelt-Sevenheck proved that $G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \nbigm^{IC}(Z_{\nbiga,U})\bigr)$ is naturally isomorphic to the image of $G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}!} (\nbigo_{Z_{\nbiga,U}},F)\bigr) \lrarr G_0\FL^{\loc}_{S_{\nbiga}} \bigl(\pi^0_{V_1\times S_{\nbiga}\dagger} \iota_{Z_{\nbiga,U}\ast} (\nbigo_{Z_{\nbiga,U}},F)\bigr)$. As mentioned in \S, $\nbigv_{\nbiga\min}$ is naturally isomorphic to the image of $\nbigv_{\nbiga!}\lrarr\nbigv_{\nbiga\ast}$ by Corollary . Hence, we have the desired isomorphism. \hfill\qed \subsection{Some mixed twistor $\nbigd$-modules in mirror symmetry} \subsubsection{Toric data} Let $X$ be an $n$-dimensional smooth weak Fano toric variety. Let $\Sigma$ be a fan of $X$. Let $\Sigma(1)=\{\rho_1,\ldots,\rho_m\}$ denote the set of the $1$-dimensional cones in $\Sigma$. Let $[\rho_i]\in\seisuu^n$ be the primitive generator of $\rho_i\cap\seisuu^n$. Let $K_X$ denote the canonical bundle of $X$. Let $\nbigl_j$ $(j=1,\ldots,r)$ be nef line bundles on $X$ such that $\bigl( K_X\otimes \bigotimes_{j=1}^{r}\nbigl_j \bigr)^{\lor}$ is nef. We may assume that $\nbigl_j=\nbigo(\sum_{i=1}^m\beta_{ji}D_i)$ $(j=1,\ldots,r)$ for some $\beta_{ji}\in\seisuu_{\geq 0}$, where $D_i$ are the hypersurfaces corresponding to the cones $\rho_i$. Let $\seisuu^{n+r}=\seisuu^n\oplus\seisuu^r$. Let $n_1,\ldots,n_r$ denote the standard basis of $\seisuu^r$. We set $\veca_i\in\seisuu^n\oplus\seisuu^r$ as follows: \[ \veca_i:= \left\{ \begin{array}{ll} [\rho_i]+\sum_{j=1}^r\beta_{ji}n_j & (i=1,\ldots,m),\\ n_{i-m} & (i=m+1,\ldots,m+r). \end{array} \right. \] We put $\nbiga:=\{\veca_1,\ldots,\veca_{r+m}\}$. We also set $\veca_{r+m+1}:=-\sum_{j=1}^r n_j$ and $\nbigatilde:=\nbiga\cup\{\veca_{r+m+1}\}$. Note that $0$ is contained in the interior part of the convex hull of $\Conv(\nbigatilde)$. (We shall use $\nbigatilde$ in \S.) \subsubsection{$\nbigrtilde$-modules with filtration and graded pairing associated to $\nbiga$} Applying the construction in \S, we obtain the integrable mixed twistor $\nbigd$-modules $\nbigt_{\nbiga,\star}$ $(\star=\ast,!)$ with real structure on $S_{\nbiga}$. They are equipped with the graded polarization $\vecnbigs_{\nbiga,\star,\veca_{r+m+1}}$. Let $\nbigv_{\nbiga,\star}$ denote their underlying $\nbigrtilde$-modules. They are $\cnum^{\ast}$-homogeneous as in \S. By the procedure in \S--, we obtain the filtration $\Wtilde$ and the graded pairing $(P_{\nbiga,\star,k}\,|\,k\in\seisuu)$ on $\nbigv_{\nbiga,\star|S_{\nbiga}^{\reg}}$. \vspace{.1in} We apply the same construction to $\nbigt_{\nbiga,\star}\otimes\newTate(n+r)$. Let $\gbigv_{\nbiga,\star}$ be the underlying $\nbigrtilde$-modules. They are $\cnum^{\ast}$-homogeneous. The restriction $\gbigv_{\nbiga,\star|S_{\nbiga}^{\reg}}$ is equipped with the filtration $\Wtilde$ and the graded pairings $\gbigp_{\nbiga,\star}=(\gbigp_{\nbiga,\star,k}\,\big|\,k\in\seisuu)$. Thus, we obtain mixed TEP-structures $\bigl( \gbigv_{\nbiga,\star|S_{\nbiga}^{\reg}}, \Wtilde, \gbigp_{\nbiga,\star} \bigr)$ on $S_{\nbiga}^{\reg}$. By construction, we have $\gbigv_{\nbiga,\star}= \lambda^{n+r}\nbigv_{\nbiga,\star}$ and $\Wtilde_k\gbigv_{\nbiga,\star}= \lambda^{n+r} \Wtilde_{k+2(n+r)}\nbigv_{\nbiga,\star}$. We have $\Gr^{\Wtilde}_k\gbigv_{\nbiga,\star} =\lambda^{n+r} \Gr^{\Wtilde}_{k+2(n+r)}\nbigv_{\nbiga,\star}$. The graded pairings on $\Gr^{\Wtilde}_k\gbigv_{\nbiga,\star}(\ast\lambda) =\Gr^{\Wtilde}_{k+2(n+r)}\nbigv_{\nbiga,\star}(\ast\lambda)$ are equal. \subsubsection{Expression as the systems of differential equations} Let $[z_0:\cdots:z_{m+r}]$ be the homogeneous coordinate system on $\proj^{m+r}$. We set $W:= \bigl\{\sum_{i=1}^{m+r}\alpha_iz_i \bigr\} \subset H^0\bigl(\proj^{m+r},\nbigo(1)\bigr) =M^{\lor}_{\nbiga}\otimes\cnum$. The splitting $S_{\nbiga}=L^{\lor}_{\nbiga}\otimes\cnum^{\ast} \lrarr M^{\lor}_{\nbiga}\otimes\cnum^{\ast}$ induces a morphism $\gamma_{\nbiga}:S_{\nbiga}^{\reg}\lrarr W$. We set $W^{\reg}:=W\cap U_{\nbiga}^{\reg}$. We have $\gamma_{\nbiga}(S^{\reg}_{\nbiga})\subset W^{\reg}$. We use the notation in \S. By applying the construction in \S to $\nbiga=\{\veca_{1},\ldots,\veca_{m+r}\}\subset\seisuu^{n+r}$ with $\vecbeta=0\in\cnum^{n+r}$, we obtain the $\nbigrtilde_{\cnum^{m+r}}$-modules $\nbigm^{\GKZ}(\nbiga,K(\nbiga),0)$ and $\nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0)$ on $W=\cnum^{m+r}=\{(\alpha_1,\ldots,\alpha_{m+r})\}$. We have the induced morphism $\id\times\gamma_{\nbiga}: \cnum_{\lambda}\times S_{\nbiga}^{\reg} \lrarr \cnum_{\lambda}\times W^{\reg}$. The restrictions of $\nbigm^{\GKZ}(\nbiga,K(\nbiga),0)$ and $\nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0)$ to $W^{\reg}$ give locally free $\nbigo_{\cnum\times W^{\reg}}$-modules. We take the pull back of them by $\id\times\gamma_{\nbiga}$ as $\nbigo$-modules. Then, $(\id\times\gamma_{\nbiga})^{\ast} \nbigm^{\GKZ}(\nbiga,K(\nbiga),0)$ and $(\id\times\gamma_{\nbiga})^{\ast} \nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0)$ are naturally equipped with the meromorphic flat connection with which they are $\nbigrtilde_{S_{\nbiga}^{\reg}}$-modules. \begin{thm} We have the following commutative diagram. \[ \begin{CD} \lambda^n(\id\times\gamma_{\nbiga})^{\ast} \nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) @>{\simeq}>> \lambda^{-r}\gbigv_{\nbiga!} \\ @VVV @VVV \\ \lambda^n(\id\times\gamma_{\nbiga})^{\ast} \nbigm^{\GKZ}(\nbiga,K(\nbiga),0) @>{\simeq}>> \lambda^{-r}\gbigv_{\nbiga\ast} \end{CD} \] \end{thm} \pf By Proposition and Proposition , we obtain the following commutative diagram: \begin{equation} \begin{CD} (\id\times\gamma_{\nbiga})^{\ast} \nbigm^{\GKZ}\bigl(\nbiga,K(\nbiga)^{\circ},0 \bigr) @>{\simeq}>> \nbigv_{\nbiga!}\\ @VVV @VVV \\ (\id\times\gamma_{\nbiga})^{\ast} \nbigm^{\GKZ}(\nbiga,K(\nbiga),0) @>{\simeq}>> \nbigv_{\nbiga\ast} \end{CD} \end{equation} Then, we obtain the claim of the theorem by construction. \hfill\qed \vspace{.1in} According to , we have $K(\nbiga)=\seisuu_{\geq 0}\nbiga$ and $K(\nbiga)^{\circ}=K(\nbiga)+\sum_{i=1}^r\veca_{n+i}$. We set $\vecc_1=\sum_{i=1}^r\veca_{n+i}$. Let $\nbigi(\nbiga,0)$ and $\nbigi(\nbiga,-\vecc_1)$ be the left ideals of $\nbigrtilde_W$ given as in \S. As remarked in \S, we have the following commutative diagram on $W=\cnum^{m+r}$: \[ \begin{CD} \nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) @>{\simeq}>> \nbigrtilde_{W}/\nbigi(\nbiga,-\vecc_1) \\ @VVV @VVV \\ \nbigm^{\GKZ}(\nbiga,K(\nbiga),0) @>{\simeq}>> \nbigrtilde_{W}\big/\nbigi(\nbiga,0) \end{CD} \] Here, the right vertical arrow is induced by the multiplication of $\prod_{i=1}^r(\lambda\del_{m+i})$. \subsubsection{Relation with quantum $\nbigd$-modules of complete intersections (Appendix)} Inspired by the work of E. Mann and T. Mignon , Reichelt and Sevenheck {\rm} constructed the $\nbigrtilde$-modules ${}^{\ast}_0\widehat{\nbign}^{(0,\underline{0},\underline{0})}_{\nbiga}$ and ${}^{\ast}_0\widehat{\nbigm}^{(-r,\underline{0},\underline{0})}_{\nbiga}$ on $W\setminus\{0\}$ in terms of the differential systems. (See {\rm} for the precise definition of the $\nbigrtilde$-modules. But, we remark the commutativity () below.) See and for the precise relation of the $\nbigrtilde$-modules and the reduced quantum $\nbigd$-module of complete intersection of general sections of $\nbigl_i$ $(i=1,\ldots,r)$. By formal computations, we can obtain the following commutative diagram on $(\cnum^{\ast})^{m+r}=\bigl\{ (\alpha_1,\ldots,\alpha_{m+r})\,\big|\, \alpha_i\neq 0 \bigr\}$: \begin{equation} \begin{CD} \bigl( \lambda^n \nbigrtilde_{W}\big/ \nbigi(\nbiga,-\vecc_1) \bigr)_{|(\cnum^{\ast})^{m+r}} @>{\simeq}>> {}^{\ast}_0\widehat{\nbign}^{(0,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}} \\ @VVV @VVV \\ \bigl( \lambda^n \nbigrtilde_{W}\big/ \nbigi(\nbiga,0) \bigr)_{|(\cnum^{\ast})^{m+r}} @>{\simeq}>> {}^{\ast}_0\widehat{\nbigm}^{(-r,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}} \end{CD} \end{equation} The vertical arrows are the natural morphisms. \begin{cor} We have the following commutative diagram: \[ \begin{CD} \lambda^{-r}\gbigv_{\nbiga!} @>{\simeq}>> (\id\times\gamma_{\nbiga})^{\ast}\bigl( {}^{\ast}_0\widehat{\nbign}^{(0,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}}\bigr) \\ @VVV @VV{a_1}V \\ \lambda^{-r}\gbigv_{\nbiga\ast} @>{\simeq}>> (\id\times\gamma_{\nbiga})^{\ast}\bigl( {}^{\ast}_0\widehat{\nbigm}^{(-r,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}}\bigr) \end{CD} \] In particular, the $\nbigrtilde$-modules $(\id\times\gamma_{\nbiga})^{\ast}\bigl( {}^{\ast}_0\widehat{\nbign}^{(0,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}}\bigr)$ and $(\id\times\gamma_{\nbiga})^{\ast}\bigl( {}^{\ast}_0\widehat{\nbigm}^{(-r,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}}\bigr)$ underlie mixed twistor $\nbigd$-modules, and the $\Image a_1$ underlies a pure twistor $\nbigd$-module. \hfill\qed \end{cor} \begin{rem} In {\rm}, Reichelt and Sevenheck constructed $\nbigrtilde$-modules by using the partial Fourier-Laplace transform of GKZ-hypergeometric systems and the Brieskorn lattices associated to the Hodge filtrations of the GKZ-hypergeometric systems. They conjectured in {\rm\cite[Conjecture {\rm 6.13}]{Reichelt-Sevenheck2}} that the $\nbigrtilde$-modules are isomorphic to $(\id\times\gamma_{\nbiga})^{\ast}\bigl( {}^{\ast}_0\widehat{\nbign}^{(0,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}}\bigr)$ and $(\id\times\gamma_{\nbiga})^{\ast}\bigl( {}^{\ast}_0\widehat{\nbigm}^{(-r,\underline{0},\underline{0})} _{\nbiga|(\cnum^{\ast})^{m+r}}\bigr)$. In {\rm \S}, we review their construction of $\nbigrtilde$-modules. In Proposition {\rm}, we compare their $\nbigrtilde$-modules with the underlying $\nbigrtilde$-modules of mixed twistor $\nbigd$-modules in a general setting. See also Proposition {\rm}. So, Theorem {\rm } also verifies their conjecture. Reichelt and Sevenheck also verified their conjecture {\rm } in a different way. \hfill\qed \end{rem} \begin{rem} In the first version of this preprint, we used some results in {\rm} for the comparison with some $\nbigrtilde$-modules associated to differential systems, which required us to go a roundabout way with careful comparison of dualities. In the second version, as explained \S, we give a direct comparison of $\nbigrtilde$-modules associated to differential systems and $\nbigrtilde$-modules underlying mixed twistor $\nbigd$-modules. Hence, the argument is made more transparent. \hfill\qed \end{rem} \subsection{Description as a specialization} This subsection is the continuation of \S--\S, and it is the preparation for Theorem . \subsubsection{The $\nbigrtilde$-module with pairing associated to $\nbigatilde$} Applying the construction in \S, we obtain the integrable pure twistor $\nbigd$-module $\nbigt_{\nbigatilde}$ on $S_{\nbigatilde}$. Let $\nbigv_{\nbigatilde}$ denote the underlying $\nbigrtilde$-module which is $\cnum^{\ast}$-homogeneous. We have the filtration $\Wtilde$ and the graded pairing $P_{\nbigatilde}$ induced by the real structure and the graded polarization on $\nbigv_{\nbigatilde}$. In this case, $\Wtilde$ is pure of weight $n+r$. We apply the same construction to $\nbigt_{\nbigatilde}\otimes\newTate(n+r)$. We obtain the underlying $\nbigrtilde$-module $\gbigv_{\nbigatilde}$ on $S_{\nbigatilde}$ which is $\cnum^{\ast}$-homogeneous. The restriction $\gbigv_{\nbigatilde|S^{\reg}_{\nbigatilde}}$ is equipped with the filtration $\Wtilde$ which is pure of weight $-n-r$, and the pairing $\gbigp_{\nbigatilde}$ of weight $-n-r$. In other words, we obtain a ${\rm TEP}(n+r)$-structure $\bigl(\gbigv_{\nbigatilde|S^{\reg}_{\nbigatilde}}, \gbigp_{\nbigatilde} \bigr)$. By construction, we have $\gbigv_{\nbigatilde}= \lambda^{n+r}\nbigv_{\nbigatilde}$, and $P_{\nbigatilde}=\gbigp_{\nbigatilde}$ on $\nbigv_{\nbigatilde}(\ast\lambda) =\gbigv_{\nbigatilde}(\ast\lambda)$. \subsubsection{Relation of $\gbigv_{\nbiga}$ and $\gbigv_{\nbigatilde}$} We set $\Stilde_{\nbiga}:=S_{\nbiga}\times\cnum_{\tau}$. Applying the construction in \S, we obtain the integrable pure twistor $\nbigd$-module $\nbigt(\nbigatilde,\Stilde_{\nbiga},\gammatilde_{\nbiga})$ and the underlying $\nbigrtilde$-module $\nbigl(\nbigatilde,\Stilde_{\nbiga},\gammatilde_{\nbiga})$ on $\proj^{m+r+1}\times \Stilde_{\nbiga}$. Let $\pitilde: \proj^{m+r+1}\times\Stilde_{\nbiga} \lrarr \Stilde_{\nbiga}$ be the projection. We set $\gbigvtilde_{\nbiga}:= \lambda^{n+r}\pitilde_{\dagger}^0 \nbigl(\nbigatilde,\Stilde_{\nbiga},\gammatilde_{\nbiga})(\ast\tau)$. It is equipped with the pairing $\gbigptilde_{\nbiga}$ of weight $-n-r$ as in the case of \S. \begin{lem} We can naturally identify $S_{\nbigatilde}$ with $S_{\nbiga}\times\cnum_{\tau}^{\ast} \subset \Stilde_{\nbiga}$ under which $\gbigvtilde_{\nbiga|S_{\nbigatilde}} =\gbigv_{\nbigatilde}$. In other words, $\gbigvtilde_{\nbiga}$ is a meromorphic extension of $\gbigv_{\nbigatilde}$ on $\Stilde_{\nbiga}$. We also have the $\cnum^{\ast}$-action on $\Stilde_{\nbiga}$ for which the inclusions $S_{\nbiga}\simeq S_{\nbiga}\times\{0\}\subset\Stilde_{\nbiga}$ and $S_{\nbigatilde}\lrarr \Stilde_{\nbiga}$ are equivariant. \end{lem} \pf We use the notation in \S. We naturally have $M_{\nbigatilde}=M_{\nbiga}\oplus\seisuu e_{m+r+1}$ and $N_{\nbigatilde}=N_{\nbiga}$. The natural inclusion $M_{\nbiga}\lrarr M_{\nbigatilde}$ induces $L_{\nbiga}\lrarr L_{\nbigatilde}$, and hence $L_{\nbigatilde}^{\lor}\lrarr L_{\nbiga}^{\lor}$. We also consider $M_X:=\seisuu^m$ and $N_X:=\seisuu^n$ and a morphism $\Theta_X:M_X\lrarr N_X$ determined by $\Theta_X(e_i)=[\rho_i]$, where the tuple $e_1,\ldots,e_m$ is the standard basis of $M_X$, and $[\rho_i]$ are as in \S. Let $L_X$ be the kernel of $\Theta_X$. The projections $M_{\nbigatilde}\lrarr M_X$ and $N_{\nbigatilde}\lrarr N_X$ induces $L_{\nbigatilde}\lrarr L_X$. It is well known and easy that the composite of $L_{\nbiga}\lrarr L_{\nbigatilde}\lrarr L_X$ is an isomorphism. Let $\Upsilon$ denote the image of $e_{m+r+1}^{\lor}$ by $M_{\nbigatilde}^{\lor}\lrarr L_{\nbigatilde}^{\lor}$. Then, $\Upsilon$ is the frame of the kernel of $L_{\nbigatilde}^{\lor}\lrarr L_{\nbiga}^{\lor}$. Hence, $L_{\nbigatilde}^{\lor}$ is the direct sum of the image of $L_X^{\lor}$ and $\seisuu\Upsilon$. The direct sum induces \[ S_{\nbigatilde} =L^{\lor}_{\nbigatilde}\otimes\cnum^{\ast} =\bigl( L^{\lor}_{\nbiga}\otimes\cnum^{\ast} \bigr) \times \cnum^{\ast} =S_{\nbiga}\times\cnum^{\ast}. \] We may assume that the splitting $\kappa_{\nbigatilde}:L_{\nbigatilde}^{\lor}\lrarr M_{\nbigatilde}^{\lor}$ is given by the direct sum of $\kappa_{\nbiga}:L_{\nbiga}^{\lor}\lrarr M_{\nbiga}^{\lor}$ and $\seisuu\Upsilon\lrarr \seisuu e_{m+r+1}^{\lor}$. Then, we have $F_{\gamma_{\nbigatilde}}(x,\tau,t) =F_{\gamma_{\nbiga}}(x,t)+ \tau t^{\veca_{m+r+1}}$. Then, by the construction of $\gammatilde_{\nbiga}$, we have $\gammatilde_{\nbiga|S_{\nbigatilde}} =\gamma_{\nbigatilde}$. We obtain $\nbigt(\nbigatilde,\Stilde_{\nbiga},\gammatilde_{\nbiga}) _{|\proj^m\times S_{\nbigatilde}} =\nbigt(\nbigatilde,S_{\nbigatilde},\gamma_{\nbigatilde})$, and hence $\gbigvtilde_{\nbiga|S_{\nbigatilde}} =\gbigv_{\nbigatilde}$. The claim on the $\cnum^{\ast}$-action is also easy to see. \hfill\qed \vspace{.1in} Recall the following due to Reichelt-Sevenheck who proved it in a more general situation. \begin{prop}[] $\gbigvtilde_{\nbiga}$ is regular along $\tau$. \hfill\qed \end{prop} \begin{lem} We consider the morphism $\lambda N:\psi_{\tau,-\vecdelta}\gbigvtilde_{\nbiga} \lrarr \psi_{\tau,-\vecdelta}\gbigvtilde_{\nbiga}$, where $N$ is induced by $\tau\del_{\tau}$. \begin{itemize} \item The kernel and the cokernel of $\lambda N$ are isomorphic to $\gbigv_{\nbiga,!}$ and $\gbigv_{\nbiga,\ast}$ respectively. \item Let $W(N)$ denote the weight filtration of $\lambda N$ on $\psi_{\tau,-\vecdelta}\gbigvtilde_{\nbiga}$. The induced filtrations on $\Ker(\lambda N)$ and $\Cok(\lambda N)$ are also denoted by $W(N)$. Then, \[ W(N)_{k}\bigl( \Cok(\lambda N)\bigr) \simeq \Wtilde_{k-n-r}\bigl( \gbigv_{\nbiga,\ast} \bigr) \] \[ W(N)_{k}\bigl(\Ker(\lambda N)\bigr) \simeq \Wtilde_{k-n-r}\bigl( \gbigv_{\nbiga,!} \bigr) \] \end{itemize} \end{lem} \pf The first follows from the isomorphisms in Corollary and the constructions of $\gbigv_{\nbiga}$ and $\gbigvtilde_{\nbiga}$. The second also follows from the isomorphisms in Corollary , the comparison of the filtrations $W(\nbign)$ and $W$ on $\psi_{\tau,-\vecdelta}(\nbigt)$, and the constructions of the filtrations $\Wtilde$. \hfill\qed \vspace{.1in} Let $\Btilde$ be any open set in $\Stilde_{\nbiga}$ such that $\gbigvtilde_{\nbiga}$ is locally free $\nbigo_{\cnum_{\lambda}\times \Btilde}(\ast\tau)$-module, and regular singular along $\tau=0$. The polarization and the real structure of $\pi_{\dagger}^0 \nbigt(\nbigatilde,\Stilde_{\nbiga},\gammatilde_{\nbiga})$ induce a pairing $\gbigptilde_{\nbiga}$ of weight $-n-r$ on $\gbigvtilde_{\nbiga|\Btilde}$. We clearly have $\gbigptilde_{\nbiga|\Btilde\setminus\{\tau=0\}} =\gbigp_{\nbigatilde|\Btilde\setminus\{\tau=0\}}$. We obtain the induced graded pairing $\sp_{\tau}(\gbigptilde_{\nbiga})$ on $(\Cok(\lambda N),\Wtilde)$ by the procedure in \S. We also have $\sp_{\tau}(\gbigptilde_{\nbiga})$ on $(\Ker(\lambda N),\Wtilde)$. \begin{lem} We have $\sp_{\tau}(\gbigptilde_{\nbiga}) =\gbigp_{\nbiga,\ast}$ under the isomorphism $(\Cok(\lambda N),\Wtilde) \simeq (\gbigv_{\nbiga\ast},\Wtilde)$. We also have $\sp_{\tau}(\gbigptilde_{\nbiga}) =\gbigp_{\nbiga,!}$ under the isomorphism $(\Ker(\lambda N),\Wtilde) \simeq (\gbigv_{\nbiga!},\Wtilde)$. \end{lem} \pf It follows from Proposition . \hfill\qed \subsubsection{Cylindrical ends} We set $Y:=\proj\bigl(\nbigo_X\oplus\bigoplus_{j=1}^{r}\nbigl_j\bigr)$ with the projection $\pi:Y\lrarr X$. It is the projective completion of the vector bundle $\bigoplus_{j=1}^{r}\nbigl_j^{\lor}$. Recall that $Y$ is also weak Fano. Indeed, for the canonical bundle $K_Y$ of $Y$, we have $K_{Y}^{\lor} =\pi^{\ast}\bigl( K_X\otimes\bigotimes_{i=1}^r\nbigl_i\bigr)^{\lor} \otimes \nbigo_Y(1)^{\otimes (r+1)}$. By the assumption, $\pi^{\ast}(K_X\otimes \bigotimes_{i=1}^r\nbigl_i )^{\lor}$ is nef. Because $\nbigl_i$ are nef line bundles on a toric variety, they are globally generated. Then, it is easy to see that $\nbigo_Y(1)$ is also globally generated. In particular, it is nef. As a result, $K_Y^{\lor}$ is nef. We have the natural morphisms: \[ \begin{CD} L_{X}^{\lor} @>>> L_{\nbigatilde}^{\lor} @>>> L_{\nbiga}^{\lor}\\ @V{\simeq}VV @V{\simeq}VV @V{\simeq}VV \\ H^2(X,\seisuu) @>{\pi^{\ast}}>> H^2(Y,\seisuu) @>{i_0^{\ast}}>> H^2(X,\seisuu) \end{CD} \] We take a frame $\eta_1,\ldots,\eta_{\ell}$ of $H^2(X,\seisuu)$ such that each $\eta_i$ is the first Chern class of a nef line bundle. Let $\eta_{\ell+1}\in H^2(Y,\seisuu)$ be the first Chern class of the tautological line bundle $\nbigo(1)$ of $Y$ over $X$. By the assumption on the line bundles $\nbigl_j$, $\eta_{\ell+1}$ is also a nef class. The tuple $\eta_1,\ldots,\eta_{\ell+1}$ gives a frame of $H^2(Y,\seisuu)$. Let $\xi_i\in L_{\nbiga}^{\lor}$ be the elements corresponding to $\eta_i$. We have $\xi_{\ell+1}=\Upsilon$. The tuple $\xi_1,\ldots,\xi_{\ell}$ gives a frame of $L_{\nbiga}^{\lor}$. It gives a coordinate system $(x_1,\ldots,x_{\ell})$ on $S_{\nbiga}$ with which $S_{\nbiga}\simeq (\cnum^{\ast})^{\ell}$. Similarly, the tuple $\xi_1,\ldots,\xi_{\ell+1}$ gives a frame of $L_{\nbigatilde}^{\lor}$. It gives a coordinate system $(x_1,\ldots,x_{\ell+1})$ on $S_{\nbigatilde}$ with which $S_{\nbigatilde}\simeq (\cnum^{\ast})^{\ell+1}$. We have $\tau=x_{\ell+1}$ under the identification in Lemma . Recall the following due to Iritani. \begin{lem}[] There exists $\epsilon>0$ such that the following holds: \begin{itemize} \item $B_{\nbigatilde}=\bigl\{ (x_1,\ldots,x_{\ell+1})\,\big|\, 0<|x_i|<\epsilon\,\,(i=1,\ldots,\ell+1) \bigr\}$ is contained in $S_{\nbigatilde}^{\reg}$. \item $B_{\nbiga}=\bigl\{ (x_1,\ldots,x_{\ell})\,\big|\, 0<|x_i|<\epsilon\,\,(i=1,\ldots,\ell) \bigr\}$ is contained in $S_{\nbiga}^{\reg}$. \end{itemize} \end{lem} \pf The first claim is proved in \cite[Appendix A1]{Iritani}. Let $\sigma$ be a face of $\Conv(\nbiga\cup\{0\})$ such that $0$ is not contained in $\sigma$. Then, $\sigma$ is a face of $\Conv(\nbigatilde)$. We have the equality $F_{\gamma_{\nbiga},\sigma} =F_{\gamma_{\nbigatilde},\sigma}$, and $F_{\gamma_{\nbiga},\sigma}$ does not contain $x_{\ell+1}$. Hence, the second claim follows. \hfill\qed \vspace{.1in} We consider the following subset $\Btilde_{\nbiga}$ in $\Stilde_{\nbiga}$: \[ \Btilde_{\nbiga}= \bigl\{ (x_1,\ldots,x_{\ell+1})\,\big|\, 0<|x_i|<\epsilon\,\,(i=1,\ldots,\ell),\,\, |x_{\ell+1}|<\epsilon \bigr\} \] The restriction of $\gbigvtilde_{\nbiga|B_{\nbigatilde}}$ is a locally free $\nbigo_{\cnum_{\lambda}\times B_{\nbigatilde}} (\ast x_{\ell+1})$-module, and it is regular along $x_{\ell+1}$. \subsubsection{Coverings} We set $\gbigs_{\nbiga}:= \cnum\otimes L_{\nbiga}^{\lor}$. The exponential map $\cnum\lrarr \cnum^{\ast}$ induces the covering map $\chi_{\nbiga}: \gbigs_{\nbiga} \lrarr S_{\nbiga}$. We obtain an $\nbigrtilde$-module $\chi^{\ast}_{\nbiga}\gbigv_{\nbiga\star}$ equipped with the filtration $\chi^{\ast}_{\nbiga}\Wtilde$ on $\gbigs_{\nbiga}$. The restriction to $\gbigs^{\reg}_{\nbiga}:= \chi^{-1}_{\nbiga}(S^{\reg}_{\nbiga})$ are equipped with graded pairings $\chi_{\nbiga}^{\ast}\gbigp_{\nbiga\star}$. We set $\widetilde{\gbigs}_{\nbiga}:= \cnum\otimes L^{\lor}_{\nbigatilde}$. By the decomposition $L_{\nbigatilde}^{\lor}=L_{\nbiga}^{\lor} \oplus\seisuu\Upsilon$ in the proof of Lemma , we naturally have $\widetilde{\gbigs}_{\nbiga} =\gbigs_{\nbiga}\times\cnum$. The map $\chi_{\nbiga}$ and the identity on $\cnum$ induces $\chitilde_{\nbiga}:\widetilde{\gbigs}_{\nbiga}\lrarr \Stilde_{\nbiga}$. Take a frame $\xi_1,\ldots,\xi_{\ell}$ of $L_{\nbiga}^{\lor}$ and an open subset $\Btilde_{\nbiga}$ as in \S. The restriction of $\chitilde_{\nbiga}^{\ast}\gbigvtilde_{\nbiga}$ to $\chitilde_{\nbiga}^{-1}(\Btilde_{\nbiga})$ is equipped with the pairing $\chitilde_{\nbiga}^{\ast}\gbigptilde_{\nbiga}$ of weight $-n-r$. \begin{lem} The restriction of the mixed TEP-structure $\chi^{\ast}_{\nbiga}\bigl( \gbigv_{\nbiga\star|S_{\nbiga}^{\reg}}, \Wtilde,\gbigp_{\nbiga\ast} \bigr)$ to $\chitilde_{\nbiga}^{-1}(B_{\nbiga})$ $(\star=\ast,!)$ are obtained from $\chitilde^{\ast}_{\nbiga}(\gbigvtilde_{\nbiga},\gbigptilde_{\nbiga}) _{|\Btilde_{\nbiga}}$ by the procedure in {\rm\S}. \hfill\qed \end{lem} \subsubsection{Logarithmic extension and endomorphisms of $\gbigv_{\nbigatilde}$} This is the continuation of \S. Let us recall the explicit description of the logarithmic extension of $\gbigv_{\nbigatilde}$ due to Reichelt and Sevenheck in . See for more details. We also give an easy remark on uniqueness of automorphisms, although it is also essentially implied in . Let $B_{\nbigatilde}$ be as in Lemma . We set \[ \Bbar_{\nbigatilde}:= \bigl\{ (x_1,\ldots,x_{\ell+1})\,\big|\, |x_i|<\epsilon\,\,(i=1,\ldots,\ell+1) \bigr\}. \] First, the following holds. \begin{prop}[\S3.2 of ] There exists a locally free $\nbigo_{\cnum_{\lambda}\times \Bbar_{\nbigatilde}}$-module $\gbigq_{\nbigatilde}$ with an isomorphism \[ \gbigq _{\nbigatilde|\cnum_{\lambda}\times B_{\nbigatilde}} \simeq \gbigv_{\nbigatilde|B_{\nbigatilde}} \] such that the following holds: \begin{itemize} \item The meromorphic flat connection $\nabla$ of $\gbigv_{\nbigatilde|B_{\nbigatilde}}$ gives a meromorphic flat connection $\nabla^{\gbigq_{\nbigatilde}}$ of $\gbigq_{\nbigatilde}$ which is logarithmic along $\{x_i=0\}$ $(i=1,\ldots,\ell+1)$. \item The residues $\Res_{x_i}(\nabla^{\gbigq_{\nbigatilde}}) _{|\cnum_{\lambda}\times \{0\}}$ are nilpotent. \end{itemize} Such $\gbigq_{\nbigatilde}$ is unique up to canonical isomorphisms. \hfill\qed \end{prop} They considered the residual connection $\nabla^E$ on $E:=\gbigq_{\nbigatilde|\cnum_{\lambda}\times\{0\}}$. They give a quite explicit description of the meromorphic flat bundle $\bigl( E,\nabla^E \bigr)$. Let $N_{i}:E\lrarr \lambda^{-1}E$ $(i=1,\ldots,\ell+1)$ denote $\Res_{x_i}(\nabla^{\gbigq_{\nbigatilde}}) _{|\cnum_{\lambda}\times\{0\}}$. We have the fundamental vector fields $\underline{\gminiv}=\sum_{i=1}^{\ell+1} k_ix_i\del_{x_i}$ of the $\cnum^{\ast}$-action on $S_{\nbigatilde}$ (see \S). Note that for the $\cnum^{\ast}$-homogeneity of $\gbigv_{\nbigatilde}$, the fundamental vector field of the $\cnum^{\ast}$-action on $\cnum_{\lambda}\times S_{\nbigatilde}$ is $\lambda\del_{\lambda}+\underline{\gminiv}$. Let us consider the $\cnum$-homomorphism $L:E\lrarr E$ given by $\nabla^E(\lambda\del_{\lambda}) +\sum_{i=1}^{\ell+1} k_iN_i$. Because $[L,N_i]=0$, we have $[L,\lambda N_i]=\lambda N_i$. For the following proposition, and more detailed description in terms of the cohomology group of $Y$, see Lemma 3.8 and its proof in . \begin{prop}[\S3.2 of ] There exists a frame $u_1,\ldots,u_{\rank E}$ for which we have the following: \begin{itemize} \item $L(u_j)=c_ju_j$ for $c_j\in\seisuu$. \item We have $c_1<c_i$ $(i\neq 1)$. \item We have $N_iu_j=\sum \alpha^i_{kj}u_k$ for some $\alpha^i_{kj}\in\seisuu$. In particular, $\Sym\Bigl( \bigoplus_{i=1}^{\ell+1} \cnum N_i \Bigr)$ acts on $F:=\bigoplus_{i=1}^{\rank E} \cnum u_i$. \item The map $\Sym\Bigl( \bigoplus_{i=1}^{\ell+1} \cnum N_i \Bigr) \lrarr F$ given by $P(N_1,\ldots,N_{\ell+1}) \longmapsto P(N_1,\ldots,N_{\ell+1})u_{1}$ is surjective. \hfill\qed \end{itemize} \end{prop} By the first property in Proposition , we have the $\cnum^{\ast}$-action on $E$ which induces $L$. As a result, we have a vector space $H$ with an increasing filtration $F$ indexed by integers such that the $\cnum^{\ast}$-equivariant bundle $E$ is isomorphic to the analytification of the Rees module of $(H,F)$. Let $d_0:=\min\{d\,|\,\Gr^F_d(H)\neq 0\}$. The second property implies that $\dim F_{d_0}(H)=1$. The third and the fourth properties imply that $F_{d_0}(H)$ generates $H$ over the induced actions of the residues. The following is also essentially implied in the proof of \cite[Lemma 3.8]{Reichelt-Sevenheck1}. \begin{cor} For any endomorphism $\varphi$ of the $\nbigrtilde$-module $\gbigv_{\nbigatilde|B_{\nbigatilde}}$, there exists a complex number $\alpha$ such that $\varphi=\alpha\id$. \end{cor} \pf Because $\gbigq_{|\cnum_{\lambda}^{\ast}\times \Bbar_{\nbigatilde}}$ is the Deligne extension of $\gbigv_{\nbigatilde|\cnum_{\lambda}^{\ast}\times B_{\nbigatilde}}$ to $\cnum_{\lambda}^{\ast}\times \Bbar_{\nbigatilde}$, the morphism $\varphi$ is extended to an endomorphism of the $\nbigo$-module $\gbigq_{\nbigatilde}$ compatible with $\nabla^{\gbigq_{\nbigatilde}}$. Hence, it induces an endomorphism $\varphi^E$ of $(E,\nabla^E)$ compatible with the actions of the residues. Because the $\cnum^{\ast}$-action is determined by $\nabla^E$ and the residues, $\varphi^E$ is $\cnum^{\ast}$-equivariant. Hence, it induces an endomorphism $\varphi^H$ of $(H,F)$ compatible with the induced actions of residues. Because $F_{d_0}$ generates $H$ over the actions of the residues, $\varphi^H$ is uniquely determined by the restriction $\varphi^H_{|F_{d_0}}$. Because $\dim F_{d_0}=1$, we have $\alpha\in\cnum$ such that $\varphi^H_{|F_{d_0}}$ is the multiplication of $\alpha$. If $\alpha=0$, then $\varphi=0$. Hence, we have $\varphi=\alpha\id$. \hfill\qed \begin{cor} Let $U\subset S_{\nbigatilde}$ be any connected open subset such that $B_{\nbigatilde}\subset U$. For any endomorphism $\varphi$ of the $\nbigrtilde$-module $\gbigv_{\nbigatilde|\cnum_{\lambda}\times U}$, there exists a complex number $\alpha$ such that $\varphi=\alpha\id$. \hfill\qed \end{cor} \begin{cor} Let $P':\gbigv_{\nbigatilde|S^{\reg}_{\nbigatilde}} \otimes j^{\ast}\gbigv_{\nbigatilde|S^{\reg}_{\nbigatilde}} \lrarr \lambda^{n+r}\nbigo_{\cnum_{\lambda}\times S^{\reg}_{\nbiga}}$ be a morphism of $\nbigrtilde$-modules. Then, there exists a complex number $\alpha$ such that $P'=\alpha \gbigp_{\nbigatilde}$. \hfill\qed \end{cor} \subsubsection{Logarithmic extension and endomorphisms of $\gbigv_{\nbiga}$} We can easily deduce the same property for $\gbigv_{\nbiga\star}$ ($\star=\ast,!$). Let $B_{\nbiga}$ be as in Lemma . We set \[ \Bbar_{\nbiga}:= \bigl\{ (x_1,\ldots,x_{\ell})\,\big|\, |x_i|<\epsilon\,\,(i=1,\ldots,\ell) \bigr\}. \] We regard $\Bbar_{\nbiga}$ as $\bigl\{ (x_1,\ldots,x_{\ell+1})\in\Bbar_{\nbigatilde}\,\big|\, x_{\ell+1}=0 \bigr\}$. We obtain the vector bundle $\gbigq_{\nbigatilde|\Bbar_{\nbiga}}$. It is equipped with the endomorphism $\lambda\Res_{x_{\ell+1}}(\nabla^{\gbigq_{\nbigatilde}})$, and the induced connection. \begin{lem} The conjugacy classes of $\lambda\Res_{x_{\ell+1}}(\nabla^{\gbigq_{\nbigatilde}})_{|(\lambda,P)}$ are independent of the choice of $(\lambda,P)\in \cnum_{\lambda}\times \Bbar_{\nbiga}$. \end{lem} \pf By the construction, $\gbigv_{\nbigatilde}$ comes from a harmonic bundle. It is easy to see that $\gbigq_{\nbigatilde}$ is equal to the prolongment of the family of $\lambda$-flat bundles $\nbigp_0\bigl( \gbigv_{\nbigatilde|B_{\nbigatilde}} \bigr)$ associated to tame harmonic bundles, as remarked in . Then, the claim follows from a general theory of harmonic bundles . \hfill\qed \vspace{.1in} Let $\gbigq_{\nbiga}$ be the cokernel of $\lambda\Res_{x_{\ell+1}}(\nabla^{\gbigq})$. It is a locally free $\nbigo_{\cnum_{\lambda}\times\Bbar_{\nbiga}}$-module. The following is clear by the relation of $\gbigv_{\nbiga\ast}$ and $\gbigv_{\nbigatilde}$. \begin{lem} We have natural isomorphisms of $\nbigrtilde$-modules $\gbigq_{\nbiga|\cnum_{\lambda}\times B_{\nbiga}} \simeq \gbigv_{\nbiga\ast|\cnum_{\lambda}\times B_{\nbiga}}$. \hfill\qed \end{lem} Set $E_{\nbiga}:= \gbigq_{\nbiga|\cnum_{\lambda}\times\{0\}}$. It is equipped with the induced connection $\nabla^{E_{\nbiga}}$. We obtain an explicit description of $(E_{\nbiga},\nabla^{E_{\nbiga}})$ from the description of $(E,\nabla^E)$ in . Let $N_{i}:E_{\nbiga}\lrarr \lambda^{-1}E_{\nbiga}$ $(i=1,\ldots,\ell)$ denote $\Res_{x_i}(\nabla^{\gbigq_{\nbiga}}) _{|\cnum_{\lambda}\times\{0\}}$. We have the fundamental vector fields $\underline{\gminiv}=\sum_{i=1}^{\ell} k_ix_i\del_{x_i}$ of the $\cnum^{\ast}$-action on $S_{\nbiga}$ (see \S). Note that for the $\cnum^{\ast}$-homogeneity of $\gbigv_{\nbiga\ast}$, the fundamental vector field of the $\cnum^{\ast}$-action on $\cnum_{\lambda}\times S_{\nbiga}$ is $\lambda\del_{\lambda}+\underline{\gminiv}$. Let us consider the $\cnum$-homomorphism $L:E_{\nbiga}\lrarr E_{\nbiga}$ given by $\nabla^E(\lambda\del_{\lambda}) +\sum_{i=1}^{\ell} k_iN_i$. Because $[L,N_i]=0$, we have $[L,\lambda N_i]=\lambda N_i$. The following is a direct consequence of Proposition . \begin{prop} There exists a frame $u_1,\ldots,u_{\rank E_{\nbiga}}$ for which we have the following: \begin{itemize} \item $L(u_j)=c_ju_j$ for $c_j\in\seisuu$. \item We have $c_1<c_i$ $(i\neq 1)$. \item We have $N_iu_j=\sum \beta^{i}_{kj}u_k$ for $\beta^{i}_{kj}\in\seisuu$. In particular, $\Sym\Bigl( \bigoplus_{i=1}^{\ell} \cnum N_i \Bigr)$ acts on $F_{\nbiga}:=\bigoplus_{i=1}^{\rank E_{\nbiga}} \cnum u_i$. \item The map $\Sym\Bigl( \bigoplus_{i=1}^{\ell} \cnum N_i \Bigr) \lrarr F_{\nbiga}$ given by $P(N_1,\ldots,N_{\ell}) \longmapsto P(N_1,\ldots,N_{\ell})u_{1}$ is surjective. \hfill\qed \end{itemize} \end{prop} We omit a more detailed description in terms of the cohomology group of $X$. By the first property in Proposition , we have the $\cnum^{\ast}$-action on $E_{\nbiga}$ which induces $L$. As a result, we have a vector space $H_{\nbiga}$ with an increasing filtration $F$ indexed by integers such that the $\cnum^{\ast}$-equivariant bundle $E_{\nbiga}$ is isomorphic to the analytification of the Rees module of $(H_{\nbiga},F)$. Let $d_0:=\min\{d\,|\,\Gr^F_d(H)\neq 0\}$. The second property implies that $\dim F_{d_0}(H_{\nbiga})=1$. The third and the fourth properties imply that $F_{d_0}(H_{\nbiga})$ generates $H_{\nbiga}$ over the induced actions of the residues. \begin{cor} Let $U\subset S_{\nbiga}$ be any connected open subset such that $B_{\nbiga}\subset U$. For any endomorphism $\varphi$ of the $\nbigrtilde$-module $\gbigv_{\nbiga\star|\cnum_{\lambda}\times U}$ $(\star=\ast,!)$, there exists a complex number $\alpha$ such that $\varphi=\alpha\id$. \end{cor} \pf We obtain the claim for $\gbigv_{\nbiga,\ast}$ by the argument in the proof of Corollary and Corollary . Because $\gbigv_{\nbiga,!}\simeq \lambda^{2n-m+r}j^{\ast}\DDD\gbigv_{\nbiga\ast}$, we obtain the claim for $\gbigv_{\nbiga,!}$. \hfill\qed \vspace{.1in} We have an isomorphism $\Psi: j^{\ast}\bigl( \gbigv_{\nbiga,\ast|S_{\nbiga}^{\reg}} \bigr)^{\lor} \simeq \lambda^{-n-r} \gbigv_{\nbiga,!|S_{\nbiga}^{\reg}}$ induced by the duality isomorphism $\gbigv_{\nbiga,!} \simeq \lambda^{2n-m+r}j^{\ast}\DDD\gbigv_{\nbiga,\ast}$. \begin{cor} Let $U\subset S_{\nbiga}^{\reg}$ be any open subset such that $B_{\nbiga}\subset U$. For any morphism of $\nbigrtilde$-modules $\Psi': j^{\ast}\bigl( \gbigv_{\nbiga,\ast|\cnum_{\lambda}\times U} \bigr)^{\lor} \lrarr \lambda^{-n-r} \gbigv_{\nbiga,!|\cnum_{\lambda}\times U}$, there exists a complex number $\alpha$ such that $\Psi'=\alpha\Psi$. \hfill\qed \end{cor} \subsection{Preliminary on the A-side} We recall some basic matters on the quantum products of toric varieties. For example, see , , , , , for more details. \subsubsection{The cohomology ring of projective bundles} Let $X$ be a complex manifold. For simplicity, we assume $H^{2i+1}(X,\seisuu)=0$ for any $i$. We shall omit the coefficient of the cohomology group if the coefficient is $\cnum$. Let $\nbige$ be a locally free $\nbigo_X$-module of rank $r$. Let $\nbige^{\lor}$ denote the dual. Let $Y:=\proj(\nbige\oplus\nbigo_X)= \Proj\bigl( \Sym^{\bullet}(\nbige\oplus\nbigo_X)\bigr)$ denote the projective completion of $\nbige^{\lor}$. We have the natural inclusions $i_0:X\simeq \proj(\nbigo)\lrarr Y$ and $i_{\infty}: H_{\infty}:=\proj(\nbige)\lrarr Y$. We have $i_0(X)\subset \nbige^{\lor}$ and $Y=\nbige^{\lor}\cup H_{\infty}$. Let $\pi:Y\lrarr X$ denote the projection. Let $\pi^{\ast}$ and $i_0^{\ast}$ denote the pull back of the cohomology. We have $H^{\ast}(Y) =\Ker i_0^{\ast}\oplus \Image \pi^{\ast}$. Let $i_{0\ast}:H^i(X)\lrarr H^{i+2r}(Y)$ and $\pi_{\ast}:H^i(Y)\lrarr H^{i-2r}(X)$ denote the Gysin map. We have $H^{\ast}(Y) =\Image i_{0\ast} \oplus \Ker \pi_{\ast}$. Let $\nbigo_Y(1)$ denote the tautological line bundle of $Y$ over $X$. We set $\gamma:=c_1\bigl(\nbigo_Y(1)\bigr) \in H^2(Y)$. We have $H^{\ast}(Y) =\bigoplus_{j=0}^r \gamma^j\cdot \pi^{\ast}\bigl( H^{\ast}(X)\bigr)$. Note that $\gamma$ is the cohomology class representing $H_{\infty}$. In particular, we have $i_0^{\ast}\gamma=0$. Hence, $\Ker i_0^{\ast}= \bigoplus_{j=1}^{r} \pi^{\ast}H^{\ast}(X)\gamma^j$. Because $c_{r+1}(\nbige^{\lor}\oplus \nbigo)=0$, we have \begin{equation} \gamma\cdot \sum_{j=0}^{r} \gamma^{r-j} \pi^{\ast}c_j(\nbige^{\lor}) =0. \end{equation} Let $N:H^{\ast}(Y)\lrarr H^{\ast}(Y)$ be determined by $N(\sigma)=\gamma\sigma$. By (), we have $\Image N= \bigoplus_{i=1}^r \gamma^i\cdot \pi^{\ast}\bigl(H^{\ast}(X)\bigr)$, and $\Cok N\simeq \Image\pi^{\ast}$. \begin{lem} We have $\Image i_{0\ast}=\Ker N$. \end{lem} \pf Because $H_{\infty}\cap i_0(X)=0$, we have $\Image (i_{0\ast})\subset\Ker N$. By comparison of the dimension, we obtain $\Image i_{0\ast}=\Ker N$. \hfill\qed \vspace{.1in} In particular, $\Ker(N)\cap H^{2r}(Y) =i_{0\ast}H^0(X,\cnum)$ is one dimensional, and that $\sum_{j=0}^r\gamma^{r-j} \pi^{\ast}c_j(\nbige^{\lor})$ is a base. By considering the restriction to a ball in $X$, we obtain that $i_{0\ast}(1) =\sum_{j=0}^r\gamma^{r-j} \pi^{\ast}c_j(\nbige^{\lor})$. In particular, we have $i_{0\ast}\sigma = \pi^{\ast}\sigma \sum_{j=0}^r\gamma^{r-j} \pi^{\ast}c_j(\nbige^{\lor})$. \subsubsection{Quantum products of $Y$ in the weak Fano toric case} Suppose that (i) $X$ is a smooth weak Fano toric variety, (ii) $\nbige$ is the direct sum of holomorphic line bundles $\nbigl_i$ $(i=1,\ldots,r)$, (iii) $\nbigl_i$ $(i=1,\ldots,r)$ and $(\bigotimes_{i=1}^r\nbigl_i \otimes K_X)^{\lor}$ are nef. Then, $Y$ is also a weak Fano toric manifold. Let $\Eff^{\ast}(Y)\subset H_2(Y,\seisuu)$ denote the subset of the homology classes of non-empty algebraic curves. We set $\Eff(Y):=\Eff^{\ast}(Y)\cup\{0\}$. We take a homogeneous base $\phi_1,\ldots,\phi_{\chi(Y)}$ of $H^{\ast}(Y,\cnum)$. Let $\phi^1,\ldots,\phi^{\chi(Y)}$ be the dual base of $H^{\ast}(Y)$ with respect to the Poincar\'e pairing. For $d\in\Eff(Y)$, let $Y_{0,3,d}$ denote the moduli stack of stable maps $f:C\lrarr Y$ where $C$ denotes a $3$-pointed pre-stable curve with genus $0$, and $f$ denotes a morphism such that the homology class of $f(C)$ is $d$. Let $p:\nbigc_{0,3,d}\lrarr Y_{0,3,d}$ be the universal curve, and let $\ev_i:\nbigc_{0,3,d}\lrarr X$ be the evaluation map at the $i$-th marked point. Let $[Y_{0,3,d}]^{\vir}$ denote the virtual fundamental class of $Y_{0,3,d}$. For $\alpha_i\in H^{\ast}(Y)$ $(i=1,2,3)$, we obtain the number $\langle \alpha_1,\alpha_2,\alpha_3 \rangle^Y_{0,3,d} := \int_{[Y_{0,3,d}]^{\vir}} \prod_{i=1,2,3}\ev_i^{\ast}\alpha_i$ called the Gromov-Witten invariants of $Y$. According to a result of Iritani (see also and ), there exists an open subset $\nbigu\subset H^{2}(Y,\cnum)$ such that the following holds: \begin{itemize} \item $\nbigu$ is of the form $\bigl\{ u\in H^2(Y,\cnum)\,\big|\, \Re\langle u,d\rangle<-M,\,\,\, \forall d\in \Eff^{\ast}(Y) \bigr\}$ for some $M>0$. \item For any $u\in\nbigu$, the quantum product $\bullet_u$ on $H^{\ast}(Y,\cnum)$ is convergent: \[ \alpha\bullet_u\beta :=\sum_{d\in\Eff(Y)} \sum_{i=1}^{\chi(Y)} \langle \alpha,\beta,\phi_i \rangle^Y_{0,3,d} \cdot \phi^i \cdot e^{\langle u,d\rangle} \] \end{itemize} We have the decomposition $H^2(Y,\seisuu) =\pi^{\ast}\bigl( H^2(X,\seisuu) \bigr)\oplus \seisuu\gamma$, which induces $H^2(Y) =\pi^{\ast}\bigl( H^2(X)\bigr) \oplus \cnum\gamma$. \begin{lem} For a large $M'>0$, we have \[ \pi^{\ast} \bigl\{ \sigma\in H^2(X,\cnum)\,\big|\, \Re(\langle\sigma,d\rangle) <-M',\,\,\forall d\in\Eff^{\ast}(X) \bigr\} \times \{c\cdot \gamma\,|\, c \in\cnum,\,\,\Re(c)<-M'\} \subset\nbigu. \] \end{lem} \pf Any algebraic curve $C$ is homologous to a curve $C_0\cup \bigcup_{j=1}^r C_j \cup\bigcup_{i=1}^{\ell} F_i$ where $C_0$ is contained in $i_0(X)$, $C_{j}$ are contained in $\proj(\nbigl_j)\subset Y$, and $F_i$ $(i=1,\ldots,\ell)$ are contained in fibers of $\pi:Y\lrarr X$. We have \[ \sum_{j=0}^r \Re \langle \sigma+c\gamma,[C_j] \rangle +\sum_{i=1}^{\ell} \Re\langle\sigma+c\gamma,[F_i] \rangle \leq \sum_{j=0}^r \Re \langle \sigma,[C_j] \rangle + \sum_{i=1}^{\ell} \Re(c)\langle \gamma,[F_i]\rangle \] Then, the claim of the lemma is clear. \hfill\qed \vspace{.1in} We set $\nbigu_X:= \bigl\{ \sigma\in H^2(X,\cnum)\,\big|\, \Re(\langle\sigma,d\rangle) <-M' \bigr\}$, $\nbigu_{\gamma}:= \bigl\{ c\cdot\gamma\,\big|\, c\in\cnum,\,\,\, \Re(c)<-M' \bigr\}$ and $\nbigu_Y:=\nbigu_X\times\nbigu_{\gamma}$. A point $\sigma+c\gamma\in \nbigu_Y$ is denoted by $(\sigma,c)$. We have a natural action of $2\pi\sqrt{-1}H^2(Y,\seisuu)$ on $\nbigu_{Y}$ by the addition. The quantum product $\bullet_{\sigma,c}$ depends only on the equivalence class of $(\sigma,c)$ in $\nbigu_Y\big/2\pi\sqrt{-1}H^2(Y,\seisuu)$. \subsubsection{Degenerated quantum products on $H^{\ast}(Y)$ and $H^{\ast}(X)$} Let $\Eff(Y,\gamma)$ denote the set of $[C]\in \Eff(Y)$ such that $\langle \gamma,[C]\rangle=0$. We set $\Eff^{\ast}(Y,\gamma):=\Eff(Y,\gamma)\cap\Eff^{\ast}(Y)$. Let $\sigma\in\nbigu_X$. As in , by taking the limit of $\bullet_{(\sigma,c)}$ for $\Re(c)\lrarr -\infty$, we obtain the following product $\bullet_{\sigma}$ on $H^{\ast}(Y)$: \[ \alpha\bullet_{\sigma}\beta :=\sum_{d\in\Eff(Y,\gamma)} \sum_{i=1}^{\chi(Y)} \langle \alpha,\beta,\phi_i \rangle^{Y}_{0,3,d} \cdot \phi^i \cdot e^{\langle \sigma,\pi_{\ast}d\rangle} \] As in \S, let $N:H^{\ast}(Y)\lrarr H^{\ast}(Y)$ be the endomorphism given by the cup product of $\gamma$, i.e., $N(\beta):=\gamma\cup\beta$ for $\beta\in H^{\ast}(Y)$. Note that $\gamma\bullet_{\sigma}\beta =\beta\bullet_{\sigma}\gamma= \beta\cup\gamma$ for any $\beta\in H^{\ast}(Y)$, which follows from the divisor axiom. Hence, $\Image(N)$ and $\Ker(N)$ are the ideal of the algebra $(H^{\ast}(Y),\bullet_{\sigma})$. By the natural isomorphism of vector spaces $H^{\ast}(X)\simeq \Cok(N)$, we obtain the induced product $\bullet^{\nbige^{\lor}}_{\sigma}$ on $H^{\ast}(X)$. The algebra $(H^{\ast}(X),\bullet^{\nbige^{\lor}}_{\sigma})$ is denoted by $Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$ in this paper. The multiplication $\Cok(N)\times \Ker(N)\lrarr \Ker(N)$ and the natural isomorphism $\Ker(N)\simeq H^{\ast}(X)$ induces a structure of $Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$-module on $H^{\ast}(X)$. The $Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$-module is denoted by $K_{\sigma}H^{\ast}(X,\nbige^{\lor})$. The multiplication of $b_1\in Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$ and $b_2\in K_{\sigma}H^{\ast}(X,\nbige^{\lor})$ is denoted by $b_1\bullet^{\nbige^{\lor}}_{K,\sigma}b_2$. \subsubsection{Filtrations and the graded pairings} Let $W(N)$ be the weight filtration of the nilpotent map $N$ on $H^{\ast}(Y)$. For any $\sigma\in\nbigu_X$, by shifting the filtrations of Konishi-Minabe in , we set \[ \Wtilde_{k}Q_{\sigma}H^{\ast}(X,\nbige^{\lor}):= \Image\Bigl( W(N)_{k+n+r}H^{\ast}(Y)\lrarr Q_{\sigma}H^{\ast}(X,\nbige^{\lor}) \Bigr). \] Here, $Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$ is identified with $\Cok(N)$. Note that \[ \Gr^{\Wtilde}_{k-n-r}Q_{\sigma}H^{\ast}(X,\nbige^{\lor}) \simeq \left\{ \begin{array}{ll} P\Gr^{W(N)}_kH^{\ast}(Y) & (k\geq 0) \\ 0 & (k<0) \end{array} \right. \] Here, $P\Gr^{W(N)}_kH^{\ast}(Y)$ is the primitive part, i.e., the kernel of $N^{k+1}:\Gr^{W(N)}_kH^{\ast}(Y) \lrarr \Gr^{W(N)}_{-k-2}H^{\ast}(Y)$. The Poincar\'e pairing $(\cdot,\cdot)_Y$ on $H^{\ast}(Y)$ and the nilpotent map $N$ induces the following symmetric pairing on $\Gr^{\Wtilde}_{k-n-r}Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$: \[ P^{\nbige^{\lor}}_{k-n-r} \bigl(a,b \bigr):= \bigl( a,N^{k}b \bigr)_Y \] It is easy and standard to see that the pairings are non-degenerate. We also have the filtration $W$ on $K_{\sigma}H^{\ast}(X,\nbige^{\lor})$ given as follows: \[ \Wtilde_kK_{\sigma}H^{\ast}(X,\nbige^{\lor}):= K_{\sigma}H^{\ast}(X,\nbige^{\lor}) \cap W(N)_{k+n+r}H^{\ast}(Y). \] Here $K_{\sigma}H^{\ast}(X,\nbige^{\lor})$ is identified with $\Ker(N)$. We have \[ \Gr^{\Wtilde}_{k-n-r}K_{\sigma}H^{\ast}(X,\nbige^{\lor}) \simeq \left\{ \begin{array}{ll} P'\Gr_{k}^{W(N)}H^{\ast}(Y) & (k\leq 0),\\ 0 & (k>0). \end{array} \right. \] Here, $P'\Gr_{k}^{W(N)}H^{\ast}(Y)$ is the image of $P\Gr_{-k}^{W(N)}H^{\ast}(Y)$ by $N^{-k}$. Hence, we have the induced pairings $\bigl(\cdot,\cdot\bigr)^{\nbige^{\lor}}_{k-n-r}$ on $\Gr^{\Wtilde}_{k-n-r}K_{\sigma}H^{\ast}(X,\nbige^{\lor})$. \vspace{.1in} Note that the filtrations $\Wtilde$ and the symmetric pairings $P^{\nbige^{\lor}}_k$ are independent of the choice of $\sigma$. Namely, we have the filtrations $\Wtilde$ on $\Cok(N)$ and $\Ker(N)$ such that $\Wtilde_k\Cok(N)= \Wtilde_kQ_{\sigma}H^{\ast}(X,\nbige^{\lor})$ and $\Wtilde_k\Ker(N)= \Wtilde_kK_{\sigma}H^{\ast}(X,\nbige^{\lor})$. We also have symmetric pairings $P^{\nbige^{\lor}}_k$ on $\Gr^{\Wtilde}_k\Cok(N)$ and $\Gr^{\Wtilde}_k\Ker(N)$ which are equal to the pairings on $\Gr^{\Wtilde}_kQ_{\sigma}H^{\ast}(X,\nbige^{\lor})$ and $\Gr^{\Wtilde}_kK_{\sigma}H^{\ast}(X,\nbige^{\lor})$ for any $\sigma$. \subsubsection{ Degenerated quantum products and local Gromov-Witten invariants (Appendix)} Let us recall the relation between the degenerated quantum products and local Gromov-Witten invariants in some special cases. We essentially follow . Let $d\in \Eff^{\ast}(X)$. Let $X_{0,3,d}$ be the moduli stack of stable maps $f:C\lrarr X$ where $C$ denotes a $3$-pointed pre-stable curve with genus $0$, and $f$ denotes a morphism such that the homology class of $f(C)$ is $d$. Let $p:\nbigc_{0,3,d}\lrarr X_{0,3,d}$ be the universal curve, and let $\mu:\nbigc_{0,3,d}\lrarr X$ be the universal map. Let us consider the following concavity condition for $d$: \begin{description} \item[(B)] $H^0(C,f^{\ast}\nbige^{\lor})=0$ for any $(C,f)\in X_{0,3,d}$. \end{description} Recall that $\nbige$ is the direct sum of the nef line bundles $\nbigl_i$ $(i=1,\ldots,r)$. We recall the following standard lemma. \begin{lem} Suppose that $\langle c_1(\nbigl_i^{\lor}),d\rangle<0$ for any $i$. Then, Condition {\rm(B)} holds for $d$. \end{lem} \pf Take $d\in \Eff^{\ast}(X)$ and $(C,f)\in X_{0,3,d}$. Let $C=\bigcup C_a$ be the irreducible decomposition. Because $\nbigl_i$ is nef, the degree of $f^{\ast}(\nbigl_i^{\lor})_{|C_a}$ is non-positive for any $a$. By the assumption $\langle c_1(\nbigl_i^{\lor},d\rangle)<0$, we have an irreducible component $C_{a_0}$ of $C$ such that the degree of $f^{\ast}(\nbigl_i^{\lor})_{|C_a}$ is strictly negative. We also remark that $C$ is connected. Then, we have $H^0(C,f^{\ast}\nbigl_i)=0$, i.e., Condition (B) holds for $d$. \hfill\qed \vspace{.1in} If Condition (B) holds for $d$, we obtain the locally free sheaf $R^1p_{\ast}\mu^{\ast}\nbige^{\lor}$ on $X_{0,3,d}$. Let $[X_{0,3,d}]^{\vir}$ denote the virtual fundamental class of $X_{0,3,d}$. Let $\ev_i:X_{0,3,d}\lrarr X$ be the evaluation map at the $i$-th marked point. For any vector bundle $E$, let $e(E)$ denote the Euler class. Then, for any $\alpha_i\in H^{\ast}(X)$ $(i=1,2,3)$, we set \[ \bigl\langle \alpha_1,\alpha_2,\alpha_3 \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} := \int_{[X_{0,3,d}]^{\vir}} \prod_{i=1,2,3}\ev_i^{\ast}(\alpha_i) \cdot e(Rp_{\ast}\mu^{\ast}\nbige^{\lor}). \] They are called $e^{-1}$-twisted Gromov-Witten invariants or local Gromov-Witten invariants. For any $\alpha,\beta\in H^{\ast}(X)$, let $\alpha\cup\beta$ denote the ordinary cup product. We take a homogeneous base $\rho_1,\ldots,\rho_{\ell}$ of $H^{\ast}(X)$. Let $\rho^{1},\ldots,\rho^{\ell}$ denote the dual base of $H^{\ast}(X)$ with respect to the Poincar\'e pairing of $X$. \begin{prop} Suppose one of the following holds: \begin{description} \item[Case 1.] $X$ is Fano, and $\nbigl_i$ $(i=1,\ldots,r)$ are ample. Note that Condition {\rm(B)} holds for any $d\in\Eff^{\ast}(X)$. \item[Case 2.] $X$ is a weak Fano surface, $r=1$, and $\nbigl_1=K_X^{-1}$. Note that Condition {\rm(B)} holds for any $d\in\Eff^{\ast}(X)$ with $\langle c_1(K_X^{\lor}),d\rangle\neq 0$. \end{description} Then, we have the following for $\sigma\in \nbigu_X$ and for $\alpha,\beta\in Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$: \begin{equation} \alpha\bullet^{\nbige^{\lor}}_{\sigma}\beta = \left\{ \begin{array}{ll} {\displaystyle \alpha\cup\beta +\sum_{d\in\Eff^{\ast}(X)} \sum_{i=1}^{\ell} \bigl\langle \alpha,\beta,e(\nbige^{\lor})\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle \sigma,d \rangle}\rho^i } & (\mbox{\rm Case 1}) \\ {\displaystyle \alpha\cup\beta +\sum_{ \substack{d\in\Eff^{\ast}(X),\\ \langle c_1(K_X),d\rangle\neq 0}} \sum_{i=1}^{\ell} \bigl\langle \alpha,\beta,e(\nbige^{\lor})\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle \sigma,d \rangle}\rho^i } & (\mbox{\rm Case 2}) \end{array} \right. \end{equation} We also have the following for $\sigma\in\nbigu_X$ and for $\alpha\in Q_{\sigma}H^{\ast}(X,\nbige^{\lor})$, and $\beta\in K_{\sigma}H^{\ast}(X,\nbige^{\lor})$: \begin{equation} \alpha\bullet^{\nbige^{\lor}}_{K,\sigma}\beta = \left\{ \begin{array}{ll} {\displaystyle \alpha\cup\beta +\sum_{d\in\Eff^{\ast}(X)} \sum_{i=1}^{\ell} \bigl\langle \alpha,e(\nbige^{\lor})\beta,\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle \sigma,d \rangle}\rho^i } & (\mbox{\rm Case 1})\\ {\displaystyle \alpha\cup\beta +\sum_{\substack{d\in\Eff^{\ast}(X)\\ \langle c_1(K_X),d\rangle\neq 0 \\}} \sum_{i=1}^{\ell} \bigl\langle \alpha,e(\nbige^{\lor})\beta,\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle \sigma,d \rangle}\rho^i } & (\mbox{\rm Case 2}) \end{array} \right. \end{equation} \end{prop} \pf For the proof of the proposition, we study the degenerated quantum products $\bullet_{\sigma}$ on $H^{\ast}(Y)$. Any element $\alpha\in H^{\ast}(Y)$ has the expression $\alpha=\sum_{j=0}^r\pi^{\ast}\alpha_j\cdot\gamma^j$ where $\alpha_j\in H^{\ast}(X)$. \begin{lem} Let $\sigma\in\nbigu_X$. Suppose Case {\rm 1} or Case {\rm 2}. For $\alpha=\sum_{j=0}^r \pi^{\ast}(\alpha_j)\gamma^j$ and $\beta=\sum_{j=0}^r \pi^{\ast}(\beta_j)\gamma^j$, we have the following: \begin{equation} \alpha\bullet_{\sigma}\beta = \left\{ \begin{array}{ll} {\displaystyle \alpha\cup\beta +\sum_{d\in\Eff^{\ast}(X)} \sum_{i=1}^{\ell} \langle \alpha_0, \beta_0,\rho_i \rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle \sigma,d\rangle} i_{0\ast}(\rho^i)} & \mbox{\rm(Case 1)} \\ {\displaystyle \alpha\cup\beta +\sum_{\substack{d\in\Eff^{\ast}(X)\\ \langle c_1(K_X),d\rangle\neq 0 }} \sum_{i=1}^{\ell} \langle \alpha_0, \beta_0,\rho_i \rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle \sigma,d\rangle} i_{0\ast}(\rho^i) } & \mbox{\rm(Case 2)} \end{array} \right. \end{equation} \end{lem} Before giving a proof of Lemma , we deduce Proposition from Lemma . We give an argument in Case 1. The other case can be given similarly. Because $i_0^{\ast}i_{0\ast}(\rho^j)=e(\nbige^{\lor})\rho^j$, we obtain the following for $\alpha,\beta\in H^{\ast}(X)$ from (): \begin{multline} \alpha\bullet^{\nbige^{\lor}}_{\sigma}\beta =\alpha\cup\beta +\sum_{d\in\Eff^{\ast}(X)}\sum_{i=1}^{\ell} \bigl\langle \alpha,\beta,\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle\sigma,d\rangle} e(\nbige^{\lor})\rho^i \\ =\alpha\cup\beta +\sum_{d\in\Eff^{\ast}(X)}\sum_{i=1}^{\ell} \bigl\langle \alpha,\beta, e(\nbige^{\lor})\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle\sigma,d\rangle} \rho^i \end{multline} Thus, we obtain (). We identify $\Ker (N)=i_{0\ast}H^{\ast}(X)$. We have $i_{0\ast}(\beta)= e(\nbige^{\lor})\beta +\sum_{j=1}^{r}c_j\gamma^j$ for some $c_j\in H^{\ast}(X)$. Hence we have the following for $\alpha\in H^{\ast}(X)$ and $\beta\in H^{\ast}(X)$: \begin{multline} i_{0\ast}\bigl( \alpha\bullet^{\nbige^{\lor}}_{K,\sigma} \beta\bigr) =\pi^{\ast}(\alpha)\bullet_{\sigma} i_{0\ast}(\beta) =\pi^{\ast}(\alpha)\cup i_{0\ast}\beta +\sum_{d\in \Eff^{\ast}(X)} \sum_{i=1}^{\ell} \bigl\langle \alpha,e(\nbige^{\lor})\beta,\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle\sigma,d\rangle} i_{0\ast}\rho^i \\ =i_{0\ast}\Bigl( \alpha\cup\beta +\sum_{d\in \Eff^{\ast}(X)} \sum_{i=1}^{\ell} \bigl\langle \alpha,e(\nbige^{\lor})\beta,\rho_i \bigr\rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} e^{\langle\sigma,d\rangle} \rho^i \Bigr) \end{multline} Thus, we obtain (). It remains to prove Lemma . \vspace{.1in} Let us prove Lemma in Case 1. For any algebraic curve $C$ in $Y$, let $[C]\in H_2(Y)$ denote the homology class of $C$. \begin{lem} Let $C$ be a non-empty algebraic curve in $Y$. Then, we have $\langle \gamma,[C]\rangle=0$ if and only if $C\subset i_0(X)$. \end{lem} \pf Because $\nbigo_Y(1)=\nbigo(H_{\infty})$, the second condition implies the first. Let us prove that the first condition implies the second. We may assume that $C$ is irreducible. If $C\subset H_{\infty}$, we clearly have $\langle \gamma,[C]\rangle>0$. Hence, we have $C\not\subset H_{\infty}$. If $C\cap H_{\infty}\neq\emptyset$, we have $\langle \gamma,[C]\rangle>0$. Hence, we obtain that $C\cap H_{\infty}=\emptyset$. Let $\varphi:\Ctilde\lrarr C\subset\nbige^{\lor}$ be the normalization. Let $q:\nbige^{\lor}\lrarr X$ be the projection. Note that $q(C)$ is not a point. We obtain the morphism $q\circ\varphi:\Ctilde\lrarr X$ and a section $s$ of $(q\circ\varphi)^{\ast}\nbige^{\lor}$. Because $(q\circ\varphi)^{\ast}\nbigl_i$ are ample, we have $s=0$. It means $C\subset i_0(X)$. Thus, we obtain Lemma . \hfill\qed \begin{lem} For $d\in\Eff^{\ast}(Y,\gamma)$ and for $i>0$, we have $\langle \alpha_1,\alpha_2,\gamma^i\alpha_3 \rangle^Y_{0,3,d}=0$. \end{lem} \pf The support of the cohomology class $\gamma^i\alpha_3$ is contained in $H_{\infty}$. For $(C,f)\in Y_{0,3,d}$, we have $f(C)\subset i_{0}(X)$. Then, the claim of Lemma is clear. \hfill\qed \vspace{.1in} By Lemma , we have $\Eff^{\ast}(Y,\gamma)=\Eff^{\ast}(X)$. By Lemma , for $d\in \Eff^{\ast}(Y,\gamma)$ and for $\alpha_i=\sum_{j=0}^{r} \pi^{\ast}\alpha_{i,j}\gamma^j$, we have \begin{multline} \langle \alpha_1,\alpha_2,\alpha_3 \rangle^{Y}_{0,3,d} =\langle \pi^{\ast}\alpha_{1,0}, \pi^{\ast}\alpha_{2,0}, \pi^{\ast}\alpha_{3,0} \rangle^{Y}_{0,3,d} = \int_{[X_{0,3,d}]^{\vir}} \prod \ev_i^{\ast}(\alpha_{i,0}) \cdot e\bigl(R^1p_{\ast}\mu^{\ast}\nbige^{\lor}\bigr) \\ = \langle \alpha_{1,0},\alpha_{2,0}, \alpha_{3,0} \rangle^{X,\nbige^{\lor},e^{-1}}_{0,3,d} \end{multline} We set $\phi_{ij}:=\rho_i\gamma^j$ $(1\leq i\leq \ell,\,\, 0\leq j\leq r)$ which give a frame of $H^{\ast}(Y)$. Let $\phi^{ij}$ $(1\leq i\leq \ell,\,\, 0\leq j\leq r)$ be the dual base with respect to the Poincar\'e pairing of $Y$. Note that $\phi^{i0}=i_{0\ast}(\rho^i)$. Indeed, it is enough to check $\langle \phi_{kj}, i_{0\ast}(\rho^i) \rangle=0$ $(j>0)$, $\langle \phi_{k0}, i_{0\ast}(\rho^i) \rangle=0$ $(k\neq i)$ and $\langle \phi_{k0}, i_{0\ast}(\rho^i) \rangle=1$, which can be checked by direct computations. Then, we obtain the equality () in Case 1. \vspace{.1in} Let us prove () in Case 2. It is exactly the case given in . We just revisit it in a slightly different way. Let $d\in\Eff(Y,\gamma)$. Note that the virtual dimension of $Y_{0,3,d}$ is $3$. Indeed, because we have $K_Y=\nbigo_Y(2)$, the virtual dimension of $Y_{0,n,d}$ is $n-3+\dim Y-\langle c_1(K_Y),d\rangle=n$. \begin{lem} For $\alpha_i=\alpha_{i0}+\alpha_{i1}\gamma$ $(\alpha_{ij}\in H^{\ast}(X),\,i=1,2,3)$, we have \[ \langle \alpha_1,\alpha_2,\alpha_3 \rangle^{Y}_{0,3,d} =\langle \alpha_{10},\alpha_{20},\alpha_{30} \rangle^Y_{0,3,d}. \] \end{lem} \pf By the dimension reason, we have $\langle\alpha_1,\alpha_2,\alpha_3 \rangle^{Y}_{0,3,d}=0$ unless the cohomological degree of $\alpha_i$ are $2$. Moreover, by the divisor axiom, we have $\langle\alpha_1,\alpha_2,\alpha_3 \rangle^{Y}_{0,3,d}=0$ if one of $\alpha_i$ is $\gamma$. Then, the claim of Lemma follows. \hfill\qed \begin{lem} Suppose that $d\in\Eff^{\ast}(Y,\gamma)$ satisfies $\langle \pi^{\ast}c_1(K_X),d\rangle=0$. Then, we have $\langle \alpha_1,\alpha_2,\alpha_3\rangle_{0,3,d}^Y=0$ for any $\alpha_i\in H^{\ast}(Y)$. \end{lem} \pf It is enough to consider the case where $\alpha_i\in \pi^{\ast}H^2(X)$. Let us observe that the induced morphism $Y_{0,3,d}\lrarr X_{0,3,d}$ is smooth and the fibers are $\proj^1$. Take $(C,g)\in X_{0,3,d}$. Because $K_X^{\lor}$ is nef, and because we have $\langle c_1(K_X),[g(C)]\rangle=0$, the restriction of $g^{\ast}K_X$ to each irreducible component of $C$ is of degree $0$. In particular, we have $H^1(C,g^{\ast}K_X)=0$, and $\dim H^0(C,g^{\ast}K_X)=1$. Let $\nbigc_{0,3,d}$ denote the universal curve over $X_{0,3,d}$. Let $\mu:\nbigc_{0,3,d}\lrarr X$ be the universal morphism. Let $p:\nbigc_{0,3,d}\lrarr X_{0,3,d}$ denote the projection. Then, $p_{\ast}\mu^{\ast}(\nbigo\oplus K_X^{\lor})$ is a locally free sheaf, and the projectivization is naturally isomorphic to $Y_{0,3,d}$. Then, it is easy to see that $[Y_{0,3,d}]^{\vir}$ is the pull back of $[X_{0,3,d}]^{\vir}$ via the natural morphism $q:Y_{0,3,d}\lrarr X_{0,3,d}$. Because $q_{\ast}([Y_{0,3,d}]^{\vir})=0$, we obtain $\langle \alpha_1,\alpha_2,\alpha_3\rangle_{0,3,d}^Y=0$. Thus, we obtain Lemma . \hfill\qed \begin{lem} Let $d\in \Eff(Y,\gamma)$. If $\langle \pi^{\ast}c_1(K_X),d\rangle< 0$, then we have $f(C)\subset i_0(X)$ for any $(C,f)\in Y_{0,3,d}$. In particular, $d$ comes from $\Eff^{\ast}(X)$. \end{lem} \pf Let $(C,f)\in Y_{0,3,d}$. We have $f(C)=\sum m_i[C_i]$ $(m_i>0)$ for algebraic curves in $Y$. Note that $C$ is connected. Suppose $f(C)\cap H_{\infty}\neq \emptyset$. If $f(C)\not\subset H_{\infty}$, we have $\langle \gamma,[C_i]\rangle\geq 0$ for any $i$, and $\langle\gamma,[C_{i_0}]\rangle>0$ for some $i_0$, and hence $\langle \gamma,[f(C)]\rangle>0$ which contradicts with the assumption. If $f(C)\subset H_{\infty}$, we have $\langle \gamma,f(C) \rangle =\langle c_1(K_X^{\lor}), d\rangle>0$ which also contradicts with the assumption. Hence, we have $f(C)\cap H_{\infty}=\emptyset$. Then, $f$ is equivalent to a morphism $g:C\lrarr X$ and a section $s$ of $g^{\ast}(K_X)$. But, because $\langle c_1(K_X),d\rangle<0$, we have $s=0$. Thus, we obtain Lemma . \hfill\qed \vspace{.1in} We can deduce the equality () in Case 2 from Lemma , Lemma and Lemma as in Case 1. Thus, Lemma and Proposition are proved. \hfill\qed \subsection{Associated quantum $\nbigd$-modules} We continue to use the notation in \S. We recall some basic matters on the quantum $\nbigd$-modules of toric varieties. Again, see , , , , , for more details. \subsubsection{Quantum $\nbigd$-modules associated to degenerated quantum products on $H^{\ast}(X)$} We set $\QDM(X,\nbige^{\lor}):= \nbigo_{\cnum_{\lambda}\times\nbigu_X} \otimes H^{\ast}(X)$. As in , we have the meromorphic connection $\nabla^{\nbige^{\lor}}$ on $\QDM(X,\nbige^{\lor})$ which is a variant of the connections of Dubrovin and Givental. It is given as follows. Let $\mu_X$ be the grading operator on $H^{\ast}(X)$ defined by $\mu(a):=ka$ for $a\in H^{2k}(X)$. We can naturally identify $H^{2}(X)\otimes\nbigo_{\nbigu_X}$ with the tangent sheaf of $\nbigu_X$, and so we can regard $\xi\in H^{2}(X)$ as a vector field on $\nbigu_X$. We can naturally regard $b\in H^{\ast}(X)$ as a section of $\QDM(X,\nbige^{\lor})$. Then, $\nabla^{\nbige^{\lor}}$ is determined by the following on $\cnum^{\ast}\times \nbigu_X$: \[ \bigl( \nabla^{\nbige^{\lor}}_{\xi}b \bigr)_{(\lambda,\sigma)} =-\lambda^{-1}\xi\bullet^{\nbige^{\lor}}_{\sigma}b, \quad \bigl( \nabla^{\nbige^{\lor}}_{\lambda\del_{\lambda}}b \bigr)_{(\lambda,\sigma)} =\lambda^{-1}E\bullet^{\nbige^{\lor}}_{\sigma}b +\mu_X(b), \quad\quad ((\lambda,\sigma)\in \cnum^{\ast}\times \nbigu_X). \] The filtrations $\Wtilde$ of $Q^{\nbige^{\lor}}_{\sigma}H^{\ast}(X)$ $(\sigma\in\nbigu_X)$ give a filtration $\Wtilde$ on $\QDM(X,\nbige^{\lor})$. It is preserved by the connection $\nabla^{\nbige^{\lor}}$. We have the induced meromorphic connection on $\Gr^{\Wtilde}_k\QDM(X,\nbige^{\lor})$, denoted by the same notation. We regard $b_i\in \Gr^{\Wtilde}_{k}\Cok(N)$ $(i=1,2)$ as sections of $\Gr^{\Wtilde}_k\QDM(X,\nbige^{\lor})$. We set \[ \nbigp^{\nbige^{\lor}}_k(b_1,j^{\ast}b_2) :=\lambda^{-k}P^{\nbige^{\lor}}_k(b_1,b_2). \] Thus, we obtain a morphism $\nbigp^{\nbige^{\lor}}_k: \Gr^{\Wtilde}_k\QDM(X,\nbige^{\lor}) \otimes j^{\ast} \Gr^{\Wtilde}_k\QDM(X,\nbige^{\lor}) \lrarr \lambda^{-k} \nbigo_{\cnum\times\nbigu_X}$. We will give a proof of the following lemma by using the description as a specialization of the quantum $\nbigd$-module of $Y$ in \S. \begin{lem} $\nabla^{\nbige^{\lor}}$ is flat, and $\nbigp^{\nbige^{\lor}}_k$ are pairings of weight $k$ on $\bigl( \Gr^{\Wtilde}_k\QDM(X,\nbige^{\lor}),\nabla^{\nbige^{\lor}} \bigr)$. In other words, $\bigl( \QDM(X,\nbige^{\lor}),\Wtilde, \bigl\{ \nbigp^{\nbige^{\lor}}_k \bigr\} \bigr)$ is a mixed TEP-structure. \end{lem} We also set $\nbigk(X,\nbige^{\lor}):= \nbigo_{\cnum_{\lambda}\times\nbigu_X} \otimes H^{\ast}(X)$. As in the case of $\QDM(X,\nbige^{\lor})$, it is equipped with the meromorphic flat connection $\nabla^{\nbige^{\lor}}$ satisfying the following: \[ \bigl( \nabla^{\nbige^{\lor}}_{\xi}b \bigr)_{(\lambda,\sigma)} =-\lambda^{-1}\xi\bullet^{\nbige^{\lor}}_{K,\sigma}b, \quad \bigl( \nabla^{\nbige^{\lor}}_{\lambda\del_{\lambda}}b \bigr)_{(\lambda,\sigma)} =\lambda^{-1}E\bullet^{\nbige^{\lor}}_{K,\sigma}b +\mu_X(b)+rb, \quad \quad ((\lambda,\sigma)\in \cnum^{\ast}\times\nbigu_X) \] Here, $b\in H^{\ast}(X)$ and $\xi\in H^{2}(X)$. The filtration $\Wtilde$ is preserved by $\nabla^{\nbige^{\lor}}$. We obtain a morphism $\nbigp_k^{\nbige^{\lor}}: \Gr^{\Wtilde}_k\nbigk(X,\nbige^{\lor}) \times j^{\ast} \Gr^{\Wtilde}_k\nbigk(X,\nbige^{\lor}) \lrarr \lambda^{-k}\nbigo_{\cnum\times\nbigu_X}$ by the formula: \[ \nbigp^{\nbige^{\lor}}_k(b_1,j^{\ast}b_2) =(-1)^{k-n-r}\lambda^{-k} P^{\nbige^{\lor}}_k(b_1,b_2) \] We will give a proof of the following lemma by using the description as a specialization of the quantum $\nbigd$-module of $Y$ in \S. \begin{lem} $\nabla^{\nbige^{\lor}}$ is flat, and $\nbigp_k^{\nbige^{\lor}}$ are pairings of weight $k$ on $\bigl( \Gr^{\Wtilde}_k\nbigk(X,\nbige^{\lor}), \nabla^{\nbige^{\lor}} \bigr)$. In other words, $\bigl( \nbigk(X,\nbige^{\lor}),\Wtilde, \{\nbigp_k^{\nbige^{\lor}}\} \bigr)$ is a mixed TEP-structure. \end{lem} \subsubsection{Quantum $\nbigd$-module of $Y$} We recall the quantum $\nbigd$-module associated to the quantum products of $Y$. We set $\nbigu_{\gamma 1}:= \nbigu_{\gamma}/2\pi\sqrt{-1}\seisuu\gamma$. We set $\nbigu_{Y1}:=\nbigu_X\times\nbigu_{\gamma 1}$ which is the quotient of $\nbigu_{Y}$ by the action of $2\pi\sqrt{-1}\seisuu\gamma$. We embed $\nbigu_{\gamma 1}\lrarr \cnum^{\ast}$ induced by $c\longmapsto \exp(c)$. The quantum products $\bullet_{\sigma,c}$ depends only on $(\sigma,e^c)$. So, we denote them by $\bullet_{\sigma,e^c}$. We set $\QDM(Y) :=\nbigo_{\cnum_{\lambda}\times\nbigu_{Y1}}\otimes H^{\ast}(Y)$. Recall that we have the meromorphic flat connection $\nabla^G$ on $\QDM(Y)$ due to Dubrovin and Givental. Let $\mu_Y$ be the grading operator on $H^{\ast}(Y)$, given by $\mu_Y(b):=kb$ for $b\in H^{2k}(Y)$. We naturally regard $\xi\in H^{2}(Y)$ as a vector field on $\nbigu_Y$, which induces a vector field on $\nbigu_{Y1}$. We also naturally regard $b\in H^{\ast}(Y)$ as a section of $\QDM(Y)$. Then, $\nabla^G$ is determined by the following on $\cnum_{\lambda}^{\ast}\times\nbigu_{Y1}$: \[ \bigl( \nabla^G_{\xi}b \bigr)_{(\lambda,\sigma,t)} =-\lambda^{-1}\xi\bullet_{(\sigma,t)}b, \quad\quad \bigl( \nabla^G_{\lambda\del_{\lambda}}b \bigr)_{(\lambda,\sigma,t)} =\lambda^{-1}E\bullet_{(\sigma,t)}b +\mu_Y(b), \quad\quad ((\lambda,\sigma,t)\in\cnum^{\ast}\times\nbigu_{Y1}). \] It is equipped with the pairing $\nbigp_Y$ of weight $-n-r$ induced by the Poincar\'e pairing: \begin{equation} \nbigp_Y(b_1,j^{\ast}b_2):=\lambda^{n+r}(b_1,b_2) \end{equation} It is well known that $\nbigp_Y$ is $\nabla^G$-flat. It is clearly equivariant with respect to the action of $2\pi\sqrt{-1}H^2(X,\seisuu)$. \subsubsection{Logarithmic extension and the specialization} Put $\nbigubar_{\gamma 1}:=\nbigu_{\gamma 1}\cup\{0\}$ which is a neighbourhood of $0$ in $\cnum$. We set $\nbigubar_{Y1}:=\nbigu_X\times\nbigubar_{\gamma 1}$. We set $\QDMbar(Y):= \nbigo_{\cnum_{\lambda}\times\nbigubar_{Y1}}\otimes H^{\ast}(Y)$. We naturally regard $\QDMbar(Y)_{|\nbigu_{Y1}}=\QDM(Y)$. The meromorphic flat connection $\nabla^G$ of $\QDM(Y)$ naturally gives a meromorphic flat connection on $\QDMbar(Y)$ which is logarithmic along $t=0$. The pairing $\nbigp_Y$ is naturally extended to the following morphism which is also denoted by $\nbigp_Y$: \[ \QDMbar(Y) \otimes j^{\ast}\QDMbar(Y) \lrarr \lambda^{n+r}\nbigo_{\cnum_{\lambda}\times\nbigubar_{Y1}} \] We set $\QDM(Y)^{\sp}:= \nbigo_{\cnum_{\lambda}\times\nbigu_X} \otimes H^{\ast}(Y)$. We naturally have \[ \QDM(Y)^{\sp}=\QDMbar(Y)_{|\cnum_{\lambda}\times\nbigu_X\times{0}}. \] The residue $\Res_{t}(\nabla^{G}): \lambda\QDM(Y)^{\sp} \lrarr \QDM(Y)^{\sp}$ is given by the multiplication of $-\lambda^{-1}\gamma$. The specialization of the connection $\nabla^{G\sp}$ is induced by the degenerated quantum products: \[ \bigl( \nabla^{G\sp}_{\xi}b \bigr)_{(\lambda,\sigma)} =-\lambda^{-1}\xi\bullet_{\sigma}b, \quad\quad \bigl( \nabla^{G\sp}_{\lambda\del_{\lambda}}b \bigr)_{(\lambda,\sigma)} =\lambda^{-1}E\bullet_{\sigma}b +\mu_Y(b), \quad\quad \bigl( (\lambda,\sigma)\in \cnum_{\lambda}^{\ast}\times\nbigu_X \bigr) \] It is equipped with the pairing $\nbigp_Y$ of weight $-n-r$ given by the formula (). \vspace{.1in} By the construction $\psi_{t,-\vecdelta}\bigl(\QDMbar(Y)(\ast t)\bigr)$ is naturally isomorphic to $\QDM(Y)^{\sp}$ with $\nabla^{G\sp}$. By the construction, we have the following isomorphisms which are compatible with the meromorphic connections: \[ \Cok\bigl(\lambda\Res_t(\nabla^G)\bigr) \simeq \QDM(X,\nbige^{\lor}), \quad\quad \Ker\bigl(\lambda\Res_t(\nabla^G)\bigr) \simeq \nbigk(X,\nbige^{\lor}) \] By the construction, we have the following which implies Lemma and Lemma . \begin{lem} The weight filtration $\Wtilde$ on $\QDM(X,\nbige^{\lor})$ and the pairings $\nbigp^{\nbige^{\lor}}_k$ on $\Gr^{W(N)}_k\QDM(X,\nbige^{\lor})$ are constructed from $\bigl(\QDM(Y),\nabla^G,\nbigp_Y\bigr)$ by the procedure in {\rm \S}. \hfill\qed \end{lem} \subsection{Local mirror isomorphism} Recall that the isomorphism $L_X^{\lor}\simeq H^2(X,\seisuu)$ is induced by $M_X^{\lor}\lrarr H^2(X,\seisuu)$ given by $\sum_{i=1}^m a_ie_i^{\lor}\longmapsto \sum a_i[D_i]$, where $e_i^{\lor}$ $(i=1,\ldots,m)$ is the dual frame of the standard basis $e_1,\ldots,e_m$. Let $\eta_1,\ldots,\eta_{\ell}$ be a frame of $H^2(X,\seisuu)$ such that the first Chern class of $K_X^{\lor}\otimes \bigotimes_{i=1}^r\nbigl_i^{\lor}$ is contained in the cone generated by $\eta_i$ $(i=1,\ldots,\ell)$. We can check the existence of such a frame in an elementary way. We have a frame $\xi_1,\ldots,\xi_{\ell}$ of $L_X^{\lor}$ corresponding to $\eta_1,\ldots,\eta_{\ell}$ by the above isomorphism. For $M\in\real$, we set \[ \nbigb_M(\veceta):= \Bigl\{ \sum_{i=1}^{\ell} \alpha_i\eta_i\in H^2(X,\cnum) \,\Big|\, \Re(\alpha_i)<-M\,\,(i=1,\ldots,\ell) \Bigr\}, \] \[ \nbigb'_M(\vecxi):=\Bigl\{ \sum_{i=1}^{\ell} \beta_i\xi_i\in \gbigs_A=L_X^{\lor}\otimes\cnum \,\Big|\, \Re(\beta_i)<-M\,\,(i=1,\ldots,\ell) \Bigr\}. \] We obtain the following theorem as a corollary of the theorem of Givental, according to which we have isomorphisms between the mixed TEP-structures associated to local A-models and local B-models in . It is also related with \cite[Conjecture 6.14]{Reichelt-Sevenheck2}. We use the notation in \S and \S. \begin{thm} We have the following: \begin{itemize} \item An open subset $U_1\subset \nbigu_X\subset H^2(X,\cnum)$; It contains $\nbigb_{M_1}(\veceta)$ for a positive number $M_1$; It is preserved by the natural action of $2\pi\sqrt{-1}H^2(X,\seisuu)$. \item An open subset $U_2\subset \gbigs_{\nbiga}^{\reg}$; It contains $\nbigb_{M_2}(\vecxi)$ for a positive number $M_2$; It is preserved by the natural action of $(2\pi\sqrt{-1}\seisuu) L_X^{\lor}$. \item A holomorphic isomorphism $\varphi:U_1\simeq U_2$; It preserves the Euler vector fields. \item Isomorphism of mixed TEP-structures \[ \Phi_{\ast}: \bigl( \QDM(X,\nbige^{\lor}), \Wtilde, \{\alpha \nbigp^{\nbige^{\lor}}_k\,|\,k\in\seisuu\} \bigr)_{|U_1} \simeq \varphi^{\ast} \chi_{\nbiga}^{\ast} \bigl( \gbigv_{\nbiga,\ast}, \Wtilde, \{\gbigp_{\nbiga,\ast,k}\,|\,k\in\seisuu\} \bigr)_{|U_2} \] \[ \Phi_!:\bigl( \nbigk(X,\nbige^{\lor}), \Wtilde, \{\alpha \nbigp^{\nbige^{\lor}}_k\,|\,k\in\seisuu\} \bigr)_{|U_1} \simeq \varphi^{\ast} \chi_{\nbiga}^{\ast} \bigl( \gbigv_{\nbiga,!}, \Wtilde, \{\gbigp_{\nbiga,!,k}\,|\,k\in\seisuu\} \bigr)_{|U_2} \] Here, $\alpha$ is a non-zero complex number. We also have the following commutative diagram of mixed TE-structures: \[ \begin{CD} \bigl( \nbigk(X,\nbige^{\lor}),\Wtilde \bigr)_{|U_1} @>{\Phi_{!}}>{\simeq}> \varphi^{\ast}\chi_{\nbiga}^{\ast} (\gbigv_{\nbiga!},\Wtilde)_{|U_2} \\ @VVV @VVV \\ \bigl( \QDM(X,\nbige^{\lor}),\Wtilde \bigr)_{|U_1} @>{\Phi_{\ast}}>{\simeq}> \varphi^{\ast}\chi_{\nbiga}^{\ast} (\gbigv_{\nbiga\ast},\Wtilde)_{|U_2} \end{CD} \] \end{itemize} The isomorphisms $\varphi$ and $\Phi_{\star}$ $(\star=\ast,!)$ are equivariant with respect to the actions of $2\pi\sqrt{-1}H^2(X,\seisuu) \simeq (2\pi\sqrt{-1}\seisuu)L_X^{\lor}$. \end{thm} \pf Let $B_{\epsilon}(t):=\bigl\{|t|<\epsilon\bigr\}$. According to , as reformulated in and , we have the following. \begin{itemize} \item An open subset $U_Y\subset \nbigu_Y$; It contains $\nbigb_{M_1}(\veceta)\times B_{\epsilon_1}(t_1)$ for a large $M_1$ and a small $\epsilon_1>0$; It is preserved by the action of $2\pi\sqrt{-1}H^2(X,\seisuu)$. \item An open subset $U_{\nbigatilde}\subset \gbigstilde_{\nbiga} =L_{\nbigatilde}\otimes\cnum$; It contains $\nbigb_{M_2}(\vecxi)\times B_{\epsilon_2}(t_2)$; It is preserved by the action of $2\pi\sqrt{-1}L_X^{\lor}$; The restriction of $\nbigv_{\nbiga}$ to $U_{\nbigatilde}$ is locally free $\nbigo_{\cnum_{\lambda}\times U_{\nbigatilde}}(\ast t_2)$-module. \item A holomorphic isomorphism $\widehat{\varphi}:U_{Y}\simeq U_{\nbigatilde}$; It preserves the Euler vector fields; We have $\widehat{\varphi}^{-1}\{t_2=0\}=\{t_1=0\}$. \item An isomorphism of TEP-structures $\widehat{\Phi}: \bigl( \QDMbar(Y)(\ast t), \alpha\nbigp_Y \bigr)_{|U_Y} \simeq \widehat{\varphi}^{\ast} \chitilde_{\nbiga}^{\ast} \bigl( \gbigvtilde_{\nbiga},\gbigptilde_{\nbiga} \bigr)_{|U_{\nbigatilde}}$. \item The isomorphisms $\widehat{\varphi}$ and $\widehat{\Phi}$ are equivariant with respect to the action of $(2\pi\sqrt{-1})L_X^{\lor} \simeq H^2(X,\seisuu)$. \end{itemize} We set $U_1:=U_Y\cap\{t_1=0\}$ and $U_2:=U_{\nbigatilde}\cap\{t_2=0\}$. As the restriction of $\widehat{\varphi}$, we obtain a holomorphic isomorphism $\varphi:U_1\simeq U_2$ which is equivariant with respect to the action of the lattices. It also preserves the Euler vector fields. Then, the claim follows from Lemma and Lemma . \hfill\qed \begin{rem} If the Euler vector fields are $0$, then the $\nbigrtilde$-modules are $\cnum^{\ast}$-homogeneous with respect to the trivial action on the base space. So, they are equipped with the Hodge filtration. Because the Hodge filtration can be recovered from the connection, the isomorphisms in Theorem {\rm} also preserve the Hodge filtrations. \hfill\qed \end{rem} \begin{rem} If $r=1$ and $\nbigl_1=K_X^{\lor}$, the mixed TEP-structure is expressed as in {\rm\S} in terms of the variation of Hodge structure associated to $\nbigb$. Here, $\nbigb$ is related to $\nbiga$ as in {\rm\S.} \hfill\qed \end{rem} \begin{rem} We set $U_1':=U_1\big/\bigl(2\pi\sqrt{-1}H^2(X,\seisuu)\bigr)$ and $U_2':=U_2\big/\bigl(2\pi\sqrt{-1}L_X^{\lor}\bigr)$. We have the induced isomorphism $\varphi':U_1'\simeq U_2'$. We may regard $U_1'\subset H^2(X,\seisuu)\otimes\cnum^{\ast}$ and $U_2'\subset L_X^{\lor}\otimes\cnum^{\ast}$. They are naturally extended to neighbourhoods $\overline{U}_1'\subset H^2(X,\cnum)$ and $\overline{U}_2'\subset L_X^{\lor}\otimes\cnum$ of $0$. According to {\rm }, the isomorphism $\varphi'$ is extended to an isomorphism $\overline{U}_1'\simeq \overline{U}_2'$. We can consider the descent of the objects and the isomorphisms in Theorem {\rm} to $U_i'$. The objects are extended on $\overline{U}_i'$ in the logarithmic way such that the residues endomorphisms are nilpotent. Then, by the uniqueness of such extensions, the isomorphisms given on $U_i'$ are extended to the isomorphisms on $\overline{U}_i'$. \hfill\qed \end{rem} \begin{rem} In {\rm}, it is announced that a result in {\rm } implies the comparison of the TE-structures in Theorem {\rm}. At this moment, it is not clear to the author if we could also deduce the comparison of the weight filtrations from {\rm }. \hfill\qed \end{rem} \subsection{Isomorphism} Let $X$ be a complex manifold with a hypersurface $H$. Let $d_X:=\dim X$. Let $M$ be a reflexive meromorphic flat connection on $(X,H)$, i.e., $M$ is a $\nbigd_X$-module and that $M$ is a coherent and reflexive $\nbigo_X(\ast H)$-module. We set $M^{\lor}:= \nhom_{\nbigo_X(\ast H)} \bigl(M,\nbigo_X(\ast H)\bigr)$ which is naturally a reflexive meromorphic flat connection. In this section, let $\DDD$ denote the duality functor on the category of holonomic $\nbigd_X$-modules. As well known, $\DDD_{X(\ast H)}(M):= \DDD(M)(\ast H)$ is isomorphic to $M^{\lor}$ as a meromorphic flat connection. Let us recall the construction of the isomorphism $\nu_{M}: \DDD_{X(\ast H)}(M)\stackrel{\simeq}{\lrarr} M^{\lor}$ as in . Let $\Theta_X$ denote the tangent sheaf of $X$. We set $\Theta^{-p}_X:=\bigwedge^p\Theta_X$ for $p\geq 0$. We have the Spencer resolution $\nbigd_X\otimes\Theta^{\bullet}_X$ of $\nbigo_X$ by a locally free left $\nbigd$-modules. We set $\Omega_X:=\Omega_X^{d_X}$. The dual line bundle is denoted by $\Omega_X^{-1}$. We have the following natural identification: \[ \DDD_{X(\ast H)}(M) \simeq \nhom_{\nbigd_X(\ast H)} \bigl( \nbigd_X\otimes\Theta_X^{\bullet}\otimes M, \nbigd_X\otimes\Omega^{-1}_X(\ast H) \bigr)[d_X]. \] The degree $0$ term in the right hand side is $\nhom_{\nbigd_X(\ast H)} (\nbigd_X\otimes\Omega_X^{-1}\otimes M, \nbigd_X\otimes\Omega_X^{-1}(\ast H)) \simeq \nbigd_X\otimes M^{\lor}$. The isomorphism $\nu_{M}$ is induced by the canonical isomorphism $\DDD_{X(\ast H)}(M) \simeq \nbigd_X\otimes\Theta^{\bullet}\otimes M^{\lor}$ whose degree $0$-term is the identity. We have another description of the isomorphism $\nu_{M}$ in the case $H=\emptyset$. Let $L_{M^{\lor}}$ be the sheaf of flat sections of $M^{\lor}$ which is a $\cnum$-local system. We have a natural isomorphism $\DR\DDD_{X} M= \nrhom_{\nbigd_X}(M,\nbigo_X[d_X]) \simeq L_{M^{\lor}}[d_X]$. We have $\DR(\nu_{M})=(-1)^{d_X}\id$ which characterizes $\nu_{M}$. Under the natural identification $\DDD_{X(\ast H)}\circ \DDD_{X(\ast H)}(M)=M$, we have the induced morphism $\DDD_{X(\ast H)}\nu_{M}: \DDD_{X(\ast H)}(M^{\lor})\simeq M$. We shall use the following lemma implicitly. (See .) \begin{lem} We have $\DDD_{X(\ast H)}\nu_{M} =(-1)^{d_X}\nu_{M^{\lor}}$. \hfill\qed \end{lem} \subsubsection{A commutative diagram} For any morphism of meromorphic flat bundles $f:M_1\lrarr M_2$ on $(X,H)$, the following diagram is commutative: \[ \begin{CD} \DDD_{X(\ast H)} M_2@>{\DDD_{X(\ast H)} f}>> \DDD_{X(\ast H)} M_1\\ @V{\nu_{M_2}}VV @V{\nu_{M_1}}VV \\ M_2^{\lor} @>{f^{\lor}}>> M_1^{\lor} \end{CD} \] Let $M_i$ be reflexive meromorphic flat connections on $(X,H)$. Let $\rho:M_1\lrarr \DDD_{X(\ast H)} M_2$ be a morphism of $\nbigd$-modules. We have the induced morphisms $\nu_{M_2}\circ\rho: M_1\lrarr M_2^{\lor}$ and $\nu_{M_1}\circ\DDD_{X(\ast H)}\rho: M_2\lrarr M_1^{\lor}$. We shall use the following lemma in Proposition . \begin{lem} We have $\nu_{M_1}\circ\DDD_{X(\ast H)}\rho =(-1)^{d_X}(\nu_{M_2}\circ\rho)^{\lor}$. \end{lem} \pf We have the following commutative diagram: \[ \begin{CD} \DDD_{X(\ast H)} (M_1) @<{\DDD_{X(\ast H)}(\nu_{M_2}\circ\rho)}<< \DDD_{X(\ast H)} (M_2^{\lor}) \\ @V{\nu_{M_1}}VV @VVV \\ M_1^{\lor} @<{(\nu_{M_2}\circ\rho)^{\lor}}<< M_2 \end{CD} \] Here, the right vertical arrow is the composite of $\DDD_{X(\ast H)} (M_2^{\lor}) \stackrel{\nu_{M_2^{\lor}}}\simeq (M_2^{\lor})^{\lor} \simeq M_2$. Note that $\nu_{M_1}\circ\DDD_{X(\ast H)} \rho$ precisely means the composite of $M_2 \simeq \DDD_{X(\ast H)} (\DDD_{X(\ast H)} M_2) \stackrel{\DDD_{X(\ast H)} \rho}{\lrarr} \DDD_{X(\ast H)} M_1 \stackrel{\nu_{M_1}}{\simeq} M_1^{\lor}$. We remark that the composite of the following natural isomorphisms is the multiplication of $(-1)^{d_X}$. \[ \begin{CD} M_2 @>>> (M_2^{\lor})^{\lor} @>>> \DDD_{X(\ast H)}(M_2^{\lor}) @>{\DDD_{X(\ast H)}\nu_{M_2}}>> \DDD_{X(\ast H)} (\DDD_{X(\ast H)} M_2) @>>>M_2 \end{CD} \] Then, the claim of Lemma follows. \hfill\qed \subsection{Duality and localizations of flat bundles} This subsection is a preliminary for \S and \S. \subsubsection{Statement} Let $X$ be a complex manifold with a smooth hypersurface $Y$. Set $d_X:=\dim X$. Let $\iota:Y\lrarr X$ be the inclusion. Let $M$ be a flat bundle on $X$, i.e., a locally free $\nbigo_X$-module with a flat connection. The pull back of $M$ to $Y$ is denoted by $M_Y$. For any $\nbigd_X$-module $N$, we set $N(\ast Y):=N\otimes\nbigo_X(\ast Y)$ and $N(!Y):=\DDD_X\bigl(\DDD_X(N)(\ast Y)\bigr)$. We have the natural morphisms $\rho_{M,1}:M(\ast Y) \lrarr \iota_{+}M_Y$ and $\rho_{M,2}:\iota_{+}M_Y \lrarr M(!Y)$. Locally, they are given as follows. Let $(x_1,\ldots,x_n)$ be a holomorphic local coordinate system of $X$ such that $Y=\{x_1=0\}$. We have $\iota_{+}M_Y \simeq \bigoplus_{n=0}^{\infty} \iota_{\ast}M_Y(dx_1)^{-1}\,\del_{x_1}^n$. Then, we have $\rho_{M,1}(f\cdot x_1^{-1}) =\iota_{\ast}(f_{|Y}(dx_1)^{-1})$ for any section $f$ of $M$. And, we have $\rho_{M,2}(g\iota_{\ast}(dx_1)^{-1}) =-\del_{x_1}(\gtilde)$ in $M(!Y)$ for any section $g$ of $M_Y$, where $\gtilde$ is a section of $M$ satisfying $\gtilde_{|Y}=g$ and $\del_{x_1}(\gtilde)=0$ in $M$. They are independent of the choice of coordinate systems, and we have the global morphisms. The isomorphism $\nu_M:\DDD M\simeq M^{\lor}$ induces $\nu_{M\ast}: \DDD\bigl( M(\ast Y)\bigr) \simeq M^{\lor}(!Y)$ and $\nu_{M!}: \DDD\bigl( M(!Y)\bigr) \simeq M^{\lor}(\ast Y)$. We have $\nu_{M_Y}:\DDD M_Y\lrarr M^{\lor}_Y$ determined by a similar condition. The composite $\DDD\iota_{+}M_Y \simeq \iota_{+}\DDD M_Y \simeq \iota_{+}M_Y^{\lor}$ is denoted by $\iota_{+}\nu_{M_Y}$. We shall prove the following proposition in \S--. \begin{prop} We have $\nu_{M\ast}\circ\DDD \rho_{M,1} =-\rho_{M^{\lor},2}\circ\iota_{+}\nu_{M_Y}$ and $\iota_{+}\nu_{M_Y}\circ \DDD\rho_{M,2} =\rho_{M^{\lor},1}\circ \nu_{M!}$. \end{prop} Clearly, it is enough to prove the case $M=\nbigo_X$ for which we naturally identify $M^{\lor}=\nbigo_X$. We shall $\rho_{M,i}$ by $\rho_i$, and $\nu_{M_X}$ and $\nu_{M_Y}$ by $\nu_X$ and $\nu_Y$, respectively. We set $d_Y:=\dim Y=d_X-1$. Because $\DDD\nu_X=(-1)^{d_X}\nu_X$ and $\DDD\nu_Y=(-1)^{d_Y}\nu_Y$, the two equalities are equivalent. Hence, it is enough to prove $\iota_{+}\nu_{Y}\circ \DDD\rho_2 =\rho_1\circ \nu_{X!}$. We have only to prove the claim locally around any point of $Y$. We may assume that $X=\cnum_t\times Y$. We have natural isomorphisms $\nbigo_X \simeq \nbigo_{\cnum_t}\boxtimes\nbigo_Y$ and $\DDD\nbigo_X \simeq \DDD\nbigo_{\cnum_t} \boxtimes \DDD\nbigo_Y$. The equality $\iota_{+}\nu_{Y}\circ \DDD\rho_2 =\rho_1\circ \nu_{X!}$ means the commutativity of the following diagram: \[ \begin{CD} \DDD\bigl( \nbigo_{\cnum_t}(!t) \bigr) \boxtimes \DDD\nbigo_Y @>{\nu_{\cnum_t!}\boxtimes\nu_Y}>> \nbigo_{\cnum_t}(\ast t) \boxtimes \nbigo_Y\\ @VV{\DDD\rho_2\boxtimes\id}V @VV{\rho_1\boxtimes\id}V \\ \DDD(\iota_{+}\nbigo_{\{0\}}) \boxtimes \DDD\nbigo_Y @>{\iota_{+}\nu_{\{0\}}\boxtimes\nu_Y}>> \iota_{+}\nbigo_{\{0\}} \boxtimes \nbigo_Y \end{CD} \] Hence, it is enough to consider the case $\dim X=1$, i.e., $X=\cnum_t$ and $Y=\{0\}$, which we assume in the following. \subsubsection{Commutativity of duality and push-forward} First, let us describe the isomorphism $\DDD\iota_{+}\nbigo_Y \simeq \iota_{+}\DDD\nbigo_Y$. We naturally have $\iota_{+}\nbigo_Y =\bigoplus_{j=0}^{\infty} \iota_{\ast} \cnum \cdot\del_t^{j}(dt)^{-1}$. We have the exact sequence $0\lrarr\nbigd_X\stackrel{\kappa_0}{\lrarr} \nbigd_X \stackrel{\kappa_1}{\lrarr} \iota_{+}\nbigo_Y \lrarr 0$, where $\kappa_0(P)=P\cdot t$ and $\kappa_1(1)=\iota_{\ast}(dt)^{-1}$. It induces the exact sequence: \[ 0\lrarr \nhom_{\nbigd_X}(\nbigd_X,\nbigd_X\otimes\Omega_X^{-1}) \stackrel{\kappa_0'}\lrarr \nhom_{\nbigd_X}(\nbigd_X,\nbigd_X\otimes\Omega_X^{-1}) \stackrel{\kappa_1'}\lrarr \DDD(\iota_{+}\nbigo_Y) \lrarr 0 \] Here, $\kappa_0'(g)(P)=g(Pt)$. Under the natural identifications $\nhom_{\nbigd_X}(\nbigd_X,\nbigd_X\otimes\Omega_X^{-1}) \simeq \nbigd_X\otimes\Omega_X^{-1}$, we have $\kappa_0'(Q)=tQ$. We naturally have $\iota_{+}\DDD\nbigo_Y =\bigoplus_{j=0}^{\infty} \iota_{\ast}\Hom(\cnum,\cnum)\cdot \del_t^j\iota_{\ast}(dt)^{-1}$. \begin{lem} The isomorphism $\DDD(\iota_{+}\nbigo_Y) \simeq \iota_{+}\DDD(\nbigo_Y)$ is given by $\kappa_1'((dt)^{-1}) \longmapsto -\iota_{\ast}\bigl(\id\cdot(dt)^{-1}\bigr)$. \end{lem} \pf Note that the following diagram is commutative, due to M. Saito (see also \cite[Proposition 12.4.25]{Mochizuki-MTM}): \[ \begin{CD} \DR(\DDD\iota_{+}\nbigo_Y) @>{\simeq}>> \DR(\iota_{+}\DDD\nbigo_Y) \\ @VVV @VVV\\ \DDD \iota_{\ast}\DR\nbigo_Y @>{\simeq}>> \iota_{\ast}\DDD\DR\nbigo_Y \end{CD} \] Here, the vertical arrows are the exchange of the de Rham functors with the others, and the horizontal arrows are the exchange of the duality functor and the push-forward. We describe the induced morphism $\DR\DDD\iota_{+}\nbigo_Y \simeq \iota_{\ast}\DDD\DR\nbigo_Y$. We set $\omega_X^{top}=\distribution_X^{\bullet}[2]$ and $\omega_Y^{top}=\cnum_Y$. Then, it is described as the composite of the following: \begin{multline} \nrhom_{\nbigd_X}\bigl( \iota_{+}\nbigo_Y,\nbigo_X[1] \bigr) \stackrel{a}{\simeq} \nrhom_{\cnum_X}\bigl( \DR\iota_{+}\nbigo_Y, \omega_X^{top} \bigr) \stackrel{b}{\simeq} \nrhom_{\cnum_X} \bigl( \iota_{\ast}\DR\nbigo_Y, \omega_X^{top} \bigr) \\ \stackrel{c}{\simeq} \nhom_{\cnum_X} (\iota_{\ast}\DR \nbigo_Y,\iota_{\ast}\omega_Y^{top}) =\iota_{\ast}\Hom(\cnum,\cnum). \end{multline} Let us describe the morphisms more explicitly. We have $\nbigo_X[1]=\distribution_X^{0,\bullet}[1] =\bigl( \distribution_X^{0,0} \stackrel{-\delbar}{\lrarr} \distribution_X^{0,1} \bigr)$, where $\distribution_X^{0,0}$ sits in the degree $-1$. We have \[ \nrhom_{\nbigd_X}(\iota_{+}\nbigo_Y,\nbigo_X[1]) \simeq \nhom_{\nbigd_X}(\iota_{+}\nbigo_Y, \distribution_X^{0,\bullet}[1]) \simeq \distribution_{X,Y}^{0,\bullet}[1], \] where $\distribution_{X,Y}^{0,i}$ denotes the sheaf of $(0,i)$-currents on $X$ whose supports are contained in $Y$. The last isomorphism is given by $\Psi\longmapsto \Psi(\iota_{\ast}(dt)^{-1})$. We have the section $-\delbar(t^{-1})$ of $\distribution^{0,1}_{X,Y}$. Recall that, for a complex manifold $Z$ with $\dim Z=d_Z$, we identify the total complex of $\Omega_Z^{\bullet}[d_Z] \otimes\distribution_Z^{0,\bullet}[d_Z]$ with $\distribution_Z^{\bullet}[2d_Z]$ by the correspondence $\eta^{d_Z+i}\otimes \etabar^{d_Z+j} \longmapsto (-1)^{d_Z(d_Z-1)/2+d_Zi} \eta^{d_Z+i}\wedge\etabar^{d_Z+j}$. We have \[ a\bigl( -\delbar(1/t) \bigr) \bigl( \del_t^j\iota_{\ast}(dt)^{-1}dt \bigr) =-\del_t^jdt\delbar(t^{-1}) =\del_t^j\delbar(dt/t), \quad a\bigl( -\delbar(1/t) \bigr) \bigl( \del_t^j\iota_{\ast}(dt)^{-1} \bigr) =\del_t^j\delbar(t^{-1}). \] Hence, we have $(b\circ a)(-\delbar(t^{-1}))\bigl( \iota_{\ast}(dt)^{-1}dt \bigr) =\delbar(dt/t)$. Note that $c$ is induced by the trace morphism $\iota_{\ast}\omega_Y^{top}\lrarr \omega_X^{top}$, for which we have $\iota_{\ast}(1) \longmapsto \delbar(t^{-1}dt)$. Hence, we have $(c\circ b\circ a)(-\delbar(t^{-1}))=\iota_{\ast}1$. \vspace{.1in} We have the quasi-isomorphism: \begin{equation} \nhom(\iota_{+}\nbigo_Y,\distribution_X^{0,\bullet}[1]) \simeq \DR\DDD(\iota_{+}\nbigo_Y) \simeq \Bigl( \nhom_{\nbigd_X}(\nbigd_X,\nbigo) \stackrel{t}{\lrarr} \nhom_{\nbigd_X}(\nbigd_X,\nbigo) \Bigr) \end{equation} The complexes () are naturally quasi-isomorphic to the total complex associated to the following double complex: \[ \begin{CD} \nhom_{\nbigd_X}(\nbigd_X,\distribution_X^{0,0}) @>{t}>> \nhom_{\nbigd_X}(\nbigd_X,\distribution_X^{0,0}) \\ @V{-\delbar}VV @V{-\delbar}VV \\ \nhom_{\nbigd_X}(\nbigd_X,\distribution_X^{0,1}) @>{t}>> \nhom_{\nbigd_X}(\nbigd_X,\distribution_X^{0,1}) \end{CD} \quad\quad \left( \begin{CD} (0,-1) @>>> (1,-1)\\ @VVV @VVV \\ (0,0) @>>> (1,0) \end{CD} \right) \] Here, the right square in the parenthesis indicates the degrees of the double complex. Then, we can easily deduce $\delbar(t^{-1})$ in $\nhom(\iota_+\nbigo_Y, \distribution_X^{0,1})$ is equal to $1$ in $\nhom_{\nbigd_X}(\nbigd_X,\nbigo)$ in the cohomology of (). Hence, the isomorphism $\DR\iota_{+}\DDD\nbigo_Y \simeq \DR\DDD(\iota_{+}\nbigo_Y)$ is identified with $\iota_{\ast}\Hom(\cnum,\cnum) \simeq \Cok(\nbigo\stackrel{t}{\lrarr}\nbigo)$ for which $\iota_{\ast}\id$ is mapped to the image of $-1$ in $\Cok(\nbigo\stackrel{t}{\lrarr}\nbigo)$. Then, the claim of Lemma follows. \hfill\qed \vspace{.1in} The isomorphism $\nu_Y:\DDD\nbigo_Y\lrarr\nbigo_Y$ is given by $\id\longmapsto 1$. We have the induced isomorphism $\iota_{+}\nu_Y: \DDD(\iota_{+}\nbigo_Y) \simeq \iota_{+}\nbigo_Y$. The following lemma is an immediate corollary of Lemma . \begin{lem} We have $\iota_{+}\nu_Y\bigl( \kappa'_1((dt)^{-1})\bigr) =-\kappa_1(1)$. \hfill\qed \end{lem} \subsubsection{A description of $\nu_X$} Let us describe the isomorphism $\nu_X:\DDD\nbigo_X\simeq\nbigo_X$. We use the following natural free resolution of $\nbigo_X$: \[ \nbigo_X\simeq \bigl( \nbigd_X\stackrel{\cdot\del_t}{\lrarr} \nbigd_X \bigr) \] It induces the following free resolution of $\DDD\nbigo_X$: \[ \DDD\nbigo_X\simeq \bigl( \nbigd_X\otimes\Omega_X^{-1} \stackrel{\del_t\cdot}{\lrarr} \nbigd_X\otimes\Omega_X^{-1} \bigr) \] We have the morphism of $\nbigd_X$-modules $\mu: \nbigd_X \lrarr \nbigd_X\otimes\Omega_X^{-1}$ given by $a_j\del_t^j \longmapsto (-1)^j\del_t^ja_j\otimes(dt)^{-1}$. We have the following commutative diagram: \begin{equation} \begin{CD} \nbigd_X\otimes\Omega_X^{-1} @>{\del_t\cdot}>> \nbigd_X\otimes\Omega_X^{-1} \\ @V{-\mu^{-1}}VV @V{\mu^{-1}}VV \\ \nbigd_X @>{\cdot\del_t}>> \nbigd_X \end{CD} \end{equation} It induces $\DDD\nbigo_X \lrarr \nbigo_X$. We can check that the induced morphism $\DR\nbigo_X\simeq \DR\DDD\nbigo_X$ is given by $\cnum\simeq \Hom(\cnum,\cnum)$, $1\longmapsto -\id$. Hence, it is $\nu_X$. \subsubsection{End of the proof of Proposition } We have the following commutative diagram: \begin{equation} \begin{CD} \nbigd_X @>{\cdot\del_t}>> \nbigd_X @>>> \nbigo_X \\ @V{\id}VV @V{\cdot t}VV @VVV \\ \nbigd_X @>{\cdot \del_tt}>> \nbigd_X @>{\varphi_1}>> \nbigo_X(\ast t) \\ @V{\cdot \del_t}VV @V{\id}VV @V{\rho_1}VV \\ \nbigd_X @>{\cdot t}>> \nbigd_X @>{\kappa_1}>> \iota_{+}\nbigo_Y \end{CD} \end{equation} Here $\varphi_1(1)=t^{-1}$. We also have the following commutative diagram: \begin{equation} \begin{CD} \nbigd_X @>{\cdot t}>> \nbigd_X @>{\kappa_1}>> \iota_{+}\nbigo_Y \\ @V{-\id}VV @V{\cdot(-\del_t)}VV @V{\rho_2}VV\\ \nbigd_X @>{\cdot t\del_t}>> \nbigd_X @>{\varphi_2}>> \nbigo_X(!t) \\ @V{\cdot t}VV @V{\id}VV @VVV \\ \nbigd_X @>{\cdot \del_t}>> \nbigd_X @>>> \nbigo_X \end{CD} \end{equation} Here $\varphi_2(1)=1$. As the dual of (), we obtain the following: \begin{equation} \begin{CD} \DDD(\iota_{+}\nbigo_Y) @<{\kappa_1'}<< \nbigd_X\otimes\Omega_X^{-1} @<{t\cdot }<< \nbigd_X\otimes\Omega_X^{-1} \\ @A{\DDD\rho_2}AA @A{-\id}AA @A{(-\del_t)\cdot}AA \\ \DDD\bigl(\nbigo_X(!t)\bigr) @<<< \nbigd_X\otimes\Omega_X^{-1} @<{t\del_t\cdot }<< \nbigd_X\otimes\Omega_X^{-1} \\ @AAA @A{t}AA @A{\id}AA \\ \DDD\nbigo_X @<<< \nbigd_X\otimes\Omega_X^{-1} @<{\del_t\cdot}<< \nbigd_X\otimes\Omega_X^{-1} \end{CD} \end{equation} We have the following commutative diagram: \[ \begin{CD} \nbigd_X\otimes\Omega_X^{-1} @>{(t\del_t)\cdot}>> \nbigd_X @>>> \DDD(\nbigo_X(!t)) \\ @V{-\mu^{-1}}VV @V{\mu^{-1}}VV @V{\nu_{X!}}VV \\ \nbigd_X @>{\cdot(\del_tt)}>> \nbigd_X @>{\varphi_1}>> \nbigo_X(\ast t) \end{CD} \] Because $\varphi_1(1)=t^{-1}$, we have $\DDD\rho_2\circ \nu_{X!}^{-1}(t^{-1}) =-\kappa'_1\bigl((dt)^{-1}\bigr)$. We also have $\rho_1(t^{-1})=\kappa_1(1)$. Then, we obtain $\DDD\rho_2\circ\nu_{X!}^{-1} =(\iota_{+}\nu_Y)^{-1}\circ\rho_1$ by Lemma . Thus, the proof of Proposition is finished. \hfill\qed \subsection{Duality and nearby cycle functors} We shall generalize Proposition . This subsection is a preliminary for \S. \subsubsection{Some induced isomorphisms} Let $t$ be a function on $X$ such that the zero divisor $(t)_0$ is smooth and reduced, i.e., the zero set $H$ of $t$ is smooth and reduced as a subscheme. Let $M$ be a regular singular meromorphic flat bundle on $(X,H)$. We use the natural identification $\bigl( \Pi^{a,b}_tM \bigr)^{\lor} \simeq \Pi_t^{-b+1,-a+1}M^{\lor}$. (See \cite[\S2.3]{Mochizuki-MTM} where the identification is explained for $\nbigr$-modules. The identification for flat bundles is obtained as the specialization at $\lambda=1$.) We have the isomorphisms $\nu_{\Pi^{a,b}_tM}:(\DDD_{X(\ast H)} \Pi_t^{a,b}M) \simeq \Pi^{-b+1,-a+1}_tM^{\lor}$, which induce \[ \nu_{\Pi^{a,b}_tM,\ast}: \DDD_X \bigl( \Pi_{t\ast}^{a,b}M \bigr) \simeq \Pi^{-b+1,-a+1}_{t!}M^{\lor}, \quad\quad \nu_{\Pi^{a,b}_tM,!}: \DDD_X\bigl(\Pi_{t!}^{a,b}M\bigr) \simeq \Pi^{-b+1,-a+1}_{t\ast}M^{\lor}. \] We have the following isomorphism for a sufficiently large $N$: \[ \psi^{(a)}_t(M)\simeq \Ker\Bigl( \Pi^{a-N,a}_{t!}(M) \lrarr \Pi^{a-N,a}_{t\ast}(M) \Bigr) \simeq \Cok\Bigl( \Pi^{a,a+N}_{t!}(M) \lrarr \Pi^{a,a+N}_{t\ast}(M) \Bigr), \] Hence, we have the following canonical isomorphisms: \[ \nu_{M,\psi_t^{(a)}}: \DDD_H\psi^{(a)}_t(M) \simeq \psi^{(-a+1)}_t(M^{\lor}) \] We have the $V$-filtration $U$ of $M$ along $t$. Here, we adopt the condition that $\del_tt+\alpha$ is nilpotent on $\psi_{t,\alpha}(M):=U_{\alpha}(M)\big/U_{<\alpha}(M)$. The natural perfect pairing $M\otimes M^{\lor}\lrarr \nbigo_X(\ast H)$ naturally induces a perfect pairing of $\nbigo_X$-modules $V_{-1}(M) \otimes V_{-1}(M^{\lor}) \lrarr \nbigo_X$. It induces a perfect pairing $\psi_{t,-1}(M)\times \psi_{t,-1}(M^{\lor}) \lrarr \nbigo_H$ of flat bundles. It gives the following natural identification: \begin{equation} \psi_{t,-1}(M)^{\lor} \simeq \psi_{t,-1}(M^{\lor}). \end{equation} We have the canonical isomorphisms $\xi^{(0)}_M:\psi_t^{(0)}(M) \simeq \iota_{+}\psi_{t,-1}(M)$ and $\xi^{(1)}_M: \psi_t^{(1)}(M)\simeq \iota_{+}\psi_{t,-1}(M)$. The construction of the isomorphisms is explained in the case of $\nbigr$-modules in \cite[S4.3]{Mochizuki-MTM}. The construction for $\nbigd$-modules is obtained as the specialization to $\lambda=1$. \subsubsection{Some commutative diagrams} \begin{prop} The following diagram is commutative: \begin{equation} \begin{CD} \DDD_X\psi_t^{(1)}(M) @>{\nu_{M,\psi_t^{(1)}}}>> \psi_t^{(0)}(M^{\lor}) \\ @A{\simeq}A{\DDD_X \xi_M^{(1)}}A @V{\simeq}V{\xi_{M^{\lor}}^{(0)}}V \\ \DDD_X\iota_{+} \psi_{t,-1}(M) @>{\simeq}>> \iota_{+} \psi_{t,-1}(M^{\lor}) \end{CD} \end{equation} Here, the lower horizontal arrow is induced as the composite of the natural isomorphisms $\DDD_X\iota_{+} \psi_{t,-1}(M) \simeq \iota_{+} \DDD_H\psi_{t,-1}(M) \simeq \iota_{+}\psi_{t,-1}(M)^{\lor} \simeq \iota_{+} \psi_{t,-1}(M^{\lor})$. The following diagram is also commutative. \begin{equation} \begin{CD} \DDD_X\psi_t^{(0)}(M) @>{-\nu_{M,\psi^{(0)}}}>> \psi_t^{(1)}(M^{\lor}) \\ @A{\simeq}A{\DDD_X\xi_M^{(0)}}A @V{\simeq}V{\xi_{M^{\lor}}^{(1)}}V \\ \DDD_X\iota_{+} \psi_{t,-1}(M) @>{\simeq}>> \iota_{+}\psi_{t,-1}(M^{\lor}) \end{CD} \end{equation} Here, the lower horizontal arrow is induced as the composite of the natural isomorphisms $ \DDD_X\iota_{+} \psi_{t,-1}(M) \simeq \iota_{+}\DDD_H\psi_{t,-1}(M) \simeq \iota_{+}\psi_{t,-1}(M)^{\lor} \simeq \iota_{+}\psi_{t,-1}(M^{\lor})$. \end{prop} \pf Let $V$ be any regular singular meromorphic flat bundle on $(X,H)$. We set $V_{\star}:=V(\star H)$. Let $K_V$ and $C_V$ denote the kernel and the cokernel of $V_!\lrarr V_{\ast}$. We have the following natural commutative diagram: \[ \begin{CD} 0@>>> \DDD_X C_V @>>> \DDD_X V_{\ast} @>>> \DDD_XV_! @>>> \DDD_X K_V @>>> 0\\ @. @V{\nu_{V,C}}VV @V{\nu_{V\ast}}VV @V{\nu_{V!}}VV @V{\nu_{V,K}}VV @.\\ 0@>>> K_{V^{\lor}} @>>> (V^{\lor})_! @>>> (V^{\lor})_{\ast} @>>> C_{V^{\lor}} @>>> 0 \end{CD} \] Let $K_{0V}$ and $C_{0V}$ denote the kernel and the cokernel of the morphism $\psi_{t,-1}(V) \stackrel{N}{\lrarr} \psi_{t,-1}(V)$. We have natural isomorphisms $\xi_{C,V}:C_V\simeq \iota_{+}C_{0V}$ and $\xi_{K,V}:K_V\simeq \iota_{+}K_{0V}$. The construction of the isomorphisms is explained in \cite[\S4.3.2]{Mochizuki-MTM} for $\nbigr$-modules. The construction for $\nbigd$-modules is obtained as the specialization at $\lambda=1$. Under the natural identification $\psi_{t,-1}(V)^{\lor} =\psi_{t,-1}(V^{\lor})$, we have $C_{0V}^{\lor} =K_{0V^{\lor}}$ and $K_{0V}^{\lor} =C_{0V^{\lor}}$. \begin{lem} The following diagram is commutative: \begin{equation} \begin{CD} \DDD_X K_V @>{\nu_{V,K}}>> C_{V^{\lor}} \\ @A{\simeq}A{\DDD_X\xi_{K,V}}A @V{\simeq}V{\xi_{C,V^{\lor}}}V \\ \DDD_X(\iota_{+}K_{0V}) @>>> \iota_{+}C_{0V^{\lor}} \end{CD} \end{equation} The lower horizontal arrow is the composite of $\DDD_X(\iota_{+}K_{0V}) \simeq \iota_{+}\DDD_H K_{0V} \simeq \iota_{+}K_{0V}^{\lor} \simeq \iota_{+}C_{0V^{\lor}}$. The following diagram is commutative: \begin{equation} \begin{CD} \DDD_X C_V @>{-\nu_{V,C}}>> K_{V^{\lor}} \\ @A{\simeq}A{\DDD_X\xi_{C,V}}A @V{\simeq}V{\xi_{K,V^{\lor}}}V \\ \DDD_X(\iota_{+}C_{0V}) @>>> \iota_{+}K_{0V^{\lor}} \end{CD} \end{equation} The lower horizontal arrow is the composite of $\DDD_X(\iota_{+}C_{0V}) \simeq \iota_{+}\DDD_H C_{0V} \simeq \iota_{+}C_{0V}^{\lor} \simeq \iota_{+}K_{0V^{\lor}}$. \end{lem} \pf It is enough to consider the case that the monodromy of $V$ along the loop around $t=0$ is unipotent. Moreover, we may assume that the logarithm of the monodromy is a Jordan block. Then, we have a flat subbundle $L\subset V$ of rank one such that $K_L\simeq K_V$ and $K_{0L}\simeq K_{0V}$. We also have $V^{\lor}\lrarr L^{\lor}$ which induces $C_{0V^{\lor}}\simeq C_{0L^{\lor}}$ and $C_{V^{\lor}}\simeq C_{L^{\lor}}$. Then, we obtain () from the claim for $L$, which follows from Proposition . We obtain () similarly. \hfill\qed \vspace{.1in} We have $K_{0,\Pi^{-N,1}M}\simeq \psi_{t,-1}(M)$ and $C_{0,\Pi^{0,N+1},M^{\lor}} \simeq \psi_{t,-1}(M^{\lor})$. Hence, we obtain the commutativity of () from Lemma . We obtain the commutativity of () similarly. \hfill\qed \subsubsection{Specialization of pairings} Let $M_i$ be regular singular meromorphic flat bundles on $(X,H)$. Let $P:M_1\otimes M_2\lrarr \nbigo_X(\ast H)$ be a morphism of $\nbigd$-modules. We have the naturally induced pairing $\psi_{t,-1}(P): \psi_{t,-1}(M_1)\otimes\psi_{t,-1}(M_2)\lrarr\nbigo_H$. The pairing $P$ and the isomorphism $\nu_{M_2}$ induce a morphism $\Psi_P:M_1\simeq \DDD_{X(\ast H)} M_2$. For $a=0,1$, we have the induced morphisms \[ \psi^{(a)}(M_1) \lrarr \psi^{(a)}(\DDD_X M_2) \simeq \DDD_X\psi^{(-a+1)}(M_2). \] By the natural isomorphisms $\iota_{+}\psi_{t,-1}(M_i) \simeq \psi^{(a)}(M_i)$ and the isomorphism $\nu_{\psi_{t,-1}(M_2)}$, we obtain \[ \iota_{+}\psi_{t,-1}(M_1) \simeq \DDD_X \iota_{+}\psi_{t,-1}(M_2) \simeq \iota_{+}\DDD_H\psi_{t,-1}(M_2) \simeq \iota_{+}\psi_{t,-1}(M_2)^{\lor}. \] It comes from a morphism $\rho_a:\psi_{t,-1}(M_1)\simeq\psi_{t,-1}(M_2)^{\lor}$, which is equivalent to a pairing \[ P_a:\psi_{t,-1}(M_1)\times\psi_{t,-1}(M_2) \lrarr\nbigo_H. \] \begin{cor} We have $P_0=\psi_{t,-1}(P)$ and $P_1=-\psi_{t,-1}(P)$. \end{cor} \pf Let $\mu:\psi_{t,-1}(M_1)\lrarr \psi_{t,-1}(M_2)^{\lor}$ be the morphism induced by $\psi_{t,-1}(P)$. The push-forward $\iota_{+}\mu: \iota_{+}\psi_{t,-1}(M_1) \lrarr \iota_{+}\psi_{t,-1}(M_2)^{\lor}$ is the composite of the following morphisms: \[ \iota_{+}\psi_{t,-1}(M_1) \simeq \psi^{(a)}(M_1) \lrarr \psi^{(a)}(M_2^{\lor}) \simeq \iota_{+}\psi_{t,-1}(M_2^{\lor}) \] By the construction, $\iota_{+}\rho_a$ is composite of the following morphisms: \begin{multline} \iota_{+}\psi_{t,-1}(M_1) \simeq \psi^{(a)}(M_1) \lrarr \psi^{(a)}(M_2^{\lor}) \simeq \psi^{(a)}(\DDD_X M_2) \simeq \DDD_X\psi^{(-a+1)}(M_2) \simeq \DDD_X\iota_{+}\psi_{t,-1}(M_2) \\ \simeq \iota_{+}\DDD_X\psi_{t,-1}(M_2) \simeq \iota_{+}\psi_{t,-1}(M_2)^{\lor} \end{multline} By Proposition , we have $\iota_{+}\mu =\iota_{+}\rho_0$ and $\iota_{+}\mu =-\iota_{+}\rho_1$. \hfill\qed \subsection{Push-forward} This subsection is a preliminary for \S. \subsubsection{Statement} Let $X$ be a complex manifold. Let $M_i$ be flat bundles on $X$. Let $P:M_1\otimes M_2\lrarr\nbigo_X$ be a morphism of $\nbigd$-modules. By the isomorphism $\nu_{M_2}:\DDD_X M_2\simeq M_2^{\lor}$, $P$ is equivalent to a morphism $\varphi:M_1\lrarr \DDD_X M_2$. Let $F:X\lrarr S$ be a smooth proper morphism of complex manifolds. We have the flat bundles $F^j_{+}M_i$ on $S$. We obtain the following morphism: \[ \begin{CD} F_{+}^jM_1 @>>> F_{+}^j\DDD_X M_2 @>>> \DDD_S F_{+}^{-j}M_2 \end{CD} \] The composite is denoted by $F^j_{+}\varphi$. By the isomorphism $\nu_{F_{+}^{-j}M_2}: \DDD_S F_{+}^{-j}M_2 \simeq \bigl( F_{+}^{-j}M_2 \bigr)^{\lor}$, we obtain a flat morphism \[ P(F_{+}^{j}\varphi): F_{+}^jM_1 \otimes F_{+}^{-j}M_2 \lrarr \nbigo_S. \] We give a more direct expression of the pairing $P(F_{+}^j\varphi)$. For each $Q\in S$, we set $X_Q:=F^{-1}(Q)$, and let $M_{i,Q}$ denote the restriction of the flat bundles $M_i$ to $X_Q$. Let $d:=\dim X-\dim S$. Let $H_{\DR}^{\bullet}(X_Q,M_{i,Q})$ denote the de Rham cohomology of the flat bundle $M_{i,Q}$. The fiber of $F_{+}^j(M_i)_{|Q}$ is naturally identified with $H^{d+j}_{\DR}\bigl(X_Q,M_{i,Q}\bigr)$. The pairing $P$ naturally induces the following pairing $P_{F,j,Q}$: \[ H^{d+j}_{\DR}\bigl(X_Q,M_{1,Q}\bigr) \times H^{d-j}_{\DR}\bigl(X_Q,M_{2,Q}\bigr) \lrarr H_{\DR}^{2d}(X_Q,\cnum) \lrarr \cnum \] Here, $H_{\DR}^{2d}(X_Q,\cnum)\lrarr\cnum$ is induced by $(2\pi\sqrt{-1})^{-d}\int_{X_Q}$. The family $\{P_{F,j,Q}\,|\,Q\in S\}$ gives a flat morphism $P_{F,j}: F_{+}^jM_1\otimes F_{+}^{-j}M_2\lrarr\nbigo_S$. We shall prove the following proposition in \S--. \begin{prop} We have $P(F_{+}^j\varphi) =\epsilon(d)(-1)^{dj}P_{F,j}$, where $\epsilon(d)=(-1)^{d(d-1)/2}$. \end{prop} \subsubsection{Preliminary} We have only to prove the claim locally around any point of $S$. So, we may assume that $S$ is a multi-disc. Let $t$ be a holomorphic function on $S$ such that the zero divisor $(t)_0$ is smooth and reduced. Let $S_0:=|(t)_0|$. The pull back $t\circ F$ is also denoted by $t$. Set $X_0:=\{Q\in X\,|\,t(Q)=0\}$. The induced morphism $X_0\lrarr S_0$ is denoted by $F_0$. We naturally identify the restriction of the flat bundles $M_i$ to $X_0$ with $\psi_{t,-1}(M_i)$. We also identify the restriction of $F_{+}^jM_i$ to $S_0$ with $\psi_{t,-1}F_{+}^j(M_i)$, which is also naturally isomorphic to $F_{0+}^j\psi_{t,-1}(M_i)$. We obtain the following pairings as the restriction of $P$, $P(F^j_{+}\varphi)$ and $P_{F,j}$: \[ \psi_{t,-1}P:\psi_{t,-1}(M_1)\otimes \psi_{t,,-1}(M_2) \lrarr \nbigo_{X_0} \] \[ \psi_{t,-1}P(F^j_{+}\varphi): \psi_{t,-1}F_{+}^j(M_1) \otimes \psi_{t,-1}F_{+}^{-j}(M_2) \lrarr \nbigo_{S_0} \] \[ \psi_{t,-1}P_{F,j}: \psi_{t,-1}F_{+}^j(M_1) \otimes \psi_{t,-1}F_{+}^{-j}(M_2) \lrarr \nbigo_{S_0} \] By the construction, we have $(\psi_{t,-1}P)_{F_0,j} =\psi_{t,-1}P_{F,j}$. Let $\varphi_0:\psi_{t,-1}M_1\lrarr \DDD_{X_0}\psi_{t,-1}M_2$ be the morphism corresponding to $\psi_{t,-1}P$. \begin{lem} We have $\psi_{t,-1}P(F^j_{+}\varphi) =P(F^j_{+}\varphi_0)$. As a consequence, to prove Proposition {\rm}, it is enough to consider the case where $S$ is a point. \end{lem} \pf Let $\iota_{X_0}:X_0\lrarr X$ be the inclusion. By Proposition , the morphism $\iota_{X_0+}\varphi_0: \iota_{X_0+}\psi_{t,-1}M_1 \lrarr \iota_{X_0+}\DDD_{X_0}\psi_{t,-1}M_2$ is identified with the composite of the following morphism: \begin{equation} \begin{CD} \psi^{(0)}_{t}(M_1) @>{\psi^{(0)}_t\varphi}>> \psi^{(0)}_t(\DDD_X M_2) \simeq \DDD_X \psi^{(1)}_t(M_2) \end{CD} \end{equation} We have the morphism $\rho_j: \psi_{t,-1}F_{+}^jM_1 \lrarr \DDD_{S_0}\psi_{t,-1}F_{+}^jM_2$ corresponding to $\psi_{t,-1}P(F_{+}^j\varphi)$. Let $\iota_{S_0}:S_0\lrarr S$ be the inclusion. By Proposition , $\iota_{S_0+}\rho_j$ is identified with \begin{equation} \begin{CD} \psi_{t}^{(0)}F_{+}^jM_1 @>{\psi_t^{(0)}F_{+}^j\varphi}>> \psi_{t}^{(0)}\DDD_S F_{+}^{-j}M_2 \simeq \DDD_S\psi_t^{(1)}F_{+}^{-j}M_2. \end{CD} \end{equation} Note that the following diagram of the natural isomorphisms is commutative, which can be checked easily: \[ \begin{CD} \DDD_S \psi_t^{(1)} F_{+}^{-j}M_2 @>{\simeq}>> \psi_t^{(0)}\DDD_S F_{+}^{-j}M_2 @>{\simeq}>> \psi_t^{(0)}F_{+}^{-j}\DDD_X M_2 \\ @V{\simeq}VV @. @V{\simeq}VV \\ \DDD_S F_{+}^{-j}\psi_{t}^{(1)}M_2 @>{\simeq}>> F_{+}^{-j} \DDD_X \psi_t^{(1)}M_2 @>{\simeq}>> F_{+}^{-j}\psi_t^{(0)} \DDD_X M_2 \end{CD} \] Then, the morphism () is identified with the push-forward of () by $F$. Hence, we have $\rho_j=F_{+}^j\varphi_0$, which implies the first claim of the lemma. The second claim follows from the first claim and the equality $(\psi_{t,-1}P)_{F_0,j} =\psi_{t,-1}P_{F,j}$. \hfill\qed \subsubsection{The case where $S$ is a point} Let $d:=\dim X$. Let $L_i$ be the sheaf of flat sections of $M_i$. Let $L_i^{\lor}$ denote the local system obtained as the dual of $L_i$. The pairing $P$ induces $\langle\cdot,\cdot\rangle: L_1\times L_2\lrarr \cnum_X$, which induces $\upsilon:L_1\lrarr L_2^{\lor}$. We have the natural isomorphism $\DR M_1\simeq L_1[d]$ and $\DR \DDD M_2\simeq \nrhom_{\nbigd_X}(M_2,\nbigo_X[d]) \simeq L_2^{\lor}[d]$. The morphism $\varphi$ induces $(-1)^d\upsilon: L_1[d]\lrarr L^{\lor}_2[d]$. Let us describe the induced morphism $L_1[d]\lrarr \nrhom_{\cnum_X}\bigl( \DR M_2, \DR \nbigo_X[d] \bigr)$. A section $e$ of $L_1$ naturally gives a section of $(L_1[d])^{-d}$, which is denoted by $e'$. We have the section $(-1)^d\upsilon(e')$ of $\nhom_{\nbigd_X}(M_2,\nbigo[d])^{-d}$. The image of $(-1)^d\upsilon(e')$ in $\nhom_{\cnum_X}(\DR M_2,\DR\nbigo[d])^{-d}$ is denoted by $F_e$. Then, we have \[ F_e(\eta^{d+j}m) =(-1)^{dj}\eta^{d+j}\cdot (-1)^d\upsilon(e')(m) =(-1)^{dj+d}\eta^{d+j}\langle e,m\rangle' \] Here, $\eta^{d+j}$ is a section of $(\Omega^{\bullet}[d])^j=\Omega^{d+j}$, $m$ is a section of $M_2$, and $\langle e,m\rangle'$ denotes the section of $(\nbigo[d])^{-d}$ corresponding to $\langle e,m\rangle$. We have the natural quasi isomorphisms $\DR\nbigo[d] \simeq \Tot\bigl( \Omega^{\bullet}[d]\otimes \distribution_X^{0,\bullet}[d] \bigr) \simeq \distribution_X^{\bullet}[2d]$, where the latter is given by $\eta^{d+j}\otimes \etabar^{d+i} \longmapsto \epsilon(d)(-1)^{dj} \eta^{d+j}\wedge\etabar^{d+i}$. Let $\Ftilde_e$ denote the image of $F_e$ in $\nhom_{\cnum_X}(\DR M_2,\distribution_X^{\bullet}[2d])^{-d}$. It is given by \[ \Ftilde_e(\eta^{d+j}m) =\epsilon(d)(-1)^d \eta^{d+j} \langle e,m\rangle \] which is naturally regarded as a section of $(\distribution^{\bullet}_X[2d])^{j-d}$. The morphism $L_1[d]\lrarr \nhom_{\cnum_X}(\DR M_2,\distribution_X^{\bullet}[2d])^{-d}$ is extended to the morphism $\DR M_1 \lrarr \nhom_{\cnum_X}(\DR M_2,\distribution_X^{\bullet}[2d])^{-d}$ induced by the multiplication $\DR M_1\times \DR M_2 \lrarr \distribution_X^{\bullet}[2d]$ given by $(\xi^{d+k}e,\eta^{d+j}m) \longmapsto \epsilon(d)(-1)^{kd} \xi^{d+k}\eta^{d+j} \langle e,m\rangle$. Then, we obtain the claim of Proposition in the case where $S$ is a point. By Lemma , the proof of Proposition is finished. \hfill\qed \subsection{Mixed TEP-structures and their functoriality} \subsubsection{Pairing of weight $w$ and graded pairing} Let $X$ be a complex manifold with a hypersurface $H$. Set $\nbigx:=\cnum_{\lambda}\times X$. For any $\nbigo_{\nbigx}$-module $N$, let $N(\ast H):= N\otimes \nbigo_{\nbigx}\bigl( \ast(\cnum_{\lambda}\times H) \bigr)$. An $\nbigr_X(\ast H)$-module $\nbigv$ is called smooth if (i) $\nbigv_{|\cnum_{\lambda}\times (X\setminus H)}$ is locally free as an $\nbigo_{\cnum_{\lambda}\times (X\setminus H)}$-module, (ii) $\nbigv$ is $\nbigo_{\nbigx}(\ast H)$-coherent and reflexive. For a smooth $\nbigr_X(\ast H)$-module $\nbigv$, a pairing of weight $w$ on $\nbigv$ is defined to be an $\nbigr_X$-homomorphism $P:\nbigv\otimes j^{\ast}\nbigv\lrarr \lambda^{-w}\nbigo_{\nbigx}(\ast H)$ such that (i) $j^{\ast}P=(-1)^{w}P\circ\exchange$, (ii) $P$ is non-degenerate, i.e., the induced morphism $\nbigv\lrarr \nhom_{\nbigo_{\nbigx}(\ast H)} \bigl(j^{\ast}\nbigv,\lambda^{-w}\nbigo_{\nbigx}(\ast H)\bigr)$ is an isomorphism. Such a pair $(\nbigv,P)$ is a ${\rm TP}(-w)$-structure in the sense of . Note that $P$ naturally induces $P(\ast\lambda):\nbigv(\ast\lambda)\otimes j^{\ast}\nbigv(\ast\lambda) \lrarr \nbigo_{\nbigx}(\ast H)(\ast\lambda)$, and $P$ is the restriction of $P(\ast\lambda)$ to $\nbigv\otimes j^{\ast}\nbigv$. We also obtain pairings $P^{(a)}$ of weight $w+2a$ on $\lambda^{-a}\nbigv$ as the restriction of $P(\ast\lambda)$. We have $P^{(a)}(\lambda^{-a}v_1,j^{\ast}(\lambda^{-a}v_2)) =(-1)^a\lambda^{-2a}P(v_1,j^{\ast}v_2)$. Suppose that a smooth $\nbigr_X(\ast H)$-module $\nbigv$ is equipped with an exhaustive increasing filtration $\Wtilde$ of $\nbigr_X(\ast H)$-submodules indexed by integers such that $\Gr^{\Wtilde}_w(\nbigv)$ are also smooth $\nbigr_X(\ast H)$-modules. A graded pairing of $(\nbigv,\Wtilde)$ is defined to be a tuple of pairings $P_w$ $(w\in\seisuu)$ of weight $w$ on $\Gr^{\Wtilde}_w(\nbigv)$. Such a filtered smooth $\nbigr_X(\ast H)$-module $(\nbigv,W)$ with a graded pairing $\{P_w\,|\,w\in\seisuu\}$ is called a mixed TP-structure. For any integer $a$, we set $\Wtilde_{k+2a}(\lambda^{-a}\nbigv) :=\lambda^{-a}\Wtilde_k\nbigv$. We naturally have $\Gr^{\Wtilde}_{k+2a}(\lambda^{-a}\nbigv) =\lambda^{-a}\Gr^{\Wtilde}_k(\nbigv)$. The pairing $P_k$ on $\Gr^{\Wtilde}_k(\nbigv)$ induces $P_{k+2a}^{(a)}$ on $\Gr^{\Wtilde}_{k+2a}(\lambda^{-a}\nbigv)$. Thus, we obtain another mixed TP-structure which consists of a filtered $\nbigr_X(\ast H)$-module $(\lambda^{-a}\nbigv,\Wtilde)$ and a graded pairing $(P^{(a)}_w\,|\,w\in\seisuu)$. A mixed TP-structure $(\nbigv,\Wtilde,P)$ is called integrable if (i) $\nbigv$ is an $\nbigrtilde_X(\ast H)$-module, (ii) $\Wtilde$ is a filtration by $\nbigrtilde_X(\ast H)$-submodules, (iii) $P_w$ are also $\nbigrtilde_X(\ast H)$-homomorphisms. In that case, $(\nbigv,\Wtilde,P)$ is also called a mixed TEP-structure. \subsubsection{Specialization of pairings} Suppose that $Y$ is a hypersurface of $X$ given as $\{t=0\}$ for a holomorphic function $t$ such that $dt$ is nowhere vanishing. We introduce a procedure for pairings in the context of $\nbigrtilde_X(\ast Y)$-modules, which we call specialization. The procedure can also work in the non-integrable case. We set $\nbigx:=\cnum_{\lambda}\times X$ and $\nbigy:=\cnum_{\lambda}\times Y$. Let $p_{\lambda}:\nbigx\lrarr X$ be the projection. Let $V\nbigrtilde_X\subset\nbigrtilde_X$ denote the sheaf of subalgebras generated by $\lambda p_{\lambda}^{\ast}\Theta_X(\log Y)$ and $\lambda^2\del_{\lambda}$ over $\nbigo_{\nbigx}$. Let $\nbigv$ be a smooth $\nbigrtilde_X(\ast Y)$-module. We assume the following: \begin{description} \item[(B0)] $\nbigv$ is regular along $t$, i.e., there exists a filtration $V_{\bullet}(\nbigv)=(V_a(\nbigv)\,|\,a\in\real)$ by $V\nbigrtilde_X$-submodules such that (i) $V_a(\nbigv)$ are locally free $\nbigo_{\nbigx}$-submodules and $\Gr^V_a(\nbigv)$ are locally free $\nbigo_{\nbigy}$-modules, (ii) $\lambda\del_tV_{a}\subset V_{a+1}$ and $tV_a(\nbigv)= V_{a-1}(\nbigv)$, (iii) $\lambda \del_tt+\lambda a$ are nilpotent on $\Gr^V_a(\nbigv)$. \end{description} We obtain $\nbigrtilde_Y$-modules $\Gr^V_a(\nbigv)$ which are smooth. We are particularly interested in $\Gr^V_{-1}(\nbigv)$. It is also denoted as $\psi_{t,-\vecdelta}(\nbigv)$. We also assume the following: \begin{description} \item[(B1)] The conjugacy classes of $(\lambda N)_{|(\lambda,W)}: \psi_{t,-\vecdelta}(\nbigv)_{(\lambda,Q)} \lrarr \psi_{t,-\vecdelta}(\nbigv)_{(\lambda,Q)}$ are independent of the choice of $(\lambda,Q)\in \nbigy$. Here, $N:\psi_{t,-\vecdelta}(\nbigv) \lrarr \psi_{t,-\vecdelta}(\nbigv)\lambda^{-1}$ denote the morphism induced by the multiplication of $t\del_t$. \end{description} Let $P$ be a pairing of weight $w$ on $\nbigv$. We have the induced morphism $P:V_{-1}\nbigv\otimes j^{\ast}V_{-1}\nbigv \lrarr \lambda^{-w}\nbigo_{\nbigx}$. Then, we have the naturally induced pairing of weight $w$: \[ \psi_{t,-\vecdelta}(P): \psi_{t,-\vecdelta}(\nbigv) \otimes j^{\ast} \psi_{t,-\vecdelta}(\nbigv) \lrarr \lambda^{-w}\nbigo_{\nbigy} \] Let $W(N)$ denote the monodromy weight filtration of $\lambda N$ on $\psi_{t,-\vecdelta}(\nbigv)$. By the assumption, $W(N)$ is a filtration by subbundles. We have the induced non-degenerate pairings: \[ \psi_{t,-\vecdelta}(P)_k: \Gr^{W(N)}_{-k}\psi_{t,-\vecdelta}(\nbigv) \times j^{\ast}\Gr^{W(N)}_{k}\psi_{t,-\vecdelta}(\nbigv) \lrarr \lambda^{-w}\nbigo_{\nbigy} \] For $k\geq 0$, let $P\Gr^W_k\psi_{t,-\vecdelta}(\nbigv)$ denote the primitive part, i.e., the kernel of $N^{k+1}: \Gr^W_k\psi_{t,-\vecdelta}(\nbigv) \lrarr \lambda^{-k-1}\Gr^W_{-k-2}\psi_{t,-\vecdelta}(\nbigv)$. We have the induced pairings: \[ \sp_t(P)_{w+k}:=\psi_{t,-\vecdelta}(P)_k\circ (N^k\times \id): P\Gr^{W(N)}_k\psi_{t,-\vecdelta}(\nbigv) \times j^{\ast}P\Gr^{W(N)}_{k}\psi_{t,-\vecdelta}(\nbigv) \lrarr \lambda^{-w-k}\nbigo_{\nbigy} \] By using $\psi_{t,-\vecdelta}(P)\circ(\id\times j^{\ast}N) =-\psi_{t,-\vecdelta}(P)\circ(N\times\id)$, we can easily check that $\sp_t(P)_{w+k}$ is a pairing of weight $w+k$ on $P\Gr^{W(N)}_k\psi_{t,-\vecdelta}(\nbigv)$. \vspace{.1in} Let $k\leq 0$. Let $P'\Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv)\subset \Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv)$ denote the image of $\lambda^{-k}P\Gr^W_{-k}\psi_{t,-\vecdelta}(\nbigv)$ by $N^{-k}$. We have the isomorphism \[ \id\times j^{\ast}N^{-k}: P'\Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv) \times j^{\ast} \bigl( \lambda^{-k}P\Gr^W_{-k}\psi_{t,-\vecdelta}(\nbigv) \bigr) \simeq P'\Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv) \times j^{\ast}P'\Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv). \] We obtain the induced pairing of weight $w+k$. \[ \sp_t(P)_{w+k}:= \psi_{t,-\vecdelta}(P)_{k} \circ(\id\times j^{\ast}N^{-k})^{-1}: P'\Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv) \times j^{\ast}P'\Gr^W_{k}\psi_{t,-\vecdelta}(\nbigv) \lrarr \lambda^{-k-w}\nbigo_{\nbigy} \] It is also induced by $(-1)^k\psi_{t,-\vecdelta}(P)_{-k}\circ (N^k\times\id)^{-1}$. \subsubsection{Filtration and graded pairings} We continue to use the notation in \S. Let $W(N)$ denote the filtration on $\Cok(\lambda N)$ induced by the filtration $W(N)$ on $\psi_{t,-\vecdelta}(\nbigv)$. It is equipped with the flat connection. We set \[ \Wtilde_{k}\Cok(\lambda N):= W(N)_{k-w}\Cok(\lambda N). \] In other words, $\Wtilde_k\Cok(\lambda N)$ is the image of $W(N)_{k-w}\psi_{t,-\vecdelta}(\nbigv) \lrarr \Cok(\lambda N)$. We have \[ \Gr^{\Wtilde}_k\Cok(\lambda N)\simeq \left\{ \begin{array}{ll} P\Gr^{W(N)}_{k-w}\psi_{t,-\vecdelta}(\nbigv) & (k\geq w)\\ 0 & (k<w) \end{array} \right. \] Hence, we have the pairing of weight $k$ on $\Gr^{\Wtilde}_k\Cok(\lambda N)$ induced by $\sp_t(P)_{k}$. The induced pairing is also denoted by $\sp_t(P)_{k}$. The tuple $\bigl(\sp_t(P)_k\,\big|\,k\in\seisuu \bigr)$ is denoted by $\sp_t(P)$. Thus, we obtain a mixed TEP-structure $(\Cok(\lambda N),\Wtilde,\sp_t(P))$. Similarly, we set $\Wtilde_k\Ker(\lambda N):= W(N)_{k-w}\cap \Ker(\lambda N)$. We have \[ \Gr^{\Wtilde}_k\Ker(\lambda N) \simeq \left\{ \begin{array}{ll} P'\Gr^{W(N)}_{k-w}\psi_{t,-\vecdelta}(\nbigv) & (k\leq w)\\ 0 & (k>w) \end{array} \right. \] Hence, we have the pairing $\sp_t(P)_k$ of weight $k$ on $\Gr^{\Wtilde}_k\Ker(\lambda N)$. The induced pairing is also denoted by $\sp_t(P)_k$. Thus, we obtain a mixed TEP-structure $(\Ker(\lambda N),\Wtilde,\sp_t(P))$, where $\sp_t(P):=\bigl(\sp_t(P)_k\,\big|\,k\in\seisuu\bigr)$. \subsubsection{Dependence on defining functions} Let $\tau$ be a holomorphic function on $X$ such that (i) $Y=\{\tau=0\}$, (ii) $d\tau$ is nowhere vanishing. The $V$-filtrations for $t$ and $\tau$ are the same, and we naturally have $\psi_{\tau,-\vecdelta}(\nbigv) =\psi_{t,-\vecdelta}(\nbigv)$ as $\nbigo_{\nbigy}$-modules. Let $N_t$ and $N_{\tau}$ denote the morphisms $\psi_{\tau,-\vecdelta}(\nbigv) \lrarr \lambda^{-1}\psi_{\tau,-\vecdelta}(\nbigv)$ induced by $t\del_t$ and $\tau\del_{\tau}$. We remark the following standard lemma. \begin{lem} We have $N_t=N_{\tau}$ and $\psi_{t,-\vecdelta}(P)=\psi_{\tau,-\vecdelta}(P)$. As a result, we have $\Cok(N_t)=\Cok(N_{\tau})$ and $\Ker(N_t)=\Ker(N_{\tau})$. On $\Cok(N_t)=\Cok(N_{\tau})$ and $\Ker(N_t)=\Ker(N_{\tau})$, the induced connections are also equal. Moreover, the induced filtrations and the graded pairings are also equal. \hfill\qed \end{lem} In other words, the induced mixed TEP-structures are well defined in the sense that they are independent of the choice of a defining function of $Y$. Note that the induced connections are not the same on $\psi_{t,-\vecdelta}(\nbigv)=\psi_{\tau,-\vecdelta}(\nbigv)$, in general. \subsection{Pairings associated with real structure and graded sesqui-linear duality} \subsubsection{Compatibility of sesqui-linear duality and real structure} Let $X$ be a complex manifold. Let $\nbigt=(\nbigm',\nbigm'',C)$ be a pure twistor $\nbigd$-module of weight $w$ on $X$. Let $\nbigs:\nbigt\lrarr \nbigt^{\ast}(-w)$ be a sesqui-linear duality. Let $\kappa:\gammatilde^{\ast}\nbigt\lrarr \nbigt$ be a real structure. We say that $\nbigs$ and $\kappa$ are compatible if the following diagram is commutative. \[ \begin{CD} \gammatilde^{\ast}\nbigt @>>> \gammatilde^{\ast}(\nbigt^{\ast})(-w) \\ @V{\kappa}VV @A{\kappa^{\ast}}AA \\ \nbigt @>{\nbigs}>> \nbigt^{\ast}(-w) \end{CD} \] Recall that $\nbigs$ is expressed as a pair of morphisms $\nbigm'\stackrel{\nbigs'}{\llarr}\nbigm''\lambda^w$ and $\nbigm''\stackrel{\nbigs''}{\lrarr}\nbigm'\lambda^{-w}$ satisfying $\nbigs'=\nbigs''$, and that $\kappa$ is expressed as a pair of morphisms $\DDD_X j^{\ast}\nbigm''\stackrel{\kappa'}{\llarr}\nbigm'$ and $\DDD_X j^{\ast}\nbigm'\stackrel{\kappa''}{\lrarr}\nbigm''$ satisfying $j^{\ast}\DDD_X\kappa''\circ\kappa'=\id$ and $\kappa''\circ j^{\ast}\DDD_X\kappa'=\id$. Here, $\DDD_X$ denotes the duality functor for $\nbigr_X$-modules. Note that $\gammatilde^{\ast}\bigl(\nbigt(-w)\bigr) \simeq \gammatilde^{\ast}(\nbigt)(-w)$ is given by $\bigl((-1)^w,(-1)^w\bigr)$. Then, the commutativity of the above diagram means the commutativity of the following: \[ \begin{CD} j^{\ast}\DDD_X(\nbigm') @>{(-1)^wj^{\ast}\DDD_X\nbigs''}>> j^{\ast}\DDD_X(\nbigm'')\lambda^{-w} \\ @V{\kappa''}VV @A{\kappa'}AA \\ \nbigm'' @>{\nbigs''}>> \nbigm'\lambda^{-w} \end{CD} \] Because $\kappa''=(j^{\ast}\DDD_X\kappa')^{-1}$, the commutativity means $\kappa'\circ\nbigs'' =(-1)^wj^{\ast}\DDD_X(\kappa'\circ\nbigs'')$. \vspace{.1in} Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module with a real structure $\kappa$ and a graded sesqui-linear duality $(\nbigs_w\,|\,w\in\seisuu)$. We say that the real structure and the graded sesqui-linear duality are compatible if the induced real structure and the sesqui-linear duality on $\Gr^W_w(\nbigt)$ are compatible for any $w$. \subsubsection{The associated TEP-structure in the pure case} Let $\nbigt=(\nbigm',\nbigm'',C)$ be an integrable pure twistor $\nbigd$-module of weight $w$ on a complex manifold $X$ with a real structure $\kappa$ and a polarization $\nbigs$ which are compatible and integrable. We have the following isomorphisms of $\nbigrtilde_X$-modules: \[ \begin{CD} \nbigm'' @>{\nbigs''}>> \nbigm'\lambda^{-w} @>{\kappa'}>> \lambda^{-w}j^{\ast}\DDD_X\nbigm'' \end{CD} \] Let $H$ be a hypersurface of $X$. Suppose that $\nbigm'(\ast H)$ and $\nbigm''(\ast H)$ are smooth $\nbigr_X(\ast H)$-modules. We have a natural isomorphism \[ \nu_{\nbigm''}: (\DDD_X\nbigm'')(\ast H) \simeq \lambda^{d_X}\cdot \nbigm''(\ast H)^{\lor} \] whose restriction to $\{\lambda_0\}\times X$ $(\lambda_0\neq 0)$ is given by the morphism in \S. We consider the pairing $P$ induced by $(-1)^{d_X}\nu_{\nbigm''}$ and $\kappa'\circ\nbigs''$: \[ \begin{CD} P: \nbigm''(\ast H)\times j^{\ast}\nbigm''(\ast H) @>>> \lambda^{d_X-w}\nbigo_{\nbigx}(\ast H) \end{CD} \] \begin{prop} We have $j^{\ast}P\circ\exchange =(-1)^{d_X-w}P$. In other words, $P$ is a pairing of weight $w-d_X$ on $\nbigm''(\ast H)$. In other words, $(\nbigm''(\ast H),P)$ is a $TEP(d_X-w)$-structure. \end{prop} \pf The claim of this proposition follows from the relation $\kappa'\circ\nbigs'' =(-1)^{w}j^{\ast}\DDD(\kappa'\circ\nbigs'')$ and Lemma . \hfill\qed \vspace{.1in} For any $n\in\seisuu$, the pure twistor $\nbigd$-module $\nbigt\otimes\newTate(-n)$ of weight $w+2n$ naturally equipped with the polarization and the real structure which are compatible. The polarization is given by $((-1)^n\nbigs',(-1)^n\nbigs'')$. The real structure is given by $((-1)^n\kappa',(-1)^n\kappa'')$. Hence, the induced pairing is $P^{(n)}$ which was introduced in \S. \begin{example} Let $\nbigt(F)=\bigl( \lambda^{d_X}\nbigl(F),\nbigl(F),C \bigr)$ be the integrable pure twistor $\nbigd$-module of weight $d_X$ associated to a holomorphic function $F$ on $X$, where $d_X:=\dim X$. The polarization $\nbigs$ is given by $\bigl((-1)^{d_X},(-1)^{d_X}\bigr)$. The real structure $\gammatilde^{\ast}\nbigt(F) \simeq\nbigt(F)$ is given by $\bigl(\nu^{-1},(-1)^{d_X}\nu\bigr)$. The induced pairing $\nbigl(F)\times j^{\ast}\nbigl(F) \lrarr \nbigo_{\nbigx}$ is just the multiplication. \hfill\qed \end{example} \subsubsection{The associated mixed TEP-structure} Let $(\nbigt,W)$ be a mixed twistor $\nbigd$-module on $X$. Let $(\nbigm',\nbigm'',C)$ be the underlying $\nbigr_X$-triple of $\nbigt$. We often pick $\nbigm''$, and we say that $\nbigm''$ is the underlying $\nbigr_X$-module of $\nbigt$. We have the filtration $W$ on $\nbigm''$ such that $W_j\nbigm''$ are the underlying $\nbigr$-modules of $W_j\nbigt$, which we call the filtration of $\nbigm''$ underlying the weight filtration of $(\nbigt,W)$. In the following, we shall often use the shifted filtration $\Wtilde$ on $\nbigm''$ given by $\Wtilde_{k}(\nbigm''):=W_{k+\dim X}\nbigm''$. \vspace{.1in} Let $(\nbigt,W)$ be an integrable mixed twistor $\nbigd$-module with real structure and a graded sesqui-linear duality which are compatible and integrable. Let $H$ be a hypersurface of $X$. Let $\nbigm''$ be the underlying $\nbigrtilde_X$-module. Suppose that $\nbigm''(\ast H)$ is a smooth $\nbigr_X(\ast H)$-module. Let $W$ be the filtration of $\nbigm''$ underlying the weight filtration of $\nbigt$. We set $\Wtilde_k\bigl(\nbigm''(\ast H)\bigr):= W_{k+d_S}\nbigm''(\ast H)$. Each $\Gr^{\Wtilde}_k(\nbigm''(\ast H))$ is equipped with the pairing $P_k$ of weight $k$ induced by the real structure and the sesqui-linear duality. In this way, we obtain a mixed TEP-structure $(\nbigm''(\ast H),\Wtilde,\{P_k\})$. \begin{rem} The underlying $\nbigr$-module of $\nbigt\otimes\newTate(-n)$ is $\lambda^{-n}\nbigm''(\ast H)$. The filtration $\Wtilde(\lambda^{-n}\nbigm''(\ast H))$ induced by the mixed twistor structure $(\nbigt,W)\otimes\newTate(-n)$ is given by $\Wtilde_k(\lambda^{-n}\nbigm''(\ast H)) =\lambda^{-n}\Wtilde_{k-2n}\nbigm''(\ast H)$, which is equal to the shift of the filtration in {\rm\S}. The graded pairing on $(\lambda^{-n}\nbigm''(\ast H),\Wtilde)$ induced by the real structure and the graded sesqui-linear duality on $(\nbigt,W)\otimes\newTate(-n)$ is equal to that induced by the graded pairing on $(\nbigm''(\ast H),\Wtilde)$ by the procedure in {\rm\S}. \hfill\qed \end{rem} \subsection{Comparison of the specializations} \subsubsection{Statement} Let $X$ be a complex manifold with a smooth hypersurface $Y$. Suppose that $Y$ is given as $\{t=0\}$ for a holomorphic function $t$ such that $dt$ is nowhere vanishing. Let $\nbigt=(\nbigm_1,\nbigm_2,C)$ be an integrable pure twistor $\nbigd$-module of weight $w$ on $X$ with a sesqui-linear duality $\nbigs$ and a real structure $\kappa$ which are compatible and integrable. \begin{assumption} We assume that $\nbigm_i(\ast Y)$ are smooth $\nbigr_X(\ast Y)$-modules, and regular along $Y$. (See the condition {\bf (B0)} in {\rm\S} for the regularity.) \hfill\qed \end{assumption} Set $d:=\dim X$. We have the pairing $P$ of weight $w-d$ on $\nbigm_2(\ast Y)$. We have the endomorphism $\lambda N: \psi_{t,-\vecdelta}(\nbigm_2) \lrarr \psi_{t,-\vecdelta}(\nbigm_2)$, where $N$ is the induced by $t\del_t$. Note that the condition {\bf (B1)} is satisfied because $\nbigm_2$ comes from a pure twistor $\nbigd$-module. By the procedure in \S, we obtain filtrations $\Wtilde^{[1]}$ and graded pairings $(P^{[1]}_k\,|\,k\in\seisuu):=\sp_t(P)$ on $\Cok(\lambda N)$ and $\Ker(\lambda N)$. \vspace{.1in} Let $\nbign: \psitilde_{t,-\vecdelta}(\nbigt)\otimes\nbigu(-1,0) \lrarr \psitilde_{t,-\vecdelta}(\nbigt)\otimes\nbigu(0,-1)$ be given by $(-N,-N)$, where $N: \psi_{t,-\vecdelta}(\nbigm_i) \lrarr \lambda^{-1}\psi_{t,-\vecdelta}(\nbigm_i)$ are induced by $t\del_t$. According to the isomorphism in Proposition 4.3.1 of , we have the following commutative diagram: \[ \begin{CD} \psi^{(1)}_t(\nbigt) @>>> \psi^{(0)}_t(\nbigt) \\ @V{\simeq}VV @V{\simeq}VV \\ \iota_{\dagger}\psitilde_{t,-\vecdelta}(\nbigt) \otimes\nbigu(-1,0) @>{\nbign}>> \iota_{\dagger}\psitilde_{t,-\vecdelta}(\nbigt) \otimes\nbigu(0,-1) \end{CD} \] Here, the upper horizontal arrow is the canonical morphism. Hence, we have the following natural isomorphisms of mixed twistor $\nbigd$-modules: \[ \iota_{\dagger}\Cok(\nbign) \simeq \Cok\bigl( \psi^{(1)}_t(\nbigt) \lrarr \psi^{(0)}_t(\nbigt) \bigr) \simeq \Cok\Bigl( \nbigt[!t] \lrarr \nbigt[\ast t] \Bigr) \] \[ \iota_{\dagger}\Ker(\nbign) \simeq \Ker\bigl( \psi^{(1)}_t(\nbigt) \lrarr \psi^{(0)}_t(\nbigt) \bigr) \simeq \Ker\Bigl( \nbigt[!t] \lrarr \nbigt[\ast t] \Bigr) \] So, $\Cok(\nbign)\otimes\newTate(1)$ and $\Ker(\nbign)$ are equipped with the real structure and the graded sesqui-linear duality which are compatible and integrable. The underlying $\nbigrtilde_Y$-modules of $\Cok(\nbign)\otimes\newTate(1)$ and $\Ker(\nbign)$ are $\Cok(\lambda N)$ and $\Ker(\lambda N)$, respectively. Applying the procedure in \S, we obtain filtrations $\Wtilde^{[2]}$ and graded pairings $(P^{[2]}_k\,|\,k\in\seisuu)$ on $\Cok(\lambda N)$ and $\Ker(\lambda N)$. \begin{prop} We have $\Wtilde^{[1]}=\Wtilde^{[2]}$ and $P^{[1]}_k=P^{[2]}_k$ for any $k\in\seisuu$. In other words, the induced mixed TEP-structures are the same. \end{prop} For the monodromy weight filtration $W(\nbign)$ of $\nbign$ on $\psitilde_{t,-\vecdelta}(\nbigt)$, we have $W(\nbign)_{j}\psitilde_{t,-\vecdelta}(\nbigt) =W_{j+w}\psitilde_{t,-\vecdelta}(\nbigt)$. Then, we can check the equality for the filtrations easily. We shall compare the pairings in \S--. \subsubsection{Preliminary} Let $\nbigv$ be a smooth $\nbigrtilde_X(\ast Y)$-module which is regular along $t=0$. We have a natural isomorphism $\nu_{\nbigv}:\DDD(\nbigv)(\ast Y) \simeq \lambda^d\cdot\nbigv^{\lor}$ whose specialization at $\{\lambda_0\}\times X$ $(\lambda_0\neq 0)$ is equal to the morphism in \S. By a procedure similar to that in \S, it induces the following isomorphisms of $\nbigrtilde_X$-modules for $a=0,1$: \[ \nu_{\nbigv,\psi^{(a)}_t}: \DDD_X \psi_t^{(a)}(\nbigv) \simeq \psi_t^{(-a+1)}(\nbigv^{\lor}) \lambda^d \] \begin{lem} The following diagram is commutative: \begin{equation} \begin{CD} \DDD_X\psi_t^{(1)}(\nbigv) @>{\nu_{\nbigv,\psi_t^{(1)}}}>{\simeq}> \psi_t^{(0)}(\nbigv^{\lor})\lambda^d \\ @V{\simeq}VV @V{\simeq}VV \\ \DDD_X\iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv) @>{\simeq}>> \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv^{\lor}) \lambda^{d-1} \end{CD} \end{equation} Here, the vertical arrows are the isomorphisms in {\rm\S4.3} of {\rm}, and the lower horizontal arrow is induced as the composite of the isomorphisms $\DDD_X(\iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv)) \simeq \iota_{\dagger}\DDD_Y \psi_{t,-\vecdelta}(\nbigv) \simeq \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv)^{\lor}\lambda^{d-1} \simeq \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv^{\lor}) \lambda^{d-1}$. The following diagram is commutative: \begin{equation} \begin{CD} \DDD_X\bigl( \psi^{(0)}_t(\nbigv) \bigr) @>{-\nu_{\nbigv,\psi^{(0)}_t}}>{\simeq}> \psi_t^{(1)}(\nbigv^{\lor})\lambda^d \\ @V{\simeq}VV @V{\simeq}VV \\ \DDD_X(\iota_{\dagger} \psi_{t,-\vecdelta}(\nbigv) \lambda^{-1}) @>{\simeq}>> \iota_{\dagger} \psi_{t,-\vecdelta}(\nbigv^{\lor}) \lambda^d \end{CD} \end{equation} The vertical arrows and the lower horizontal arrows are given as in the diagram {\rm ()}. \end{lem} \pf We have only to prove the commutativity of the diagrams after taking the specializations to $\{\lambda_0 \}\times X$ for any generic $\lambda_0\neq 0$, which follows from Proposition . \hfill\qed \vspace{.1in} Let $\nbigv_i$ $(i=1,2)$ be smooth $\nbigrtilde_X(\ast Y)$-modules which are regular along $t=0$. Let $P':\nbigv_1\otimes j^{\ast}\nbigv_2\lrarr \lambda^{-m}\nbigo_{\nbigx}(\ast Y)$ be a morphism of $\nbigrtilde_X(\ast Y)$-modules. We have the induced morphism of $\nbigrtilde_Y$-modules \[ \psi_{t,-\vecdelta}(P'): \psi_{t,-\vecdelta}(\nbigv_1) \otimes j^{\ast} \psi_{t,-\vecdelta}(\nbigv_2) \lrarr \lambda^{-m} \nbigo_{\nbigy}. \] The pairing $P'$ and the isomorphism $(-1)^{d}\nu_{\nbigv_2}$ induce an $\nbigrtilde_X$-homomorphism $\nbigv_1\lrarr \lambda^{-m-d}(j^{\ast}\DDD_X\nbigv_2)(\ast Y)$. It induces the following morphisms for $a=0,1$: \[ \psi^{(a)}_t(\nbigv_1) \lrarr \lambda^{-m-d} j^{\ast}\DDD_X\psi^{(1-a)}_t(\nbigv_2) \] By the isomorphisms $\lambda^{-1+a}\cdot \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv_1) \simeq \psi^{(a)}_t(\nbigv_1)$ and $\lambda^{-a}\iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv_2) \simeq \psi^{(1-a)}_t(\nbigv_2)$, we obtain the following: \begin{equation} \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv_1) \lrarr \lambda^{-m-d+1} j^{\ast}\DDD_X\bigl( \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigv_2) \bigr) \simeq \lambda^{-m-d+1} \iota_{\dagger}j^{\ast}\DDD_Y\psi_{t,-\vecdelta}(\nbigv_2) \end{equation} The morphism () and $(-1)^{d-1}\nu_{\psi_{t,-\vecdelta}(\nbigv_2)}$ induce a pairing $P'_a:\psi_{t,-\vecdelta}(\nbigv_1) \times j^{\ast}\psi_{t,-\vecdelta}(\nbigv_2) \lrarr \lambda^{-m}\nbigo_{\nbigy}$. We obtain the following lemma from Lemma , as in the case of Corollary . We remark the twist of the signatures, i.e., we use $(-1)^{d}\nu_{\nbigv}$ and $(-1)^{d-1}\nu_{\psi_{t,-\vecdelta}(\nbigv_2)}$. \begin{lem} We have $\psi_{t,-\vecdelta}(P')=P'_1$ and $\psi_{t,-\vecdelta}(P')=-P'_0$. \hfill\qed \end{lem} \subsubsection{Proof of Proposition } For $a=0,1$, from the morphisms $\psi^{(a)}_t(\nbigs): \psi_t^{(a)}(\nbigt) \lrarr \psi_t^{(a)}(\nbigt)^{\ast}\otimes\newTate(-w-1+2a)$ and $\psi_t^{(a)}(\gammatilde^{\ast}): \gammatilde^{\ast}\psi_t^{(a)}(\nbigt) \simeq \psi_t^{(a)}(\nbigt)$, we have the following $\nbigrtilde_X$-homomorphisms: \begin{equation} \psi^{(a)}_t(\nbigm_2) \stackrel{b^{(a)}_1}{\lrarr} \lambda^{-w-1+2a}\psi^{(1-a)}_t(\nbigm_1) \stackrel{b^{(a)}_2}{\lrarr} \lambda^{-w-1+2a} j^{\ast}\DDD_X\psi^{(a)}_t(\nbigm_2) \end{equation} The composite $b_2^{(a)}\circ b_1^{(a)}$ comes from an isomorphism $\psi_{t,-\vecdelta}(\nbigm_2)\lambda^{-1} \simeq \lambda^{-w-1} j^{\ast}\DDD_X \bigl(\psi_{t,-\vecdelta}(\nbigm_2) \lambda^{-1}\bigr)$, in the case $a=0$, or $\psi_{t,-\vecdelta}(\nbigm_2) \simeq \lambda^{-w+1} j^{\ast}\DDD_X \bigl(\psi_{t,-\vecdelta}(\nbigm_2)\bigr)$ in the case $a=1$. Together with $(-1)^{d-1}\nu_{\psi_{t,-\vecdelta}(\nbigm_2)}$, we obtain a pairing $\Ptilde_a: \psi_{t,-\vecdelta}(\nbigm_2) \times j^{\ast}\psi_{t,-\vecdelta}(\nbigm_2) \lrarr \lambda^{d-w}\nbigo_{\nbigy}$. \begin{lem} We have $\psi_{t,-\vecdelta}(P)=\Ptilde_0=\Ptilde_1$. \end{lem} \pf According to \S4.3 of , we have the following commutative diagram: \begin{equation} \begin{CD} \psi^{(0)}_t(\nbigt) @>{\psi_t^{(0)}(\nbigs)}>> \psi^{(0)}_t(\nbigt)^{\ast} \otimes\newTate(-w-1) \\ @V{\simeq}VV @V{\simeq}VV \\ \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigt) \otimes\nbigu(0,-1) @>>> \Bigl( \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigt) \otimes\nbigu(0,-1) \Bigr)^{\ast} \otimes\newTate(-w-1) \end{CD} \end{equation} The lower horizontal arrow is induced by $\psi_{t,-\vecdelta}(\nbigs)$. It means that the following diagram is commutative: \begin{equation} \begin{CD} \psi^{(0)}_t(\nbigm_2) @>{b_1^{(0)}}>> \psi^{(1)}_t(\nbigm_1) \lambda^{-w-1} \\ @V{\simeq}VV @V{\simeq}VV \\ \iota_{\dagger} \lambda^{-1}\psi_{t,-\vecdelta}(\nbigm_2) @>{\psi_{t,-\vecdelta}(\nbigs'')}>> \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigm_1) \lambda^{-w-1} \end{CD} \end{equation} We have the pairing $P':\nbigm_1(\ast Y) \times j^{\ast}\nbigm_2(\ast Y) \lrarr \lambda^{d} \nbigo_{\nbigx}(\ast Y)$ induced by $\kappa':\nbigm_1\lrarr j^{\ast}\DDD\nbigm_2$ and $(-1)^{d}\nu_{\nbigm_2(\ast Y)}$. The morphism $b^{(0)}_2$ comes from $\psi_{t,-\vecdelta}(\nbigm_1) \lrarr j^{\ast}\DDD(\psi_{t,-\vecdelta}(\nbigm_2)\lambda^{-1})$. Together with the morphism $(-1)^{d-1}\nu_{\psi_{t,-\vecdelta}(\nbigm_2)}$, we obtain a pairing $P'_1: \psi_{t,-\vecdelta}(\nbigm_1) \times j^{\ast}\psi_{t,-\vecdelta}(\nbigm_2) \lrarr \lambda^{d}\nbigo_{\nbigy}$. We obtain $\psi_{t,-\vecdelta}(P)=\Ptilde_0$ from the equality $\psi_{t,-\vecdelta}(P')=P'_1$ in Lemma , and the commutativity (). We also have the following commutative diagram from (): \begin{equation} \begin{CD} \psi^{(1)}_t(\nbigm_2) @>{b_1^{(1)}}>> \psi^{(0)}_t(\nbigm_1) \lambda^{-w+1} \\ @V{\simeq}VV @V{\simeq}VV \\ \iota_{\dagger} \psi_{t,-\vecdelta}(\nbigm_2) @>{-\psi_{t,-\vecdelta}(\nbigs'')}>> \iota_{\dagger}\psi_{t,-\vecdelta}(\nbigm_1) \lambda^{-w} \end{CD} \end{equation} The morphism $b_2^{(1)}$ comes from a morphism $\lambda^{-1}\psi_{t,-\vecdelta}(\nbigm_1) \lrarr j^{\ast}\DDD\bigl( \psi_{t,-\vecdelta}(\nbigm_2) \bigr)$. It induces $P_0': \psi_{t,-\vecdelta}(\nbigm_1) \times j^{\ast}\psi_{t,-\vecdelta}(\nbigm_2) \lrarr \lambda^d\nbigo_{\nbigy}$. From the equality $\psi_{t,-\vecdelta}(P')=-P_0'$ and the commutativity of the diagram (), we obtain $\psi_{t,-\vecdelta}(P)=\Ptilde_1$. Thus, the proof of Lemma is finished. \hfill\qed \vspace{.1in} The graded sesqui-linear duality on $\Gr^W_{w+1+k}\iota_{\dagger}\Cok(\nbign)$ is induced by $\psi_t^{(0)}(\nbigs)\circ(-\nbign)^k$. Hence, by Lemma , we obtain $P^{[1]}=P^{[2]}$ on $\Cok(\nbign)$. The graded sesqui-linear duality on $\Gr^{W}_{w-1-k}\iota_{\dagger}\Ker(\nbign)$ is induced by $\psi_t^{(1)}(\nbigs)\circ \nbign^k$. Hence, by Lemma , we obtain $P^{[1]}=P^{[2]}$ on $\Ker(\nbign)$. Thus the proof of Proposition is finished. \hfill\qed \subsection{Push-forward} Let $X$ be a complex manifold. We set $d_X:=\dim X$. Let $\nbigt=(\nbigm_1,\nbigm_2,C)$ be an integrable pure twistor $\nbigd$-module of weight $w$ on $X$ with a sesqui-linear duality $\nbigs$ and a real structure $\kappa$ which are compatible and integrable. Suppose that $\nbigt$ is smooth, i.e., $\nbigm_i$ are locally free $\nbigo_{\nbigx}$-modules. We have the associated pairing $P(\nbigs,\kappa)$ on $\nbigm_2$ of weight $w-d_X$. Let $F:X\lrarr Y$ be a smooth projective morphism. We set $d_Y:=\dim Y$. We have the pure twistor $\nbigd$-modules $F_{\dagger}^i(\nbigt) =(F_{\dagger}^{-i}\nbigm_1,F_{\dagger}^i\nbigm_2,F_{\dagger}^iC)$. They are equipped with the induced real structure $\kappa_i$. We also have the induced morphisms $\nbigs_i: F_{\dagger}^i\nbigt \lrarr \bigl(F_{\dagger}^{-i}\nbigt\bigr)^{\ast} \otimes\newTate(-w)$. Note that $F_{\dagger}^j\nbigm_i$ are locally free $\nbigo_{\nbigy}$-modules. As in \S, we have the associated morphism \[ P(F_{\dagger}\nbigs,F_{\dagger}\kappa)_i: F_{\dagger}^i\nbigm_2 \otimes j^{\ast}F_{\dagger}^{-i}\nbigm_2 \lrarr \lambda^{-w+d_Y} \nbigo_{\nbigy} \] induced by the composite of the following morphisms: \begin{equation} \begin{CD} F_{\dagger}^i\nbigm_2 @>{F_{\dagger}^i\nbigs}>> \lambda^{-w}F_{\dagger}^i\nbigm_1 @>{F_{\dagger}^i\kappa}>> \lambda^{-w}F_{\dagger}^ij^{\ast}\DDD\nbigm_2 \simeq \lambda^{-w} j^{\ast}\DDD F_{\dagger}^{-i}\nbigm_2 \simeq \lambda^{-w+d_Y} j^{\ast}\bigl( F_{\dagger}^{-i}\nbigm_2 \bigr)^{\lor} \end{CD} \end{equation} The last isomorphism is given by $(-1)^{d_Y}\nu_{F_{\dagger}^{-i}\nbigm_2}$. We give an expression of $P(F_{\dagger}\nbigs,F_{\dagger}\kappa)_i$ in terms of $P$. \vspace{.1in} Let $Q\in Y$ be any point of $Y$. We set $X_Q:=F^{-1}(Q)$. Let $F_Q:X_Q\lrarr \{Q\}$ be the restriction of $F$. Let $\nbigt_Q:=(\nbigm_{1Q},\nbigm_{2Q},C_Q)$ be the restriction of $\nbigt$ to $X_Q$ as smooth $\nbigr$-triple. Let $F_{\dagger}^i(\nbigt)_Q$ be the restriction of $F_{\dagger}^i(\nbigt)$ to $Q$ as smooth $\nbigr$-triple. We naturally have $F_{Q\dagger}^i(\nbigt_Q) =F_{\dagger}^i(\nbigt)_Q$. We set $\Omegabar_{X_Q}^1:= \lambda^{-1}p_{\lambda}^{\ast}\Omega_{X_Q}^1$ and $\Omegabar_{X_Q}^p:=\bigwedge^p\Omegabar_{X_Q}^1$. We have the de Rham complex $\Omegabar_{X_Q}^{\bullet}\otimes (\nbigm_{2Q})$ for $\nbigm_{2Q}$. We have \[ F_{Q\dagger}^i(\nbigm_{2Q})= R^{i+d}F_{\ast} \Bigl( \Omegabar_{X_Q}^{\bullet}\otimes(\nbigm_{2Q}). \Bigr) \] Set $d:=d_X-d_Y=\dim X_Q$. Let $\Omega^{0,q}_{X_Q}$ denote the sheaf of $C^{\infty}$ $(0,q)$-forms on $X_Q$. Let $\Omega_{\cnum_{\lambda}\times X_Q/\cnum_{\lambda}}^{p,q}:= p_{\lambda}^{\ast}(\Omega_{X_Q}^p) \otimes_{p_{\lambda}^{-1}\nbigo_{X_Q}} p_{\lambda}^{-1}\bigl(\Omega_{X_Q}^{0,q}\bigr)$. The pairing $P$ induces the following: \[ \Bigl( \lambda^{-k}\Omega^{k,p}_{\cnum\times X_Q/\cnum} \otimes\nbigm_2 \Bigr) \times j^{\ast} \Bigl( \lambda^{-(d-k)}\Omega^{d-k,d-p}_{\cnum\times X_Q/\cnum} \otimes\nbigm_2 \Bigr) \lrarr \lambda^{-d-w+d_X} \Omega^{d,d}_{\cnum\times X_Q/\cnum} \] given by $(\xi^{k,p}m_1,\xi^{d-k,d-p}m_2) \longmapsto \xi^{k,p}\xi^{d-k,d-p} P(m_1,m_2)$. Note that $-d-w+d_X=-w+d_Y$. It induces \[ P_{i|Q}: R^{d+i}F_{\ast}\bigl( \Omegabar^{\bullet}_{X_Q}(\nbigm_{2Q}) \bigr) \times j^{\ast}R^{d-i}F_{\ast}\bigl( \Omega^{\bullet}_{X_Q}(\nbigm_2) \bigr) \lrarr \lambda^{-w+d_Y} \nbigo_{\cnum_{\lambda}} \] By varying $Q$, we obtain $P_{F,i}: F_{\dagger}^i\nbigm_2 \otimes F_{\dagger}^{-i}\nbigm_2 \lrarr \lambda^{-w+d_Y} \nbigo_{\nbigy}$. \begin{prop} We have $P(F_{\dagger}\nbigs,F_{\dagger}\kappa)_i =\epsilon(d)(-1)^{di}P_{F,i}$. \end{prop} \pf It is enough to compare the specializations along $\{\lambda\}\times X$ for $\lambda\neq 0$, which follows from Proposition . \hfill\qed In an earlier version, in the proof of Theorem , we used the comparison of the duality functors for $\nbigr$-modules and $\nbigrtilde$-modules. After the proof of the theorem has been simplified, we do not need it in the current version. (See Remark .) Hence, the most results in this section are not used in the other part of this paper. However, because the author hopes that the results in this section would be useful for our understanding of the duality of integrable mixed twistor $\nbigd$-modules, we keep this section. \subsection{Preliminary} \subsubsection{Some sheaves} For any complex manifold $X$, let $\nbigx:=\cnum_{\lambda}\times X$. Let $p_{\lambda}:\nbigx\lrarr X$ be the projection. Set $d_X:=\dim X$. Let $\nbigo_{\cnum_{\lambda}}$ denote the sheaf of functions on $\nbigx$ which are locally constant in the $X$-direction, and holomorphic in the $\cnum_{\lambda}$-direction. Let $\Omega_X^1$ denote the sheaf of holomorphic $1$-forms on $X$. Let $\Omegabar^1_X:= \lambda^{-1}p_{\lambda}^{\ast}\Omega^{1}_X$ and $\Omegabar^p_X:= \bigwedge^p\Omegabar_X^1$. In particular, we set $\Omegabar_X:=\Omegabar_X^{d_X}$. As in the case of $\nbigd$-modules, for any left $\nbigr_X$-module $\nbign^{\ell}$, we naturally obtain a right $\nbigr_X$-module $\Omegabar_X\otimes\nbign^{\ell}$. Conversely, for any right $\nbigr_X$-module $\nbign^r$, we naturally obtain a left $\nbigr_X$-module $\nbign^r\otimes\Omegabar_X^{-1}$. Let $\Theta_X$ be the sheaf of holomorphic vector fields on $X$ . We set $\Thetabar_X:=\lambda p_{\lambda}^{\ast}\Theta_X$. We set $\Thetabar^{p}_X:=\bigwedge^{-p}\Thetabar_X$. We obtain the Spencer resolution $\nbigr_X\otimes\Thetabar^{\bullet}_X$ of $\nbigo_{\nbigx}$ by locally free left $\nbigr_X$-modules. (See .) Let $\nbigx^0:=\{0\}\times X$. Let $\Omega^1_{\nbigx}(\log \nbigx^0)$ be the sheaf of logarithmic holomorphic $1$-forms on $(\nbigx,\nbigx^0)$. We set $\Omegatilde^1_{\nbigx}:= \Omega^1_{\nbigx}(\log\nbigx^0) \otimes\nbigo(\nbigx^0)$, and $\Omegatilde^p_{\nbigx}:= \bigwedge^p\Omegatilde^1_{\nbigx}$. We set $\Omegatilde_{\nbigx}:= \Omegatilde_{\nbigx}^{\dim X+1}$. For any left $\nbigrtilde_X$-module $\nbign^{\ell}$, we naturally obtain a right $\nbigrtilde_X$-module $\Omegatilde_{\nbigx}\otimes\nbign^{\ell}$. Conversely, for any right $\nbigrtilde_X$-module $\nbign^{r}$, we naturally obtain a left $\nbigrtilde_X$-module $\nbign^{r}\otimes\Omegatilde_{\nbigx}^{-1}$. We have the natural isomorphism $\lambda^{-2}\Omegabar_X \simeq \Omegatilde_{\nbigx}$ given by $\tau\longmapsto \tau d\lambda$. Let $\Theta_{\nbigx}(\log \nbigx^0)$ be the sheaf of tangent vectors on $\nbigx$ which are logarithmic along $\nbigx^0$. Set $\Thetatilde_{\nbigx}:=\lambda \Theta_{\nbigx}(\log\nbigx^0)$ and $\Thetatilde^p_{\nbigx}:=\bigwedge^{-p}\Thetatilde_{\nbigx}$. We obtain the Spencer resolution $\nbigrtilde_X\otimes\Thetatilde_{\nbigx}^{\bullet}$ of $\nbigo_{\nbigx}$ by locally free left $\nbigr_{\nbigx}$-modules in the standard way. \subsubsection{Partial Spencer resolution} Let $\nbign$ be any $\nbigr_X$-module. We set $\nbign\langle \lambda^2\del_{\lambda}\rangle:= \bigoplus_{j=0}^{\infty} \del_{\lambda}^j\otimes \lambda^{2j}\nbign$. We have the isomorphism of sheaves $\Phi: \nbigrtilde_X\otimes_{\nbigr_X}\nbign \lrarr \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown \bigr)$ given by $\sum (\del_{\lambda}^j\lambda^{2j})\otimes m_j \longmapsto \sum \del_{\lambda}^j\otimes (\lambda^{2j}m_j)$. The natural $\nbigrtilde_X$-action on $\nbigrtilde_X\otimes_{\nbigr_X}\nbign$ is transferred onto $\nbign\langle\lambda^2\del_{\lambda}\rangle$. We denote it by $\triangledown$ which is explicitly described as follows. \begin{itemize} \item Let $P$ and $m$ denote local sections of $\nbigr_X$ and $\nbign$, respectively. For any $j\geq 0$, we have local sections $P_{ij}$ $(i=0,\ldots,j)$ of $\nbigr_X$ such that $P\del_{\lambda}^j\lambda^{2j} =\sum_{i=0}^j \del_{\lambda}^iP_{ij}\lambda^{2i}$ in $\nbigrtilde_X$. Then, $P\triangledown(\del_{\lambda}^j\otimes\lambda^{2j}m)= \sum_{i=0}^j\del_{\lambda}^i\otimes(P_{ij}\lambda^{2i}m)$. \item We also have $(\del_{\lambda}\lambda^2)\triangledown (\del_{\lambda}^j\otimes \lambda^{2j}m) =\del_{\lambda}^{j+1}\otimes(\lambda^{2j+2}m) -2j\del_{\lambda}^j\otimes \lambda^{2j+1}m +j(j-1)\del_{\lambda}^{j-1}\otimes \lambda^{2j}m$. \end{itemize} We have the inclusion $\iota: \nbign\lrarr \nbign\langle\lambda^2\del_{\lambda}\rangle$ given by $\iota(m)=1\otimes m$. It is an $\nbigr_X$-homomorphism. We have the decomposition: \begin{equation} \nbign\langle\lambda^2\del_{\lambda}\rangle =\bigoplus_{j=0}^{\infty} (\del_{\lambda}\lambda^2)^j\triangledown\iota(\nbign) \end{equation} We also have the $\nbigr_X$-action $\blacktriangledown$ on $\nbign\langle\lambda^2\del_{\lambda}\rangle$ given as follows. \begin{itemize} \item For any $P$ and $m$ as above, we set $P\blacktriangledown(\del_{\lambda}^j\otimes\lambda^{2j}m) =\del_{\lambda}^j\otimes P(\lambda^{2j}m)$. \end{itemize} If moreover $\nbign$ is an $\nbigrtilde_X$-module, the $\nbigr_X$-action $\blacktriangledown$ is extended to the $\nbigrtilde_X$-action $\blacktriangledown$ on $\nbign\langle\lambda^2\del_{\lambda}\rangle$ given as follows: \[ \del_{\lambda}\lambda^2\blacktriangledown (\del_{\lambda}^j\otimes \lambda^{2j}m) =\del_{\lambda}^j\otimes \del_{\lambda}\lambda^{2+2j}m -\del_{\lambda}^{j+1} \otimes\lambda^{2j+2}m. \] It is easy to check that $(\nbign\langle\lambda^2\del_{\lambda}\rangle,\blacktriangledown)$ is an $\nbigrtilde_X$-module. (See also Lemma below.) We have the decomposition: \begin{equation} \nbign\langle\lambda^2\del_{\lambda}\rangle =\bigoplus_{j=0}^{\infty} (\del_{\lambda}\lambda^2)^j\blacktriangledown \iota(\nbign). \end{equation} For any $\ell\in\seisuu$, we set $\lambda^{\ell}\nbign \langle\lambda^2\del_{\lambda}\rangle :=\bigoplus_{j=0}^{\infty} \del_{\lambda}^{j}\otimes \lambda^{2j+\ell}\nbign$. It is equipped with the natural $\nbigrtilde$-actions $\triangledown$ and $\blacktriangledown$, obtained as above. We naturally have $\lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle =\lambda^2\triangledown \nbign\langle\lambda^2\del_{\lambda}\rangle =\lambda^2\blacktriangledown \nbign\langle\lambda^2\del_{\lambda}\rangle$. We have the well defined morphisms of sheaves $\del_{\lambda}\triangledown: \lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle \lrarr \nbign\langle\lambda^2\del_{\lambda}\rangle$ and $\del_{\lambda}\blacktriangledown: \lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle \lrarr \nbign\langle\lambda^2\del_{\lambda}\rangle$. It is easy to check the following lemma. \begin{lem} \mbox{{}} Let $\nbign$ be an $\nbigrtilde_X$-module. \begin{itemize} \item For any local section $m$ of $\nbign$, we have the following relation \begin{equation} (\del_{\lambda}\lambda^2)\triangledown \iota(m) + (\del_{\lambda}\lambda^2)\blacktriangledown \iota(m) =\iota(\del_{\lambda}\lambda^2m). \end{equation} \item For the above isomorphism $\Phi:\nbigrtilde_X\otimes_{\nbigr_X}\nbign \simeq \nbign\langle\lambda^2\del_{\lambda}\rangle$, we have \[ \del_{\lambda}\lambda^2\blacktriangledown \Phi\bigl( \del_{\lambda}^j\lambda^{2j} \otimes m \bigr) =\Phi\bigl( \del_{\lambda}^j\lambda^{2j+2}(-\del_{\lambda}) \otimes m \bigr) +\Phi\bigl( \del_{\lambda}^j\lambda^{2j+2} \otimes \del_{\lambda}m \bigr). \] For any $P\in\nbigr_X$, we have the following: \[ P\blacktriangledown \Phi(\del_{\lambda}^j\lambda^{2j}\otimes m) =\Phi(\del_{\lambda}^j\lambda^{2j}\otimes Pm). \] \item $\del_{\lambda}\triangledown$ gives a morphism $\bigl( \lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr) \lrarr \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr)$, and $\del_{\lambda}\blacktriangledown$ gives a morphism $\bigl( \lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown \bigr) \lrarr \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown \bigr)$. \hfill\qed \end{itemize} \end{lem} By the lemma, for any $\nbigrtilde_X$-module $\nbign$, we obtain the following exact sequence: \[ \begin{CD} 0 @>>> (\lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown) @>{\del_{\lambda}\triangledown}>> (\nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown) @>{a_1}>> \nbign @>>> 0 \end{CD} \] Here, $a_1$ is the projection onto the $0$-th component with respect to the decomposition (). We also have the following exact sequence: \[ \begin{CD} 0 @>>> (\lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown) @>{\del_{\lambda}\blacktriangledown}>> (\nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown) @>{a_2}>> \nbign @>>> 0 \end{CD} \] Here, $a_2$ is the projection onto the $0$-th component with respect to the decomposition (). We obtain the following $\nbigrtilde_X$-resolutions of $\nbign$: \[ S_{\lambda}^{\blacktriangledown}(\nbign):= \Bigl( \bigl( \lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr) \stackrel{\del_{\lambda}\triangledown}{\lrarr} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr) \Bigr) \] \[ S_{\lambda}^{\triangledown}(\nbign):= \Bigl( \bigl( \lambda^2\nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown \bigr) \stackrel{\del_{\lambda}\blacktriangledown}{\lrarr} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown \bigr) \Bigr) \] \begin{example} Let us consider the case $\nbign=\nbigr_X$. We consider the $\nbigr_X$-action given by the left multiplication. We have the isomorphism $\nbigr_X\langle\lambda^2\del_{\lambda}\rangle \simeq \nbigrtilde_X$ given by $\sum \del_{\lambda}^j\otimes \lambda^{2j}m_j \longmapsto \sum \del_{\lambda}^j\lambda^{2j}m_j$. Under the isomorphism, $(\del_{\lambda}\lambda^2)\triangledown$ is the left multiplication of $\del_{\lambda}\lambda^2$, and $(\del_{\lambda}\lambda^2)\blacktriangledown$ is the right multiplication of $-\lambda^2\del_{\lambda}$. Hence, we naturally have $\nbigrtilde_X\otimes\Thetatilde_{\nbigx}^{\bullet} \simeq S_{\lambda}^{\triangledown} \bigl( \nbigr_X\otimes\Thetabar^{\bullet}_X\bigr)$. \hfill\qed \end{example} \subsubsection{Exchange} We define $\Psi: \nbign\langle\lambda^2\del_{\lambda}\rangle \lrarr \nbign\langle\lambda^2\del_{\lambda}\rangle$ by $\Psi\bigl( (\del_{\lambda}\lambda^2)^j\triangledown\iota(m) \bigr) := (\del_{\lambda}\lambda^2)^j\blacktriangledown\iota(m)$. By the construction, $\Psi$ gives an isomorphism $\bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown \bigr) \lrarr \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr)$. \begin{lem} $\Psi$ also gives an isomorphism $\Psi: \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown\bigr) \lrarr \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \triangledown\bigr)$. \end{lem} \pf It is enough to check that $\Psi \bigl( (\del_{\lambda}\lambda^2)\blacktriangledown (\del_{\lambda}\lambda^2)^j\triangledown \iota(m) \bigr) = (\del_{\lambda}\lambda^2)\triangledown \Psi\bigl( (\del_{\lambda}\lambda^2)^j\triangledown \iota(m) \bigr)$ due to the decomposition (). By the commutativity of $(\del_{\lambda}\lambda^2)\blacktriangledown$ and $(\del_{\lambda}\lambda^2)\triangledown$, we have only to check $\Psi\bigl( (\del_{\lambda}\lambda^2)\blacktriangledown \iota(m) \bigr) =(\del_{\lambda}\lambda^2)\triangledown \iota(m)$. It follows from (). \hfill\qed \vspace{.1in} We have the induced isomorphisms $\Psi: S_{\lambda}^{\triangledown}(\nbign) \lrarr S_{\lambda}^{\blacktriangledown}(\nbign)$ and $\Psi: S_{\lambda}^{\blacktriangledown}(\nbign) \lrarr S_{\lambda}^{\triangledown}(\nbign)$. \begin{example} If $\nbign=\nbigr_X$, under the isomorphism $\nbigrtilde_X\simeq\nbigr_X\langle\lambda^2\del_{\lambda}\rangle$, the isomorphism $\Psi:\nbigrtilde_X\lrarr\nbigrtilde_X$ is given by $\Psi\bigl(\sum \del_{\lambda}^jm_j\bigr) =\sum m_j(-\del_{\lambda})^j$. \hfill\qed \end{example} \subsubsection{Bi-modules} Let $\nbigo_{\cnum_{\lambda}} \langle\lambda^2\del_{\lambda}\rangle$ denote the sheaf of subalgebras in $\nbigrtilde_X$ generated by $\lambda^2\del_{\lambda}$ over $\nbigo_{\cnum_{\lambda}}$. Let $\nbigb$ be an $\nbigo_{\cnum_{\lambda}}$-module. An action of $\lambda^2\del_{\lambda}$ on $\nbigb$ means a morphism of sheaves $\rho(\lambda^2\del_{\lambda}):\nbigb\lrarr \nbigb$ such that $\rho(\lambda^2\del_{\lambda})(fP) =f\rho(\lambda^2\del_{\lambda})(P) +(\lambda^2\del_{\lambda}f)\,P$ for any local sections $f\in\nbigo_{\cnum_{\lambda}}$ and $P\in\nbigb$. If $\nbigb$ is a sheaf of algebras over $\nbigo_{\cnum_{\lambda}}$, we moreover impose the Leibniz rule $\rho(\lambda^2\del_{\lambda})(PQ) =\rho(\lambda^2\del_{\lambda})(P)Q +P\rho(\lambda^2\del_{\lambda})(Q)$. If we are given a sheaf of algebras $\nbigb$ over $\nbigo_{\cnum}$ equipped with an action of $\lambda^2\del_{\lambda}$, the sheaf $\nbigb\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigo\langle\lambda^2\del_{\lambda}\rangle$ is naturally a sheaf of algebras. The multiplication is given by $(P\otimes (\lambda^2\del_{\lambda})^k) \cdot (Q\otimes (\lambda^2\del_{\lambda})^{\ell}) =\sum_{j=0}^k (k:j) P\cdot \rho(\lambda^2\del_{\lambda})^j(Q)\otimes (\lambda^2\del_{\lambda})^{k+\ell-j}$, where $(k:j)$ denote the binomial coefficients. We can naturally regard $\nbigr_X$ as a left $\nbigo_{\cnum_{\lambda}} \langle\lambda^2\del_{\lambda}\rangle$. The action of $\lambda^2\del_{\lambda}$ on $\nbigr_X$ is given by $\rho(\lambda^2\del_{\lambda})(P) =[\lambda^2\del_{\lambda},P]$ in $\nbigrtilde_X$ which satisfies the Leibniz rule $\rho(\lambda^2\del_{\lambda})(PQ) =\rho(\lambda^2\del_{\lambda})(P)\cdot Q +P\cdot \rho(\lambda^2\del_{\lambda})(Q)$. The sheaf $\nbigr_X\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigo_{\cnum_{\lambda}}\langle\lambda^2\del_{\lambda}\rangle$ is naturally a sheaf of algebras. We have the isomorphism $\nbigr_X\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigo_{\cnum_{\lambda}}\langle\lambda^2\del_{\lambda}\rangle \simeq \nbigrtilde_X$ given by $\sum P_j\otimes (\lambda^2\del_{\lambda})^j \longmapsto \sum P_j(\lambda^2\del_{\lambda})^j$. Because $\nbigo_{\cnum_{\lambda}}$ is the center of $\nbigr_X$, we have the naturally defined sheaf of algebras $\nbigr_X\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigr_X$ on $\nbigx$. It is naturally an $\nbigo_{\cnum_{\lambda}} \langle\lambda^2\del_{\lambda}\rangle$-module. The action of $\lambda^2\del_{\lambda}$ is given by $\rho_1(\lambda^2\del_{\lambda}) (P\otimes Q) =\rho(\lambda^2\del_{\lambda})(P)\otimes Q +P\otimes\rho(\lambda^2\del_{\lambda})(Q)$. It satisfies the Leibniz rule. We set $\nbiga_X:= \nbigr_X\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigr_X\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigo_{\cnum_{\lambda}}\langle\lambda^2\del_{\lambda}\rangle$. It is naturally a sheaf of algebras. Let $k_i:\nbigr_X\lrarr \nbiga_X$ $(i=1,2)$ be the morphism of algebras given by the inclusion to the $i$-th component. They are extended to the morphism of algebras $k_i:\nbigrtilde_X\lrarr \nbiga_X$. Let $\nbign$ be an $\nbiga_X$-module. Let $\ell$ (resp. $r$) denote the $\nbigr_X$-action or $\nbigrtilde_X$-action on $\nbign$ induced by $k_1$ (resp. $k_2$). We set $\nbign\langle\lambda^2\del_{\lambda}\rangle:= \bigoplus_{j=0}^{\infty} \del_{\lambda}^j\otimes\lambda^{2j}\nbign$. The $\nbigrtilde_X$-action $\ell$ on $\nbign$ and the construction $\blacktriangledown$ give an $\nbigrtilde_X$-action $\ell\blacktriangledown$ on $\nbign\langle\lambda^2\del_{\lambda}\rangle$. The induced multiplication of $g\in\nbigrtilde_X$ and $c\in\nbign\langle\lambda^2\del_{\lambda}\rangle$ is denoted by $g(\ell\blacktriangledown)c$. We obtain an $\nbigr_X$-action by taking the restriction, which is also denoted by $\ell\blacktriangledown$. For $a\in\nbigrtilde_X$, let $\ell\blacktriangledown(a)$ denote the endomorphism of the sheaves $\nbign\langle\lambda^2\del_{\lambda}\rangle$ given by $m\longmapsto a(\ell\blacktriangledown)m$. We use the notation $\ell\triangledown$, $r\triangledown$ and $r\blacktriangledown$ similarly. \begin{lem} We have $r\triangledown(a_1) \circ \ell\blacktriangledown(a_2) =\ell\blacktriangledown (a_2) \circ r\triangledown (a_1)$ for any $a_1,a_2\in\nbigrtilde_X$. \end{lem} \pf By using the explicit construction of $\triangledown$ and $\blacktriangledown$, we can observe that $r\triangledown (a_1)\circ \ell\blacktriangledown(a_2) =\ell\blacktriangledown (a_2) \circ r\triangledown (a_1)$ if $a_1,a_2\in\nbigr_X$. By the assumption, we have $r(\lambda^2\del_{\lambda})=\ell(\lambda^2\del_{\lambda})$. Hence, we have $r\triangledown(\del_{\lambda}\lambda^2) =\ell\triangledown(\del_{\lambda}\lambda^2)$ and $r\blacktriangledown(\del_{\lambda}\lambda^2) =\ell\blacktriangledown(\del_{\lambda}\lambda^2)$. By using Lemma , we obtain $r\triangledown(a_1) \circ \ell\blacktriangledown(a_2) =\ell\blacktriangledown(a_2) \circ r\triangledown(a_1)$ if either one of $a_1$ or $a_2$ is $\del_{\lambda}\lambda^2$. \hfill\qed \vspace{.1in} The following lemma is clear by construction. \begin{lem} We have $r\triangledown(a_1) \circ \ell\triangledown(a_2) =\ell\triangledown(a_2) \circ r\triangledown(a_1)$ and $r\blacktriangledown(a_1) \circ \ell\blacktriangledown(a_2) =\ell\blacktriangledown(a_2) \circ r\blacktriangledown(a_1)$ for any $a_1,a_2\in\nbigr_X$. \hfill\qed \end{lem} As mentioned in the proof of Lemma , we have $r\triangledown(\del_{\lambda}\lambda^2) =\ell\triangledown(\del_{\lambda}\lambda^2)$ and $r\blacktriangledown(\del_{\lambda}\lambda^2) =\ell\blacktriangledown(\del_{\lambda}\lambda^2)$, the morphism $\Psi$ in \S is independent of the choices of $\ell$ and $r$. It exchanges $(r\triangledown,\ell\blacktriangledown)$ and $(r\blacktriangledown,\ell\triangledown)$. \begin{example} We consider $\nbign:=\nbigr_X\otimes\Omegabar_X^{-1}$ as an $\nbiga_X$-module. Here, the $\nbigr_X$-actions $\ell$ and $r$ are induced by the left and the right multiplications. The action of $\lambda^2\del_{\lambda}$ is given by $\lambda^2\del_{\lambda} (P) =[\lambda^2\del_{\lambda},P]$ in $\nbigrtilde_X$. We have the isomorphism $\bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \ell\triangledown \bigr) \simeq \nbigrtilde_X\otimes_{\nbigr_X} \bigl( \nbign,\ell \bigr) =\nbigrtilde_X\otimes\Omegabar_X^{-1} \simeq \nbigrtilde_X\otimes \Omegabar_X^{-1}(d\lambda)^{-1}$. Under the isomorphism, $\ell\triangledown$ is equal to the left multiplication, and $r\blacktriangledown$ is equal to the $\nbigrtilde_X$-action induced by the right multiplication. We have the isomorphism $\bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, r\triangledown \bigr) \simeq \nbigrtilde_X\otimes_{\nbigr_X} \bigl( \nbign,r \bigr) =\nbigrtilde_X\otimes\Omegabar_X^{-1} \simeq \nbigrtilde_X\otimes \Omegabar_X^{-1}(d\lambda)^{-1}$. Under the isomorphism, $r\triangledown$ is equal to the $\nbigrtilde_X$-action induced by the right multiplication, and $\ell\blacktriangledown$ is equal to the left multiplication. \hfill\qed \end{example} \subsubsection{Functoriality for inner homomorphisms} Let $\nbign_1$ be an $\nbigrtilde_X$-module. Let $\nbign_2$ be an $\nbiga_X$-module. Let $\nhom_{\nbigr_X}(\nbign_1,\nbign_2^{\ell})$ be the sheaf of $\nbigr_X$-homomorphisms from $\nbign_1$ to $(\nbign_2,\ell)$. It is equipped with the $\nbigr_X$-action induced by $r$. It is naturally extended to an $\nbigrtilde_X$-action, where the action of $\lambda^2\del_{\lambda}$ is given by $\lambda^2\del_{\lambda}(f)(m)= \lambda^2\del_{\lambda}(f(m)) -f\bigl(\lambda^2\del_{\lambda}m\bigr)$. We have the $\nbigrtilde_X$-actions $\triangledown$ and $\blacktriangledown$ on $\nhom_{\nbigr_X}(\nbign_1,\nbign_2^{\ell}) \langle\lambda^2\del_{\lambda}\rangle$. \vspace{.1in} Let $\nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2\langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} \bigr)$ denote the sheaf of $\nbigr_X$-homomorphisms $\nbign_1\lrarr \bigl( \nbign_2\langle\lambda^2\del_{\lambda}\rangle, \ell\blacktriangledown \bigr)$. By Lemma , we have the $\nbigrtilde_X$-action on $\nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2\langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} \bigr)$ induced by $r\triangledown$ on $\nbign_2\langle\lambda^2\del_{\lambda}\rangle$. The induced $\nbigrtilde_X$-action is also denoted by $r\triangledown$. By Lemma , we have the $\nbigr_X$-action on $\nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2\langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} \bigr)$ induced by $r\blacktriangledown$ on $\nbign_2\langle\lambda^2\del_{\lambda}\rangle$. It is extended to an $\nbigrtilde_X$-action. The action of $\lambda^2\del_{\lambda}$ is given by $(\lambda^2\del_{\lambda}g)(m) =\lambda^2\del_{\lambda}(\ell\blacktriangledown)(g(m)) -g(\lambda^2\del_{\lambda}m)$. The induced $\nbigrtilde_X$-action is denoted by $r\blacktriangledown$. \vspace{.1in} We define a morphism of sheaves $F: \nhom_{\nbigr_X}\bigl( \nbign_1,\nbign_2^{\ell} \bigr)\langle \lambda^2\del_{\lambda}\rangle \lrarr \nhom_{\nbigr_X}\bigl( \nbign_1,\nbign_2^{\ell\blacktriangledown} \langle\lambda^2\del_{\lambda}\rangle \bigr)$ given as follows. For any local section $\sum \del_{\lambda}^j\otimes\lambda^{2j}g_j$ of $\nhom_{\nbigr_X}\bigl( \nbign_1,\nbign_2^{\ell} \bigr)\langle \lambda^2\del_{\lambda}\rangle$ and any local section $m$ of $\nbign_1$, we set \[ F\Bigl( \sum\del_{\lambda}^j\otimes \lambda^{2j}g_j \Bigr)(m) :=\sum \del_{\lambda}^j\otimes \lambda^{2j}g_j(m). \] If $\nbign_1$ is $\nbigr_X$-coherent, then $F$ is an isomorphism. We can check the following lemma by a direct computation. \begin{lem} For local sections $a\in \nbigrtilde_X$ and $b\in \nhom_{\nbigr_X}(\nbign_1,\nbign_2^{\ell}) \langle\lambda^2\del_{\lambda}\rangle$, we have $F(a \triangledown b) =a(r\triangledown) F(b)$ and $F(a \blacktriangledown b) =a(r\blacktriangledown) F(b)$. \hfill\qed \end{lem} We have the following natural isomorphism by the decomposition (): \begin{equation} \nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2 \langle\lambda^2\del_{\lambda} \rangle^{\ell\blacktriangledown} \bigr) \simeq \nhom_{\nbigrtilde_X}\bigl( \nbign_1\langle\lambda^2\del_{\lambda} \rangle^{\blacktriangledown}, \nbign_2 \langle\lambda^2\del_{\lambda} \rangle^{\ell\blacktriangledown} \bigr) \end{equation} \begin{lem} We have the following commutative diagram: \[ \begin{CD} \nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2 \langle\lambda^2\del_{\lambda}\rangle^{\ell\blacktriangledown} \bigr) @>{b}>> \nhom_{\nbigrtilde_X}\bigl( \nbign_1\langle\lambda^2\del_{\lambda}\rangle^{\blacktriangledown}, \nbign_2 \langle\lambda^2\del_{\lambda}\rangle^{\ell\blacktriangledown} \bigr) \\ @V{a_1}VV @V{a_2}VV \\ \nhom_{\nbigr_X}\bigl( \lambda^2\nbign_1, \nbign_2 \langle\lambda^2\del_{\lambda}\rangle^{\ell\blacktriangledown} \bigr) @>{b}>> \nhom_{\nbigrtilde_X}\bigl( \lambda^2\nbign_1 \langle\lambda^2\del_{\lambda}\rangle^{\blacktriangledown}, \nbign_2 \langle\lambda^2\del_{\lambda}\rangle^{\ell\blacktriangledown} \bigr) \end{CD} \] The horizontal arrows $b$ are {\rm()}. The morphism $a_2$ is given by $\del_{\lambda}\triangledown$ on $\nbign_1\langle\lambda^2\del_{\lambda}\rangle$. The morphism $a_1$ is determined by $a_1(g)(m)= -\del_{\lambda}(\ell\blacktriangledown)\bigl(g(m)\bigr) +g(\del_{\lambda}m)$ for any $g\in \nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2\langle\lambda^2\del_{\lambda} \rangle^{\ell\blacktriangledown}\bigr)$ and $m\in \lambda^2\nbign_1$. \end{lem} \pf For any $g\in \nhom_{\nbigr_X}\bigl(\nbign_1, \nbign_2\langle\lambda^2\del_{\lambda} \rangle^{\ell\blacktriangledown}\bigr)$ and $m=\lambda^2m_1\in\lambda^2\nbign_1$, we have \begin{multline} (a_2\circ b)(g)\bigl(\iota(\lambda^2 m_1)\bigr) =b(g)\Bigl( (\del_{\lambda}\lambda^2) \triangledown \iota(m_1) \bigr) \Bigr) =b(g)\Bigl( -(\del_{\lambda}\lambda^2) \blacktriangledown \iota(m_1) +\iota(\del_{\lambda}\lambda^2m_1) \Bigr) \\ =(-\del_{\lambda}\lambda^2) (\ell\blacktriangledown) \bigl(g(m_1)\bigr) +g\bigl(\del_{\lambda}\lambda^2m_1\bigr) = -\del_{\lambda} (\ell\blacktriangledown) \bigl(g(m)\bigr) +g(\del_{\lambda}m) \end{multline} Thus, we obtain the claim of the lemma. \hfill\qed \begin{cor} We have the following natural isomorphism of $\nbigrtilde_X$-complexes: \begin{equation} S_{\lambda}^{\triangledown} \nhom_{\nbigr_X}\bigl( \nbign_1,\nbign_2^{\ell} \bigr) \simeq \nhom_{\nbigrtilde_X} \bigl(S_{\lambda}^{\blacktriangledown}\nbign_1, \lambda^2\nbign_2\langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} \bigr)[1] \end{equation} \begin{equation} S_{\lambda}^{\blacktriangledown} \nhom_{\nbigr_X}\bigl( \nbign_1,\nbign_2^{\ell} \bigr) \simeq \nhom_{\nbigrtilde_X} \bigl(S_{\lambda}^{\triangledown}\nbign_1, \lambda^2\nbign_2\langle\lambda^2\del_{\lambda}\rangle ^{\ell\triangledown} \bigr)[1] \end{equation} \end{cor} \pf We obtain () from Lemma and Lemma . We obtain () by exchanging $\triangledown$ and $\blacktriangledown$. \hfill\qed \subsection{Push-forward} \subsubsection{Push-forward of $\nbigr$-modules and $\nbigrtilde$-modules} We recall the push-forward of $\nbigr$-modules and $\nbigrtilde$-modules. See for more details. Let $X$ be any complex manifold. Let $\nbigc^{\infty}_{\nbigx/\cnum_{\lambda}}$ denote the sheaf of $C^{\infty}$-functions $F$ on $\nbigx$ such that $\delbar_{\lambda}F=0$. Let $\Omega^{0,q}_X$ denote the sheaf of $C^{\infty}$ $(0,q)$-forms on $X$. Let $\nbigc^{\infty}_X$ denote the sheaf of $C^{\infty}$-functions on $X$. We define \[ \nbigcbar_X^{\infty,j}:= \bigoplus_{p+q=j} p_{\lambda}^{-1}\Omega_{X}^{0,q} \otimes_{p_{\lambda}^{-1}\nbigc^{\infty}_X} \bigl( \Omegabar^{p}_X \otimes_{\nbigo_{\nbigx}} \nbigc^{\infty}_{\nbigx/\cnum_{\lambda}} \bigr). \] With the exterior derivative in the $X$-direction, we obtain the complexes $\bigl( \Omegabar^{\bullet}_X,d \bigr)$ and $\bigl( \nbigcbar_X^{\infty,\bullet},d \bigr)$. The natural inclusion gives a quasi-isomorphism $\bigl( \Omegabar^{\bullet}_X,d \bigr) \lrarr \bigl( \nbigcbar_X^{\infty,\bullet},d \bigr)$. Let $f:X\lrarr Y$ be any morphism of complex manifolds. The induced morphism $\nbigx\lrarr\nbigy$ is also denoted by $f$. We set $\nbigr_{X,Y,f}:= \nbigr_X\otimes_{\nbigo_{\cnum_{\lambda}}} f^{-1}\nbigr_Y$. It is naturally a sheaf of algebras. It is also naturally an $\nbigo_{\cnum_{\lambda}} \langle\lambda^2\del_{\lambda}\rangle$-module, and the action of $\lambda^{2}\del_{\lambda}$ satisfies the Leibniz rule. Hence, we obtain the sheaf of algebras $\nbiga_{X,Y,f}:= \nbigr_{X,Y,f}\otimes_{\nbigo_{\cnum_{\lambda}}} \nbigo_{\cnum_{\lambda}} \langle\lambda^2\del_{\lambda}\rangle$ as in the case of $\nbiga_X$. We have a natural inclusion $\nbigrtilde_X\lrarr \nbiga_{X,Y,f}$ and $f^{-1}\nbigrtilde_Y\lrarr \nbiga_{X,Y,f}$. We set $\nbigr_{Y\larr X}:= \Omegabar_{X} \otimes_{\nbigo_{\nbigx}} f^{-1}(\nbigr_Y\otimes \Omegabar_Y^{-1})$. It is naturally a left $\nbigr_{X,Y,f}$-module. We have the natural action of $\lambda^2\del_{\lambda}$ on $f^{-1}(\nbigr_Y)$ given by $\lambda^2\del_{\lambda}(Q):=[\lambda^2\del_{\lambda},Q]$ in $f^{-1}(\nbigrtilde_Y)$. It induces an action of $\lambda^2\del_{\lambda}$ on $\nbigr_{Y\larr X}$, with which $\nbigr_{Y\larr X}$ is an $\nbiga_{X,Y,f}$-module. We have a canonical locally free $\nbigr_{X,Y,f}$-resolution $\bigl(\Omegabar_X^{\bullet}[d_X] \otimes_{\nbigo_{\nbigx}}\nbigr_X\bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1})$ of $\nbigr_{Y\larr X}$. It is also an $\nbiga_{X,Y,f}$-resolution. We have the following natural quasi-isomorphism of $\nbiga_{X,Y,f}$-complexes: \[ \bigl(\Omegabar_X^{\bullet}[d_X] \otimes_{\nbigo_{\nbigx}}\nbigr_X\bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \lrarr \bigl( \nbigcbar^{\infty\bullet}_X[d_X] \otimes_{\nbigo_{\nbigx}}\nbigr_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \] Let $\nbigm^{\bullet}$ be any $\nbigr_X$-complex. We obtain the following $\nbigr_Y$-complex: \begin{multline} f_{\dagger}\nbigm^{\bullet}:= f_{!}\left( \Bigl(\bigl( \nbigcbar^{\infty,\bullet}_X[d_X] \otimes_{\nbigo_{\nbigx}}\nbigr_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \Bigr) \otimes_{\nbigr_X} \nbigm^{\bullet} \right) \\ =f_{!}\left( \Bigl( \nbigcbar^{\infty,\bullet}_X[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \Bigr) \otimes_{\nbigo_{\nbigx}} \nbigm^{\bullet} \right) \end{multline} It induces a functor $D^b(\nbigr_X)\lrarr D^b(\nbigr_Y)$. We denote the right hand side of () by $Rf_!\bigl(\nbigr_{Y\larr X} \otimes_{\nbigr_X}^L\nbigm^{\bullet}\bigr)$. If $\nbigm^{\bullet}$ is cohomologically $\nbigr_X$-coherent and good relative to $f$, then $f_{\dagger}\nbigm^{\bullet}$ is also cohomologically $\nbigr_X$-coherent. If $\nbigm^{\bullet}$ is an $\nbigrtilde_X$-complex, $f_{\dagger}\nbigm^{\bullet}$ is naturally an $\nbigrtilde_Y$-complex. Hence, () gives a functor $D^b(\nbigrtilde_X)\lrarr D^b(\nbigrtilde_Y)$. \subsubsection{Another expression of the push-forward of $\nbigrtilde$-modules} We put \[ \nbigctilde^{\infty,j}_{\nbigx} :=\bigoplus_{p+q=j} p_{\lambda}^{-1}\Omega_X^{0,q} \otimes_{p_{\lambda}^{-1}\nbigc^{\infty}_X} \bigl( \Omegatilde_{\nbigx}^{p} \otimes_{\nbigo_{\nbigx}} \nbigc^{\infty}_{\nbigx/\cnum_{\lambda}} \bigr). \] With the exterior derivative $d$, we obtain the complexes $\bigl( \Omegatilde_{\nbigx}^{\bullet},d \bigr)$ and $\bigl( \nbigctilde_{\nbigx}^{\infty,\bullet},d \bigr)$, and we have the natural quasi-isomorphism $\bigl(\Omegatilde_{\nbigx}^{\bullet},d\bigr) \lrarr \bigl(\nbigctilde_{\nbigx}^{\infty,\bullet}, d\bigr)$. Let $f:X\lrarr Y$ be a morphism of complex manifolds. We set $\nbigrtilde_{X,Y,f}:= \nbigrtilde_X\otimes_{\cnum} f^{-1}\nbigrtilde_Y$. We set $\nbigrtilde_{Y\larr X}:= \Omegatilde_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr)$. It is naturally an $\nbigrtilde_{X,Y,f}$-module. We have a canonical locally free $\nbigrtilde_{X,Y,f}$-resolution $\bigl( \Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{\nbigo_{\nbigx}} \nbigrtilde_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1})$ of $\nbigrtilde_{Y\larr X}$. We have the following natural quasi-isomorphism of $\nbigrtilde_X\otimes f^{-1}\nbigrtilde_Y$-complexes: \[ \bigl(\Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{\nbigo_{\nbigx}}\nbigrtilde_X\bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \lrarr \bigl( \nbigctilde^{\infty,\bullet}_{\nbigx}[d_X+1] \otimes_{\nbigo_{\nbigx}}\nbigrtilde_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigx}} f^{-1}\bigl(\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}\bigr) \] Let $\nbigm^{\bullet}$ be any $\nbigrtilde_X$-complex. We obtain the following $\nbigrtilde_Y$-complex: \begin{multline} \ftilde_{\dagger}\nbigm^{\bullet}:= f_!\left( \Bigl( \bigl( \nbigctilde^{\infty,\bullet}_{\nbigx}[d_X+1] \otimes_{\nbigo_{\nbigx}}\nbigrtilde_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}\bigr) \Bigr) \otimes_{\nbigrtilde_X} \nbigm^{\bullet} \right) \\ =f_!\left( \Bigl( \nbigctilde^{\infty,\bullet}_{\nbigx}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}\bigr) \Bigr) \otimes_{\nbigo_{\nbigx}} \nbigm^{\bullet} \right) \end{multline} It induces a functor $D^b(\nbigrtilde_X)\lrarr D^b(\nbigrtilde_Y)$. We denote the right hand side of () by $Rf_!\bigl( \nbigrtilde_{Y\larr X}\otimes^L_{\nbigrtilde_X}\nbigm \bigr)$. \begin{lem} We have the following natural isomorphism: \begin{multline} \Bigl( \nbigctilde^{\infty,\bullet}_{\nbigx}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}\bigr) \Bigr) \otimes_{\nbigo_{\nbigx}} \nbigm^{\bullet} \simeq \\ S_{\lambda}^{\triangledown} \left( \Bigl( \nbigcbar^{\infty\bullet}_X[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \Bigr) \otimes_{\nbigo_{\nbigx}} \nbigm^{\bullet} \right) \end{multline} \end{lem} \pf Let $\ell$ and $r$ denote the $\nbigr_Y$-action (resp. $\nbigrtilde_Y$-action) on $\nbigr_{Y}\otimes\Omegabar_Y^{-1}$ (resp. $\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}$) induced by the left and right multiplications. Let $\ell\blacktriangledown$ (resp. $r\triangledown$) denote the $\nbigrtilde_Y$-action on $\bigl( \nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \langle\lambda^2\del_{\lambda}\rangle$ induced by $\ell$ and $\blacktriangledown$ (resp. $r$ and $\triangledown$). We have the isomorphism $\Omegabar_Y^{-1}\simeq \lambda^{-2}\Omegatilde_{\nbigy}^{-1}$ given by $\tau\longmapsto \tau(d\lambda)^{-1}$. As mentioned in Example , they induce the following isomorphism: \begin{equation} \Bigl( \nbigrtilde_Y\otimes (\lambda^{-2}\Omegatilde_{\nbigy}^{-1}), \ell,r \Bigr) \simeq \Bigl( \bigl( \nbigr_Y\otimes\Omegabar_Y^{-1} \bigr) \langle\lambda^2\del_{\lambda}\rangle, \ell\blacktriangledown, r\triangledown \Bigr). \end{equation} We obtain () from (). \hfill\qed \begin{cor} We have the natural isomorphism $\ftilde_{\dagger}\nbigm^{\bullet} \simeq S_{\lambda}^{\triangledown}f_{\dagger}\nbigm^{\bullet}$ of $\nbigrtilde_Y$-complexes. In particular, we have a natural isomorphism $\ftilde_{\dagger}\nbigm^{\bullet} \lrarr f_{\dagger}\nbigm^{\bullet}$ in the derived category of $\nbigrtilde_Y$-modules. \hfill\qed \end{cor} \subsubsection{Trace morphisms} Let $\distribution_{\nbigx/\cnum_{\lambda}}$ be the sheaf of distributions $F$ on $\nbigx$ such that $\delbar_{\lambda} F=0$. We set \[ \distributionbar^j_X:= \bigoplus_{p+q=j} p_{\lambda}^{-1}\Omega^{0,q}_X \otimes_{p_{\lambda}^{-1}\nbigc^{\infty}_X} \bigl( \Omegabar_X^p\otimes_{\nbigo_{\nbigx}} \distribution_{\nbigx/\cnum_{\lambda}} \bigr). \] With the exterior derivative $d$ in the $X$-direction, we obtain the complex $\bigl( \distributionbar^{\bullet}_X,d \bigr)$. We have the natural quasi-isomorphism $(\Omegabar^{\bullet}_X,d) \lrarr (\distributionbar^{\bullet}_X,d)$. We also set \[ \distributiontilde^j_{\nbigx}:= \bigoplus_{p+q=j} p_{\lambda}^{-1}\Omega^{0,q}_X \otimes_{p_{\lambda}^{-1}\nbigc^{\infty}_X} \bigl( \Omegatilde_{\nbigx}^p\otimes_{\nbigo_{\nbigx}} \distributiontilde_{\nbigx/\cnum_{\lambda}} \bigr). \] With the exterior derivative, we obtain the complex $\bigl( \distributiontilde^{\bullet}_{\nbigx},d \bigr)$. The natural inclusion $(\Omegatilde_{\nbigx}^{\bullet},d) \lrarr (\distributiontilde_{\nbigx}^{\bullet},d)$ is a quasi-isomorphism. \vspace{.1in} Let $f:X\lrarr Y$ be a proper morphism. Recall that we have the trace morphism $\tr_f:\lambda^{d_X}f_{\dagger}\nbigo_{\nbigx}[d_X] \lrarr \lambda^{d_Y}\nbigo_{\nbigy}[d_Y]$ in the derived category of $\nbigrtilde_Y$-modules: \begin{multline} \lambda^{d_X}f_{\dagger}\nbigo_{\nbigx}[d_X] \simeq f_!\Bigl( \lambda^{d_X} \distributionbar_X^{\bullet}[2d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \Bigr) \simeq f_!\bigl( \lambda^{d_X} \distributionbar_X^{\bullet}[2d_X] \bigr) \otimes_{\nbigo_{\nbigy}} \bigl( \nbigr_Y\otimes\Omegabar_Y^{-1} \bigr) \\ \stackrel{a}{\lrarr} \lambda^{d_Y} \distributionbar_Y^{\bullet}[2d_Y] \otimes_{\nbigo_{\nbigy}} \bigl( \nbigr_Y\otimes\Omegabar_Y^{-1} \bigr) \simeq \lambda^{d_Y}\nbigo_{\nbigy}[d_Y]. \end{multline} Here, $a$ is given by the integration of currents along $f$ multiplied by $(2\pi\sqrt{-1})^{-d_X+d_Y}$. We also have the trace morphism $\trtilde_f: \lambda^{d_X}\ftilde_{\dagger}\nbigo_{\nbigx}[d_X] \lrarr \lambda^{d_Y}\nbigo_{\nbigy}[d_Y]$: \begin{multline} \lambda^{d_X}\ftilde_{\dagger}\nbigo_{\nbigx}[d_X] \simeq f_!\Bigl( \lambda^{d_X}\distributiontilde_{\nbigx}^{\bullet}[2d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \Bigr) \simeq f_!\bigl( \lambda^{d_X}\distributiontilde_{\nbigx}^{\bullet}[2d_X+1] \bigr) \otimes_{\nbigo_{\nbigy}} \bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \\ \stackrel{b}{\lrarr} \lambda^{d_Y}\distributiontilde_{\nbigy}^{\bullet}[2d_Y+1] \otimes_{\nbigo_{\nbigy}} \bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \simeq \lambda^{d_Y} \nbigo_{\nbigy}[d_Y] \end{multline} Here, $b$ is given by the integration of currents along $f$ multiplied by $(2\pi\sqrt{-1})^{-d_X+d_Y}$. The following is clear by the construction of the morphisms. \begin{lem} $\trtilde_f$ is equal to the composite of $\lambda^{d_X}\ftilde_{\dagger}\nbigo_{\nbigx}[d_X] \lrarr \lambda^{d_X}f_{\dagger}\nbigo_{\nbigx}[d_X]$ and $\tr_f$. \end{lem} \pf As in Lemma , we can identify the morphism () with that obtained from () by the functor $S_{\lambda}^{\triangledown}$. Hence, the claim is clear. \hfill\qed \subsubsection{Complement} Let $\nbign$ be any $\nbigrtilde_X$-module. We have the morphism of sheaves \begin{multline} \Bigl( \nbigcbar^{\infty\,\bullet}_X[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar^{-1}_Y) \Bigr) \otimes_{\nbigo_{\nbigx}} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr) \lrarr \\ \Bigl( \bigl( \nbigcbar^{\infty\,\bullet}_X[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar^{-1}_Y) \bigr) \otimes_{\nbigo_{\nbigx}} \nbign \Bigr) \langle\lambda^2\del_{\lambda}\rangle \end{multline} by $g\otimes(\del_{\lambda}^j\otimes\lambda^{2j}m) \longmapsto \del_{\lambda}^j\otimes (g\otimes\lambda^{2j}m)$. By construction, the following holds. \begin{lem} The morphism {\rm()} is an $\nbigrtilde_X$-homomorphism with respect to the natural $f^{-1}\nbigrtilde_Y$-action on the left hand side, and the $f^{-1}\nbigrtilde_Y$-action $\blacktriangledown$ on the right hand side. The morphism {\rm()} is compatible with the actions $(\del_{\lambda}\lambda^2)\triangledown$ on the both sides. \hfill\qed \end{lem} \begin{cor} The morphism {\rm()} induces an isomorphism $f_{\dagger}\bigl( \nbigm^{\bullet}\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr) \lrarr \bigl( f_{\dagger}(\nbigm^{\bullet})\langle\lambda^2\del_{\lambda}\rangle, \blacktriangledown \bigr)$. It also induces an isomorphism $f_{\dagger}S_{\lambda}^{\blacktriangledown}(\nbigm) \simeq S_{\lambda}^{\blacktriangledown} f_{\dagger}(\nbigm)$. \hfill\qed \end{cor} \subsection{Duality} \subsubsection{Duality of $\nbigr$-modules and $\nbigrtilde$-modules} Let $X$ be a complex manifold with a hypersurface $H$. Set $d_X:=\dim X$. We naturally regard $\lambda^{d_X}\nbigr_X(\ast H)\otimes\Omegabar_X^{-1}$ as an $\nbiga_{X}(\ast H)$-module. We take an $\nbiga_{X}(\ast H)$-injective resolution $\nbigg_0^{\bullet}$ of $\lambda^{d_X}\nbigr_X(\ast H) \otimes\Omegabar_X^{-1}[d_X]$. For any $\nbigr_X(\ast H)$-module $\nbign$, let $\nhom_{\nbigr_X(\ast H)}(\nbign,\nbigg_0^{p,\ell})$ denote the sheaf of $\nbigr_X(\ast H)$-homomorphisms $\nbign\lrarr (\nbigg_0^{p},\ell)$. It is naturally an $\nbigr_X(\ast H)$-module by $r$. If $\nbign$ is an $\nbigrtilde_X(\ast H)$-module, it is naturally an $\nbigrtilde_X(\ast H)$-module. The action of $\lambda^2\del_{\lambda}$ is given by $(\lambda^2\del_{\lambda}f)(m) =\lambda^2\del_{\lambda}(f(m)) -f(\lambda^2\del_{\lambda}m)$. For any $\nbigr_X(\ast H)$-complex $\nbigm^{\bullet}$, we obtain the following $\nbigr_X(\ast H)$-complex: \begin{equation} \DDD_{X(\ast H)}\nbigm^{\bullet}:= \nhom_{\nbigr_X(\ast H)}\bigl( \nbigm^{\bullet}, \nbigg_0^{\bullet\,\ell} \bigr) \end{equation} If $\nbigm^{\bullet}$ is an $\nbigrtilde_X(\ast H)$-complex, then $\DDD_{X(\ast H)}\nbigm^{\bullet}$ is naturally an $\nbigrtilde_X(\ast H)$-complex. We shall often denote the right hand side of () by $\nrhom_{\nbigr_X(\ast H)}\bigl(\nbigm^{\bullet}, \lambda^{d_X}\nbigr_X(\ast H)\otimes\Omegabar_X^{-1} \bigr)[d_X]$ if there is no risk of confusion. \subsubsection{Another expression of the duality of $\nbigrtilde$-modules} We give another expression of the duality functor for $\nbigrtilde$-modules. We naturally regard $\nbigrtilde_X(\ast H)\otimes\Omegatilde_{\nbigx}^{-1}$ as $\nbigrtilde_X(\ast H)\otimes_{\cnum} \nbigrtilde_X(\ast H)$-module. Let $\ell$ (resp. $r$) denote the $\nbigrtilde_X(\ast H)$-actions induced by the left multiplication (resp. the right multiplication). We take an injective $\nbigrtilde_X(\ast H) \otimes_{\cnum} \nbigrtilde_X(\ast H) $-resolution $\nbigg_1^{\bullet}$ of $\lambda^{d_X}\nbigrtilde_X(\ast H) \otimes\Omegatilde_{\nbigx}^{-1}[d_X+1]$. For any $\nbigrtilde_X(\ast H)$-module $\nbign$, let $\nhom_{\nbigrtilde_X(\ast H)}(\nbign,\nbigg_1^{p,\ell})$ denote the sheaf of $\nbigrtilde_X(\ast H)$-homomorphisms $\nbign\lrarr (\nbigg_1^{p},\ell)$. It is naturally an $\nbigrtilde_X(\ast H)$-module induced by the $\nbigrtilde_X(\ast H)$-action $r$. For any $\nbigrtilde_X(\ast H)$-complex $\nbigm^{\bullet}$, we obtain the following $\nbigrtilde_X(\ast H)$-complex: \begin{equation} \DDDtilde_{X(\ast H)}(\nbigm^{\bullet}):= \nhom_{\nbigrtilde_X(\ast H)} \bigl( \nbigm^{\bullet}, \nbigg_1^{\bullet\,\ell} \bigr) \end{equation} We shall denote the right hand side of () by $\nrhom_{\nbigrtilde_X(\ast H)}\bigl( \nbigm^{\bullet}, \nbigrtilde_X\otimes\Omegatilde_{\nbigx}^{-1} \bigr)[d_X+1]$ if there is no risk of confusion. \vspace{.1in} We have the $\nbigrtilde_X(\ast H)$-action $\ell\blacktriangledown$ on $(\nbigr_X(\ast H)\otimes\Omegabar_X^{-1}) \langle\lambda^2\del_{\lambda}\rangle$ induced by the $\nbigrtilde_X(\ast H)$-action $\ell$ and the construction $\blacktriangledown$. We also have the $\nbigrtilde_X(\ast H)$-action $r\triangledown$ on $(\nbigr_X(\ast H)\otimes\Omegabar_X^{-1})\langle \lambda^2\del_{\lambda}\rangle$ induced by the $\nbigr(\ast H)$-action $r$ and the construction $\triangledown$. Similarly, we have the $\nbigrtilde_X(\ast H)$-actions $\ell\blacktriangledown$ and $r\triangledown$ on $\nbigg^{\bullet}_0\langle\lambda^2\del_{\lambda}\rangle$. As in (), we have the isomorphism \begin{equation} \Bigl( \bigl( \nbigr_X(\ast H)\otimes\Omegabar_X^{-1} \bigr) \langle\lambda^2\del_{\lambda} \rangle, \ell\blacktriangledown, r\triangledown \Bigr) \simeq \Bigl( \nbigrtilde_X(\ast H)\otimes (\lambda^{-2}\Omegatilde_{\nbigx}^{-1}), \ell,r \Bigr). \end{equation} Hence, we may assume to have a quasi-isomorphism of $\nbigrtilde_X(\ast H) \otimes_{\cnum}\nbigrtilde_X(\ast H)$-complexes \begin{equation} \lambda^{2}\nbigg_0^{\bullet} \langle\lambda^2\del_{\lambda}\rangle[1] \lrarr \nbigg_1^{\bullet}. \end{equation} From Corollary and (), we have the following morphisms of $\nbigrtilde_X(\ast H)$-complexes: \begin{equation} S_{\lambda}^{\triangledown} \nhom_{\nbigr_X(\ast H)}\bigl( \nbigm^{\bullet}, \nbigg_0^{\bullet\,\ell} \bigr) \simeq \nhom_{\nbigrtilde_X(\ast H)}\bigl( S_{\lambda}^{\blacktriangledown} \nbigm^{\bullet}, \lambda^2\nbigg_0^{\bullet} \langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} [1] \bigr) \lrarr \nhom_{\nbigrtilde_X(\ast H)}\bigl( S_{\lambda}^{\blacktriangledown} \nbigm^{\bullet}, \nbigg_1^{\bullet\,\ell} \bigr) \end{equation} We also have the following natural quasi-isomorphisms: \begin{equation} S_{\lambda}^{\triangledown} \nhom_{\nbigr_X(\ast H)}\bigl( \nbigm^{\bullet}, \nbigg_0^{\bullet\,\ell} \bigr) \lrarr \nhom_{\nbigr_X(\ast H)}\bigl( \nbigm^{\bullet}, \nbigg_0^{\bullet,\ell} \bigr) \end{equation} \begin{equation} \nhom_{\nbigrtilde_X(\ast H)}\bigl( \nbigm^{\bullet}, \nbigg_1^{\bullet\,\ell} \bigr) \lrarr \nhom_{\nbigrtilde_X(\ast H)}\bigl( S_{\lambda}^{\blacktriangledown} \nbigm^{\bullet}, \nbigg_1^{\bullet\,\ell} \bigr) \end{equation} \begin{prop} If $\nbigm^{\bullet}$ is cohomologically $\nbigr_X(\ast H)$-coherent, then {\rm()} is a quasi-isomorphism. Together with {\rm()} and {\rm()}, we obtain the following isomorphisms in the derived category of cohomologically $\nbigrtilde_X(\ast H)$-coherent complexes: \[ \begin{CD} \DDDtilde_{X}\nbigm^{\bullet} @>{\simeq}>> \DDDtilde_X S_{\lambda}^{\blacktriangledown}\nbigm^{\bullet} @>{\simeq}>> S_{\lambda}^{\triangledown} \DDD_X\nbigm^{\bullet} @>{\simeq}>> \DDD_X\nbigm^{\bullet} \end{CD} \] \end{prop} \pf For simplicity of the description, we omit to denote $(\ast H)$ in the proof. It is enough to prove the claim for an $\nbigrtilde_X$-module $\nbigm$ which is coherent over $\nbigr_X$. We begin with the following lemma. \begin{lem} Let $0\lrarr\nbign_1\lrarr\nbign_2\lrarr\nbign_3\lrarr 0$ be an exact sequence of coherent $\nbigr_X$-modules. Let $\nbigi$ be any injective $\nbigr_X$-module. Then, the following is exact. \begin{multline} 0\lrarr \nhom_{\nbigrtilde_X} \bigl( \nbign_3\langle\lambda^2\del_{\lambda}\rangle, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \lrarr \nhom_{\nbigrtilde_X} \bigl( \nbign_2\langle\lambda^2\del_{\lambda}\rangle, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \\ \lrarr \nhom_{\nbigrtilde_X} \bigl( \nbign_1\langle\lambda^2\del_{\lambda}\rangle, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \lrarr 0 \end{multline} \end{lem} \pf The sequence () is rewritten as follows: \begin{equation} 0\lrarr \nhom_{\nbigr_X} \bigl( \nbign_3, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \lrarr \nhom_{\nbigr_X} \bigl( \nbign_2, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \stackrel{\beta}{\lrarr} \nhom_{\nbigr_X} \bigl( \nbign_1, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \lrarr 0 \end{equation} It is enough to prove that $\beta$ is an epimorphism. For any non-negative integer $L$, we set $F_L\nbigi\langle\lambda^2\del_{\lambda}\rangle:= \bigoplus_{j=0}^L\del_{\lambda}^j\otimes\lambda^{2j}\nbigi$. They are naturally an $\nbigr_X$-submodules of $\nbigi\langle\lambda^2\del_{\lambda}\rangle$. Let $F_L\nhom_{\nbigr_X} \bigl(\nbign_i, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr)$ be the image of $\nhom_{\nbigr_X} \bigl(\nbign_i, F_L\nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr)$. Because $\nbign_i$ is $\nbigr_X$-coherent, the filtration $F$ is exhaustive on $\nhom_{\nbigr_X} \bigl(\nbign_i, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr)$. By construction, we have $\Gr^L_j\nhom_{\nbigr_X}\bigl( \nbign_i, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \simeq \nhom_{\nbigr_X}\bigl( \nbign_i, \lambda^{2j}\nbigi \bigr)$. Because $\nbigi$ is $\nbigr$-injective, we have the surjectivity of $\Gr^L_j\nhom_{\nbigr_X}\bigl( \nbign_2, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr) \lrarr \Gr^L_j\nhom_{\nbigr_X}\bigl( \nbign_1, \nbigi\langle\lambda^2\del_{\lambda}\rangle \bigr)$. Then, we obtain the desired surjectivity. \hfill\qed \vspace{.1in} The claim of Proposition is reduced to the following lemma. \begin{lem} For any coherent $\nbigr_X$-module $\nbign$, the natural morphism \begin{equation} \nhom_{\nbigrtilde_X} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \lambda^2\nbigg_0^{\bullet} \langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} [1] \bigr) \lrarr \nhom_{\nbigrtilde_X} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \nbigg_1^{\bullet\,\ell} \bigr) \end{equation} is a quasi-isomorphism. \end{lem} \pf The morphism () is a quasi-isomorphism if $\nbign=\nbigr_X$. It is enough to prove that () is a quasi-isomorphism locally around any point of $X$. We may assume to have an $\nbigr_X$-free resolution $\nbigp^{\bullet}$ of $\nbign$. We have the following commutative diagram: \[ \begin{CD} \nhom_{\nbigrtilde_X} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \lambda^2\nbigg_0^{\bullet} \langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} [1] \bigr) @>{\alpha_1}>> \nhom_{\nbigrtilde_X} \bigl( \nbign\langle\lambda^2\del_{\lambda}\rangle, \nbigg_1^{\bullet\,\ell} \bigr) \\ @V{\alpha_2}VV @V{\alpha_3}VV \\ \Tot \nhom_{\nbigrtilde_X} \bigl( \nbigp^{\bullet}\langle\lambda^2\del_{\lambda}\rangle, \lambda^2\nbigg_0^{\bullet} \langle\lambda^2\del_{\lambda}\rangle ^{\ell\blacktriangledown} [1] \bigr) @>{\alpha_4}>> \Tot \nhom_{\nbigrtilde_X} \bigl( \nbigp^{\bullet}\langle\lambda^2\del_{\lambda}\rangle, \nbigg_1^{\bullet\,\ell} \bigr) \end{CD} \] Here, $\Tot\nbigc^{\bullet\bullet}$ denote the total complex of a double complex $\nbigc^{\bullet\bullet}$. As remarked above, $\alpha_4$ is a quasi-isomorphism. Because $(\nbigg_1^{\bullet},\ell)$ is $\nbigrtilde_X$-injective, $\alpha_3$ is a quasi-isomorphism. As for $\alpha_2$, note that $(\lambda^2\nbigg_0^{\bullet},\ell)$ is $\nbigr_X$-injective. We have the isomorphism exchanging $\triangledown$ and $\blacktriangledown$ as in \S. Hence, $\alpha_2$ is a quasi-isomorphism by Lemma . Then, we obtain that $\alpha_1$ is a quasi-isomorphism. Thus we obtain Lemma , and hence Proposition . \hfill\qed \subsubsection{Smooth case} Let $\nbigm$ be a smooth $\nbigrtilde_X(\ast H)$-module. Let $\nbigm^{\lor}$ be the smooth $\nbigrtilde_X(\ast H)$-module $\nhom_{\nbigo_{\nbigx}(\ast H)} \bigl(\nbigm,\nbigo_{\nbigx}(\ast H)\bigr)$ with the induced meromorphic connection. As remarked in , we have a natural isomorphism $\DDD_{X(\ast H)}\nbigm \simeq \lambda^{d_X}\nbigm^{\lor}$ in the derived category. Together with Proposition , we obtain $\DDDtilde_{X(\ast H) }\nbigm\simeq \lambda^{d_X}\nbigm^{\lor}$. We also have the isomorphism $\DDDtilde_{X(\ast H)}\nbigm \simeq \lambda^{d_X}\nbigm^{\lor}$ given in a more direct and standard way, by using the Spencer resolution $\nbigrtilde_X(\ast H) \otimes\Thetatilde^{\bullet}_{\nbigx}$ of $\nbigo_{\nbigx}(\ast H)$: \begin{multline} \nrhom_{\nbigrtilde_X(\ast H)} \bigl( \nbigm, \lambda^{d_X}\nbigrtilde_X(\ast H) \otimes \Omegatilde_{\nbigx}^{-1} \bigr)[d_X+1] \simeq \\ \nhom_{\nbigrtilde_X(\ast H)} \bigl( \nbigrtilde_X(\ast H) \otimes\Thetatilde^{\bullet}_{\nbigx} \otimes\nbigm, \lambda^{d_X}\nbigrtilde_X(\ast H) \otimes \Omegatilde_{\nbigx}^{-1} \bigr)[d_X+1] \simeq \\ \nbigrtilde_X(\ast H) \otimes\Thetatilde^{\bullet}_{\nbigx} \otimes\lambda^{d_X}\nbigm^{\lor} \simeq \lambda^{d_X}\nbigm^{\lor} \end{multline} \begin{lem} The isomorphisms $\DDDtilde_{X(\ast H)}\nbigm \simeq\lambda^{d_X}\nbigm^{\lor}$ are equal. \end{lem} \pf We have the natural quasi-isomorphisms $S^{\blacktriangledown}_{\lambda} \bigl((\nbigr_X(\ast H) \otimes_{\nbigo_{\nbigx}}\Thetabar_{X}^{\bullet}) \otimes_{\nbigo_{\nbigx}}\nbigm\bigr) \lrarr S^{\blacktriangledown}_{\lambda}(\nbigm) \lrarr \nbigm$. We have the following natural isomorphisms: \[ S^{\blacktriangledown}_{\lambda} \bigl((\nbigr_X(\ast H) \otimes_{\nbigo_{\nbigx}}\Thetabar_{\nbigx}^{\bullet}) \otimes_{\nbigo_{\nbigx}}\nbigm\bigr) \simeq S^{\triangledown}_{\lambda} \bigl((\nbigr_X(\ast H) \otimes_{\nbigo_{\nbigx}}\Thetabar_{\nbigx}^{\bullet}) \otimes_{\nbigo_{\nbigx}}\nbigm\bigr) \simeq \bigl( \nbigrtilde_X(\ast H) \otimes_{\nbigo_{\nbigx}} \Thetatilde^{\bullet}_{\nbigx} \bigr) \otimes_{\nbigo_{\nbigx}}\nbigm \] The isomorphism in Proposition is expressed as follows: \begin{multline} \nhom_{\nbigrtilde_X(\ast H)} \bigl( \nbigrtilde_{X}(\ast H) \otimes\Thetatilde^{\bullet}_{\nbigx} \otimes\nbigm, \nbigrtilde_{X}(\ast H) \otimes \Omegatilde_{\nbigx}^{-1} \bigr)[d_X+1] \simeq \\ \nhom_{\nbigrtilde_X(\ast H)} \bigl( S^{\blacktriangledown}_{\lambda}\bigl( \nbigr_X(\ast H)\otimes\Thetabar^{\bullet}_X\otimes \nbigm \bigr), \nbigrtilde_{X}(\ast H) \otimes \Omegatilde_{\nbigx}^{-1} \bigr)[d_X+1] \simeq \\ S^{\triangledown}_{\lambda} \nhom_{\nbigr_X(\ast H)} \bigl( \nbigr_X(\ast H)\otimes\Thetabar^{\bullet}_X\otimes \nbigm, \nbigr_{X}(\ast H) \otimes \Omegabar_{X}^{-1} \bigr)[d_X] \simeq \\ \nhom_{\nbigr_X(\ast H)}\bigl( \nbigr_X(\ast H)\otimes\Thetabar^{\bullet}_X\otimes \nbigm, \nbigr_{X}(\ast H) \otimes \Omegabar_{X}^{-1} \bigr)[d_X] \end{multline} Then, the claim is clear by the constructions of the isomorphisms $\DDDtilde_{X(\ast H)}\nbigm \simeq\lambda^{d_X}\nbigm^{\lor}$. \hfill\qed \subsection{Compatibility of push-forward and duality} \subsubsection{Compatibility for $\nbigr$-modules} Let us recall the compatibility of the push-forward and the duality for $\nbigr$-modules studied in . Let $Y$ be a complex manifold. We set $\nbigm_{Y0}:=(\nbigr_Y\otimes\Omegabar_Y^{-1}) \lefttop{\ell}\otimes^{\ell}_{\nbigo_{\nbigy}} (\nbigr_Y\otimes\Omegabar_Y^{-1})$, where $\lefttop{\ell}\otimes^{\ell}$ means that we consider the actions $\ell$ for the tensor product over $\nbigo_{\nbigy}$. It is equipped with the $\nbigr_Y$-actions $r_i$ $(i=1,2)$ induced by the action $r$ on the $i$-th factor. We also have the $\nbigr_Y$-action $\ell\otimes\ell$ induced by the actions $\ell$. We have the action of $\lambda^2\del_{\lambda}$ on $\nbigm_{Y0}$ induced by the actions on $\nbigr_Y\otimes\Omegabar_Y^{-1}$ with the Leibniz rule. Each $\nbigr_Y$-action with the action of $\lambda^2\del_{\lambda}$ gives an $\nbigrtilde_Y$-action. We set $\nbigm_{Y1}:= (\nbigr_Y\otimes\Omegabar_Y^{-1}) \lefttop{r}\otimes^{\ell}_{\nbigo_{\nbigy}} (\nbigr_Y\otimes\Omegabar_Y^{-1})$, where $\lefttop{r}\otimes^{\ell}$ means that we consider the action $r$ on the first factor and the action $\ell$ on the second factor for the tensor product over $\nbigo_{\nbigy}$. It is equipped with the $\nbigr_Y$-action $\ell_1$ induced by the action $\ell$ on the first factor, and the $\nbigr_Y$-action $r_2$ induced by the action $r$ on the second factor. We also have the $\nbigr_Y$-action $r\otimes\ell$ induced by the action $r$ on the first factor and the action $\ell$ on the second factor. As in the case of $\nbigm_{Y0}$, we have the action of $\lambda^2\del_{\lambda}$ on $\nbigm_{Y1}$ with which each $\nbigr_Y$-action is extended to an $\nbigrtilde_Y$-action. Recall that we have a unique isomorphism of sheaves $\Psi_1:\nbigm_{Y1}\lrarr\nbigm_{Y0}$ such that (i) $\Psi_1\circ r_2=r_2\circ \Psi_1$, $\Psi_1\circ\ell_1=(\ell\otimes\ell)\circ\Psi_1$ and $\Psi_1\circ(r\otimes\ell)=r_1\circ\Psi_1$, (ii) $\Psi_1$ induces the identity on $\Omegabar_Y^{-1}\otimes\Omegabar_Y^{-1}$. Here, $\Psi_1\circ r_2=r_2\circ\Psi_1$ means $\Psi_1(r_2(P)m)=r_2(P)\Psi_1(m)$ for any local sections $P\in\nbigr_X$ and $m\in \nbigm_{Y1}$, and the meaning of $\Psi_1\circ\ell_1=(\ell\otimes\ell)\circ\Psi_1$ and $\Psi_1\circ(r\otimes\ell)=r_1\circ\Psi_1$ are similar. The uniqueness is clear by the conditions (i) and (ii). Locally, $\Psi$ is given as follows. For any local frame $\tau$ of $\Omegabar_X^{-1}$, we have $\Psi_1 \bigl((Q\otimes\tau)\otimes m \bigr) =(\ell\otimes\ell)(Q)\bigl((1\otimes\tau)\otimes m\bigr)$ for $Q\in\nbigr_X$ and $m\in \nbigr_X\otimes\Omegabar_X^{-1}$. It is compatible with the actions of $\lambda^{2}\del_{\lambda}$. \vspace{.1in} Let $f:X\lrarr Y$ be a proper morphism. We obtain the morphism \begin{equation} \lambda^{d_X} f_{\dagger}\Bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \Bigr)[d_X] \lrarr \lambda^{d_Y} \nbigr_Y\otimes\Omegabar_Y^{-1}[d_Y] \end{equation} in the derived category of $\nbiga_Y$-modules as the composite of the following morphisms (see \S for $\nbiga_Y$): \begin{multline} \lambda^{d_X} f_{\dagger}\Bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \Bigr)[d_X] \\ \simeq \lambda^{d_X} f_{!}\Bigl( \Bigl( \bigl( \distributionbar^{\bullet}_X[2d_X] \otimes \nbigr_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr)\otimes_{\nbigr_X} \bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \bigr) \Bigr) \\ \simeq \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigm_{Y0} \Bigr) \stackrel{a_1}{\simeq} \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigm_{Y1} \Bigr) \\ \simeq \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr) \otimes_{\nbigo_{\nbigy}} \bigl( \nbigr_Y\otimes\Omegabar_{Y}^{-1} \bigr) \\ \simeq \lambda^{d_X}f_{\dagger}(\nbigo_{\nbigx}[d_X]) \otimes_{\nbigo_{\nbigy}} \bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr) \stackrel{a_2}{\lrarr} \lambda^{d_Y} \bigl(\nbigr_Y\otimes\Omegabar_Y^{-1}\bigr)[d_Y] \end{multline} Here, $a_1$ is induced by $\Psi_1^{-1}$, and $a_2$ is induced by the trace morphism (see \S). Let $\nbigm$ be a coherent $\nbigr_X$-module which is good relative to $f$. Then, we have the natural morphism \begin{equation} f_{\dagger}\DDD_X\nbigm\lrarr \DDD_Y f_{\dagger}\nbigm \end{equation} obtained as the composite of the following morphisms: \begin{multline} f_{\dagger}\DDD_X\nbigm \simeq Rf_!\nrhom_{\nbigr_X}\Bigl( \nbigm, \lambda^{d_X}\nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1})[d_X] \Bigr) \\ \lrarr Rf_!\nrhom_{f^{-1}\nbigr_Y}\Bigl( \nbigr_{Y\larr X}\otimes^L_{\nbigr_X}\nbigm, \nbigr_{Y\larr X}\otimes^L_{\nbigr_X} \bigl( \lambda^{d_X}\nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \bigr)[d_X] \Bigr) \\ \lrarr \nrhom_{\nbigr_Y}\Bigl( f_{\dagger}\nbigm, \lambda^{d_X}f_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr)[d_X] \Bigr) \\ \lrarr \nrhom_{\nbigr_Y}\Bigl( f_{\dagger}\nbigm, \lambda^{d_Y}\nbigr_Y\otimes\Omegabar_Y^{-1}[d_Y] \Bigr) \simeq \DDD_Y f_{\dagger}\nbigm \end{multline} If $\nbigm$ is also an $\nbigrtilde_X$-module, then () is a morphism in the derived category of $\nbigrtilde_Y$-modules. \subsubsection{Compatibility for $\nbigrtilde$-modules} We set $\nbigmtilde_{Y0}:= (\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \lefttop{\ell}\otimes^{\ell}_{\nbigo_{\nbigy}} (\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1})$. The tensor product is taken as in the case of $\nbigm_{Y0}$. We have the $\nbigrtilde_X$-actions $r_i$ $(i=1,2)$ and $\ell\otimes\ell$ on $\nbigmtilde_{Y0}$ as in the case of $\nbigm_{Y0}$. We set $\nbigmtilde_{Y1}:= (\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \lefttop{r}\otimes^{\ell}_{\nbigo_{\nbigy}} (\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1})$, where the tensor product is taken as in the case of $\nbigm_{Y1}$. We have the $\nbigrtilde_X$-actions $r_2$, $\ell_1$ and $r\otimes\ell$ on $\nbigmtilde_{Y1}$ as in the case of $\nbigm_{Y1}$. We have a unique isomorphism of sheaves $\Psitilde_1:\nbigmtilde_{Y1}\lrarr\nbigmtilde_{Y0}$ such that (i) $\Psitilde_1\circ r_2=r_2\circ \Psitilde_1$, $\Psitilde_1\circ\ell_1=(\ell\otimes\ell)\circ\Psitilde_1$, (ii) $\Psitilde_1$ induces the identity on $\Omegatilde_{\nbigx}^{-1} \otimes \Omegatilde_{\nbigx}^{-1}$. For any local frame $\tau$ of $\Omegatilde_{\nbigy}^{-1}$, we have $\Psitilde_1\bigl( (Q\otimes\tau)\otimes m \bigr)= (\ell\otimes\ell)(Q) \bigl((1\otimes\tau)\otimes m\bigr)$. \begin{lem} We have $\Psitilde_1\circ(r\otimes\ell) =r_1\circ\Psitilde_1$. \end{lem} \pf It is enough to prove the equality locally around any point of $\nbigx\setminus\nbigx^0$. We may assume to have a frame $\tau$ of $\Omegatilde_{\nbigy}^{-1}$ such that $r(v)(\tau^{-1})=0$ for any $v\in\Thetatilde_{\nbigx}$. Let $Q\in\nbigrtilde_Y$ and $m\in \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}$. For any $g\in\nbigo_{\nbigy}$, we have \begin{multline} r_1(g)\Psitilde_1\bigl((Q\otimes\tau)\otimes m\bigr) =r_1(g)(\ell\otimes\ell)(Q)\bigl((1\otimes\tau)\otimes m\bigr) =(\ell\otimes\ell)(Q)\bigl((g\otimes\tau)\otimes m\bigr) \\ =(\ell\otimes\ell)(Q)\bigl((1\otimes\tau)\otimes gm\bigr) =\Psitilde_1\bigl((Q\otimes\tau)\otimes gm\bigr) =\Psitilde_1\bigl( (r\otimes\ell)(g)\bigl((Q\otimes\tau)\otimes m\bigr) \bigr). \end{multline} For any $v\in \Thetatilde_{\nbigx}$, we have \begin{multline} \Psitilde_1\Bigl( (r\otimes\ell)(v)\bigl((Q\otimes\tau)\otimes m\bigr) \Bigr) =\Psitilde_1\bigl((-Qv\otimes\tau)\otimes m\bigr) +\Psitilde(Q\otimes vm) \\ =-(\ell\otimes\ell)(Qv)\bigl((1\otimes\tau)\otimes m\bigr) +(\ell\otimes\ell)(Q)\bigl((1\otimes\tau)\otimes vm\bigr) =(\ell\otimes\ell)(Q)\bigl((-v\otimes\tau)\otimes m\bigr) \\ =(\ell\otimes\ell)(Q)r_1(v) \bigl((1\otimes\tau)\otimes m\bigr) =r_1(v)\Psitilde_1\bigl((Q\otimes\tau)\otimes m\bigr) \end{multline} Thus, we are done. \hfill\qed \vspace{.1in} Let $f:X\lrarr Y$ be a proper morphism as in \S. We obtain the morphism \begin{equation} \lambda^{d_X} \ftilde_{\dagger}\Bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}\bigr) \Bigr)[d_X+1] \lrarr \lambda^{d_Y} \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}[d_Y+1] \end{equation} in the derived category of $\nbigrtilde_Y\otimes_{\cnum} \nbigrtilde_Y$-modules as the composite of the following morphisms: \begin{multline} \lambda^{d_X} \ftilde_{\dagger}\Bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}\bigr) \Bigr)[d_X+1] \\ \simeq \lambda^{d_X} f_{!}\Bigl[ \Bigl( \bigl( \distributiontilde^{\bullet}_{\nbigx/\cnum_{\lambda}}[2d_X+2] \otimes \nbigrtilde_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr)\otimes_{\nbigrtilde_X} \bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}\bigr) \bigr) \Bigr] \\ \simeq \lambda^{d_X} f_!\Bigl( \distributiontilde^{\bullet}_{\nbigx/\cnum_{\lambda}}[2d_X+2] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigmtilde_{Y0} \Bigr) \stackrel{\atilde_1}{\simeq} \lambda^{d_X} f_!\Bigl( \distributiontilde^{\bullet}_{\nbigx/\cnum_{\lambda}}[2d_X+2] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigmtilde_{Y1} \Bigr) \\ \simeq \lambda^{d_X} f_!\Bigl( \distributiontilde^{\bullet}_{\nbigx/\cnum_{\lambda}}[2d_X+2] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigy}} \bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \\ \simeq \lambda^{d_X}\ftilde_{\dagger}(\nbigo_{\nbigx}[d_X+1]) \otimes_{\nbigo_{\nbigy}} \bigl(\nbigrtilde_X\otimes\Omegatilde_{\nbigx}^{-1}\bigr) \stackrel{\atilde_2}{\lrarr} \lambda^{d_Y} \bigl(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}\bigr)[d_Y+1] \end{multline} Here, $\atilde_1$ is induced by $\Psitilde_1^{-1}$, and $\atilde_2$ is induced by the trace morphism (see \S). Let $\nbigm$ be a coherent $\nbigrtilde_X$-module which is good relative to $f$. Then, we have the natural isomorphism \begin{equation} \ftilde_{\dagger}\DDDtilde_X\nbigm\lrarr \DDDtilde_Y \ftilde_{\dagger}\nbigm \end{equation} in the derived category of $\nbigrtilde_Y$-modules, obtained as the composite of the following morphisms: \begin{multline} \ftilde_{\dagger}\DDDtilde_X\nbigm \simeq Rf_!\nrhom_{\nbigrtilde_X}\Bigl( \nbigm, \lambda^{d_X}\nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1})[d_X+1] \Bigr) \\ \lrarr Rf_!\nrhom_{f^{-1}\nbigrtilde_Y}\Bigl( \nbigrtilde_{Y\larr X}\otimes^L_{\nbigrtilde_X}\nbigm, \nbigrtilde_{Y\larr X}\otimes^L_{\nbigrtilde_X} \bigl( \lambda^{d_X}\nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \bigr)[d_X+1] \Bigr) \\ \lrarr \nrhom_{\nbigrtilde_Y}\Bigl( \ftilde_{\dagger}\nbigm, \lambda^{d_X}\ftilde_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr)[d_X+1] \Bigr) \\ \lrarr \nrhom_{\nbigrtilde_Y}\Bigl( \ftilde_{\dagger}\nbigm, \lambda^{d_Y}\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}[d_Y+1] \Bigr) \simeq \DDDtilde_Y \ftilde_{\dagger}\nbigm. \end{multline} \subsubsection{Comparison} Let $f:X\lrarr Y$ be a proper morphism of complex manifolds as above. We shall prove the following proposition in \S--. \begin{prop} Let $\nbigm$ be a coherent $\nbigrtilde_X$-module which is good relative to $f$. The following diagram is commutative: \begin{equation} \begin{CD} \ftilde_{\dagger}\DDDtilde_X \nbigm @>{\simeq}>> \DDDtilde_Y \ftilde_{\dagger}\nbigm \\ @V{\simeq}VV @V{\simeq}VV \\ f_{\dagger}\DDD_X\nbigm @>{\simeq}>> \DDD_Y f_{\dagger}\nbigm \end{CD} \end{equation} Here, the horizontal arrows are given in {\rm\S} and {\rm\S}, and the vertical arrows are given by the isomorphisms in Corollary {\rm} and Proposition {\rm}. \end{prop} \subsubsection{Step 1} We construct a natural transform $\ftilde_{\dagger}\DDDtilde_X \lrarr \DDDtilde_Y f_{\dagger}$ modifying the constructions in \S and \S. We set $\nbigm'_{Y0}:= (\nbigr_Y\otimes\Omegabar_Y^{-1}) \lefttop{\ell}\otimes^{\ell}_{\nbigo_{\nbigy}} \bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr)$. The tensor product is taken as in the case of $\nbigm_{Y0}$. It is equipped with the $\nbigrtilde_X$-action $r_2$ induced by the $\nbigrtilde_X$-action $r$ on the second factor. It is equipped with the $\nbigrtilde_X$-action $\ell\otimes\ell$ induced by the $\nbigrtilde_X$-actions $\ell$ on the first and second factors. It is equipped with the $\nbigr_X$-action $r_1$ induced by the $\nbigr_X$-action $r$ on the first factor. Together with the action of $\lambda^2\del_{\lambda}$ by $\ell\otimes\ell$, it is extended to an $\nbigrtilde_X$-action which is also denoted by $r_1$. We set $\nbigm'_{Y1}:= (\nbigr_Y\otimes\Omegabar_Y^{-1}) \lefttop{r}\otimes^{\ell}_{\nbigo_{\nbigy}} \bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr)$. The tensor product is taken as in the case of $\nbigm_{Y1}$. It is equipped with the $\nbigrtilde_X$-action $r_2$ induced by the $\nbigrtilde_X$-action $r$ on the second factor. It is equipped with the $\nbigrtilde_X$-action $r\otimes\ell$ induced by the $\nbigrtilde_X$-actions $r$ and $\ell$ on the first and second factors, respectively. It is equipped with the $\nbigr_X$-action $\ell_1$ induced by the $\nbigr_X$-action $\ell$ on the first factor. Together with the action of $\lambda^2\del_{\lambda}$ by $r\otimes\ell$, it is extended to an $\nbigrtilde_X$-action which is also denoted by $\ell_1$. We have the isomorphism $\Psi'_1:\nbigm'_{1Y}\simeq\nbigm'_{0Y}$ determined by the conditions (i) $\Psi'_1\circ\ell_1=(\ell\otimes\ell)\circ\Psi_1$ and $\Psi'_1\circ r_2=r_2\circ\Psi_1$, (ii) $\Psi'_1$ induces the identity on $\Omegabar_{Y}^{-1}\otimes\Omegatilde_{\nbigy}^{-1}$. It also satisfies the condition $\Psi'_1\circ (r\otimes\ell)=r_1\circ\Psi'_1$, which can be checked by the argument in Lemma . Then, we obtain the morphism \begin{equation} \lambda^{d_X}f_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \Bigr)[d_X+1] \lrarr \lambda^{d_Y}\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}[d_Y+1] \end{equation} in the derived category of $\nbigrtilde_Y\otimes\nbigrtilde_Y$-complexes, obtained as the composite of the following: \begin{multline} \lambda^{d_X}f_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr)[d_X+1] \\ \simeq \lambda^{d_X} f_{!}\Bigl( \Bigl( \bigl( \distributionbar^{\bullet}_X[2d_X+1] \otimes \nbigr_X \bigr) \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr)\otimes_{\nbigr_X} \bigl( \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl(\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}\bigr) \bigr) \Bigr) \\ \simeq \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigm'_{Y0} \Bigr) \stackrel{a'_1}{\simeq} \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigm'_{Y1} \Bigr) \\ \simeq \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_{Y}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigy}} \bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \\ \simeq \lambda^{d_X}f_{\dagger}(\nbigo_{\nbigx}[d_X+1]) \otimes_{\nbigo_{\nbigy}} \bigl(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}\bigr) \stackrel{a'_2}{\lrarr} \lambda^{d_Y} \bigl(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}\bigr)[d_Y+1] \end{multline} Here $a_1'$ is induced by $\Psi_1'$, and $a_2'$ is induced by the trace morphism in \S. Then, for any $\nbigrtilde_X$-module $\nbigm$ which is good relative to $f$ and coherent over $\nbigr_X$, we obtain the morphism \begin{equation} \ftilde_{\dagger}\DDDtilde_X\nbigm \lrarr \DDDtilde_Y f_{\dagger}\nbigm \end{equation} as the composite of the following morphisms: \begin{multline} \ftilde_{\dagger}\DDDtilde_X\nbigm \simeq Rf_!\nrhom_{\nbigrtilde_X}\Bigl( \nbigm, \lambda^{d_X}\nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1})[d_X+1] \Bigr) \\ \lrarr Rf_!\nrhom_{f^{-1}\nbigrtilde_Y}\Bigl( \nbigr_{Y\larr X}\otimes^L_{\nbigr_X}\nbigm, \nbigr_{Y\larr X}\otimes^L_{\nbigr_X} \bigl( \lambda^{d_X}\nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}) \bigr)[d_X+1] \Bigr) \\ \lrarr \nrhom_{\nbigrtilde_Y}\Bigl( f_{\dagger}\nbigm, \lambda^{d_X}f_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr)[d_X+1] \Bigr) \\ \lrarr \nrhom_{\nbigrtilde_Y}\Bigl( f_{\dagger}\nbigm, \lambda^{d_Y}\nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1}[d_Y+1] \Bigr) \simeq \DDDtilde_Y f_{\dagger}\nbigm \end{multline} \begin{lem} The following diagram is commutative: \begin{equation} \begin{CD} \lambda^{d_X}\ftilde_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \Bigr)[d_X+1] @>{c_1}>> \lambda^{d_Y}\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}[d_Y+1] \\ @VVV @V{=}VV \\ \lambda^{d_X}f_{\dagger}\Bigl( \nbigo_{\nbigx}\otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\bigl( \nbigrtilde_Y\otimes\Omegatilde_{\nbigy}^{-1} \bigr) \Bigr)[d_X+1] @>{c_2}>> \lambda^{d_Y}\nbigrtilde_Y\otimes \Omegatilde_{\nbigy}^{-1}[d_Y+1] \end{CD} \end{equation} Here, $c_1$ and $c_2$ are given by {\rm()} and {\rm()}, and the left vertical arrow is given in Proposition {\rm}. \end{lem} \pf We can regard $\nbigm'_{Y0}$ as an $\nbiga_{Y}$-module by $\ell\otimes\ell$ and $r_1$ with the action of $\lambda^2\del_{\lambda}$. We can also regard $\nbigm'_{Y1}$ as an $\nbiga_{Y}$-module by $r\otimes\ell$ and $\ell_1$ with the action of $\lambda^2\del_{\lambda}$. As mentioned in Example , we have the following isomorphisms: \[ \bigl( \nbigr_Y\otimes\Omegabar_Y^{-1} \langle\lambda^2\del_{\lambda}\rangle, \ell\blacktriangledown, r\triangledown \bigr) \simeq \bigl( \nbigrtilde_Y\otimes \lambda^{-2}\Omegatilde_{\nbigy}^{-1}, \ell, r\bigr) \] It induces the following isomorphisms: \begin{equation} \Bigl( \nbigm'_{Y0}\langle\lambda^2\del_{\lambda}\rangle, (\ell\otimes\ell)\blacktriangledown, r_1\triangledown \Bigr) \simeq \Bigl( \lambda^{-2}\nbigmtilde_{Y0}, \ell\otimes\ell, r_1 \Bigr) \end{equation} \begin{equation} \Bigl( \nbigm'_{Y1}\langle\lambda^2\del_{\lambda}\rangle, (r\otimes\ell)\triangledown, \ell_1\blacktriangledown \Bigr) \simeq \Bigl( \lambda^{-2}\nbigmtilde_{Y1}, r\otimes\ell, \ell_1 \Bigr) \end{equation} The isomorphisms are compatible with the $\nbigrtilde_X$-actions $r_2$. We have the following commutative diagram: \[ \begin{CD} \nbigm'_{Y1}\langle\lambda^2\del_{\lambda}\rangle @>>> \lambda^{-2}\nbigmtilde_{Y1} \\ @VVV @VVV\\ \nbigm'_{Y0}\langle\lambda^2\del_{\lambda}\rangle @>>> \lambda^{-2}\nbigmtilde_{Y0} \end{CD} \] Here, the vertical arrows are induced by $\Psi'_1$ and $\Psitilde_1$, and the horizontal arrows are as in () and (). Hence, the following diagram is commutative: \[ \begin{CD} \lambda^{d_X} f_!\Bigl( \distributiontilde^{\bullet}_{\nbigx/\cnum_{\lambda}}[2d_X+2] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigmtilde_{Y0} \Bigr) @>{\atilde_1}>{\simeq}> \lambda^{d_X} f_!\Bigl( \distributiontilde^{\bullet}_{\nbigx/\cnum_{\lambda}}[2d_X+2] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigmtilde_{Y1} \Bigr) \\ @VVV @VVV\\ \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigm'_{Y0} \Bigr) @>{a'_1}>{\simeq}> \lambda^{d_X} f_!\Bigl( \distributionbar^{\bullet}_X[2d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}\nbigm'_{Y1} \Bigr) \end{CD} \] We also have the compatibility of the trace morphism in Lemma . Then, we obtain the claim of Lemma by construction of the morphisms. \hfill\qed \vspace{.1in} The natural transform $\ftilde_{\dagger}\lrarr f_{\dagger}$ in Corollary induces $\DDDtilde_Y f_{\dagger}\lrarr \DDDtilde_Y \ftilde_{\dagger}$. \begin{lem} The composite of {\rm()} and $\DDDtilde_Y f_{\dagger}\lrarr \DDDtilde_Y \ftilde_{\dagger}$ is equal to {\rm()}. \end{lem} \pf We obtain the claim by comparing the constructions () and () and by using Lemma . \hfill\qed \subsubsection{Step 2} We consider the following diagram: \begin{equation} \begin{CD} \ftilde_{\dagger} \DDDtilde_X @>{\alpha_1}>> \ftilde_{\dagger}\DDDtilde_X S_{\lambda}^{\blacktriangledown} @>{\alpha_2}>> f_{\dagger}\DDD_X \\ @V{\beta_1}VV @V{\beta_2}VV @V{\beta_3}VV \\ \DDDtilde_Y f_{\dagger} @>{\gamma_1}>> \DDDtilde_Y f_{\dagger}S_{\lambda}^{\blacktriangledown} @>{\gamma_2}>> \DDD_Y f_{\dagger} \end{CD} \end{equation} The morphisms $\alpha_1$ and $\gamma_1$ are the natural ones. The morphisms $\beta_i$ $(i=1,2)$ are given in \S. The morphism $\beta_3$ is given in \S. The morphism $\alpha_2$ is induced by $\ftilde_{\dagger}\lrarr f_{\dagger}$ in Corollary , and $\DDDtilde_X S_{\lambda}^{\blacktriangledown} \simeq S_{\lambda}^{\triangledown}\DDD_X \lrarr \DDD_X$ in \S. The morphism $\gamma_2$ is induced by $f_{\dagger}S_{\lambda}^{\blacktriangledown} \simeq S_{\lambda}^{\blacktriangledown}f_{\dagger}$ in \S, and $\DDDtilde_Y S_{\lambda}^{\blacktriangledown} \simeq S_{\lambda}^{\triangledown}\DDD_Y \lrarr \DDD_Y$ in \S. The morphism $\beta_3$ is given in \S. Clearly, the left square is commutative. For the proof of Proposition , it is enough to prove the commutativity of the right square. We have the following expression: \begin{equation} \ftilde_{\dagger}\DDDtilde S_{\lambda}^{\blacktriangledown}\nbigm \simeq \lambda^{d_X}Rf_! \Bigl[ \Bigl( \Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigx}} \nrhom_{\nbigrtilde_X}\bigl( S_{\lambda}^{\blacktriangledown}\nbigm, \nbigrtilde_X\otimes\Omegatilde_{\nbigx}^{-1} \bigr)[d_X+1] \Bigr] \end{equation} It is rewritten as follows: \begin{equation} \lambda^{d_X} Rf_!\Bigl[ \Bigl( \Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigx}} S_{\lambda}^{\triangledown} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_{X}^{-1} \bigr)[d_X] \Bigr] \end{equation} We have the natural morphism from () to the following: \begin{multline} \lambda^{d_X}Rf_! \Bigl[ \bigl( \Omegabar^{\bullet}_{X}[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \bigr) \otimes_{\nbigo_{\nbigx}} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_X^{-1} \Bigr)[d_X] \Bigr] \simeq \\ \lambda^{d_X}Rf_!\nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigo_{\nbigx} \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr)[d_X] \end{multline} Indeed, we have \begin{multline} \lambda^{d_X} Rf_!\Bigl[ \Bigl( \Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigx}} S_{\lambda}^{\triangledown} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_{X}^{-1} \bigr)[d_X] \Bigr] \lrarr \\ \lambda^{d_X} Rf_!\Bigl[ \Bigl( \Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigx}} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_{X}^{-1} \bigr)[d_X] \Bigr] \\ \lambda^{d_X}Rf_! \Bigl[ \bigl( \Omegabar^{\bullet}_{X}[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \bigr) \otimes_{\nbigo_{\nbigx}} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_X^{-1} \Bigr)[d_X] \Bigr] \end{multline} We also have \begin{multline} \lambda^{d_X} Rf_!\Bigl[ \Bigl( \Omegatilde_{\nbigx}^{\bullet}[d_X+1] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigrtilde_{Y}\otimes\Omegatilde_{\nbigy}^{-1}) \Bigr) \otimes_{\nbigo_{\nbigx}} S_{\lambda}^{\triangledown} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_{X}^{-1} \bigr)[d_X] \Bigr] \lrarr \\ \lambda^{d_X}Rf_! \Bigl[ \bigl( \Omegabar^{\bullet}_{X}[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \bigr) \otimes_{\nbigo_{\nbigx}} S_{\lambda}^{\triangledown} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_X^{-1} \Bigr)[d_X] \Bigr] \\ \lambda^{d_X}Rf_! \Bigl[ \bigl( \Omegabar^{\bullet}_{X}[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \bigr) \otimes_{\nbigo_{\nbigx}} \nrhom_{\nbigr_X}\Bigl( \nbigm, \nbigr_X\otimes\Omegabar_X^{-1} \Bigr)[d_X] \Bigr] \end{multline} The composite of the morphisms () and () are equal. We have the following morphisms: \begin{multline} Rf_!\nrhom_{\nbigr_X}\Bigl( \nbigm, \lambda^{d_X}\nbigo_{\nbigx}[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr) \lrarr \\ \nrhom_{\nbigr_Y}\Bigl( f_{\dagger}\nbigm, f_{\dagger}\Bigl( \lambda^{d_X}\nbigo_{\nbigx}[d_X] \otimes_{f^{-1}\nbigo_{\nbigy}} f^{-1}(\nbigr_Y\otimes\Omegabar_Y^{-1}) \Bigr) \Bigr) \\ \simeq \nrhom_{\nbigr_Y}\Bigl( f_{\dagger}\nbigm, \lambda^{d_X} f_{\dagger}(\nbigo_{\nbigx})[d_X] \otimes_{\nbigo_{\nbigy}} \bigl( \nbigr_Y\otimes\Omegabar_Y^{-1} \bigr) \Bigr) \\ \lrarr \nrhom_{\nbigr_Y}\Bigl( f_{\dagger}\nbigm, \lambda^{d_Y} \nbigr_Y\otimes\Omegabar_Y^{-1}[d_Y] \Bigr) \lrarr \DDD f_{\dagger}\nbigm \end{multline} As the composite of the above morphisms, we obtain a morphism $G:\ftilde_{\dagger}\DDDtilde S_{\lambda}^{\blacktriangledown} \nbigm \lrarr \DDD f_{\dagger}\nbigm$. By construction, we can check that both $\beta_3\circ\alpha_2$ and $\gamma_2\circ\beta_2$ are equal to $G$. Indeed, $\beta_3\circ\alpha_2$ is induced by (), and $\gamma_2\circ\beta_2$ is induced by (). Thus, we obtain the commutativity of the right square of (), and the proof of Proposition is finished. \hfill\qed \subsection{Preliminary} \subsubsection{A construction} Let $W_t=\cnum[t]\langle\del_t\rangle$ be the Weyl algebra in the variable $t$. We set $W_t^{\loc}:=W_t\otimes_{\cnum[t]}\cnum[t,t^{-1}]$. Let $X$ be any smooth algebraic variety. Let $p:\cnum_t\times X\lrarr X$ be the projection. In the algebraic setting, we naturally identify $p_{\ast}\nbigd_{\cnum_t\times X}$ with $\nbigd_X\otimes W_t$. We naturally identify algebraic $\nbigd_{\cnum_t\times X}$-modules and algebraic $\nbigd_X\otimes W_t$-modules. We also naturally identify algebraic $\nbigd_{\cnum_t^{\ast}\times X}$-modules and algebraic $\nbigd_X\otimes W^{\loc}_t$-modules. We regard $\cnum_{\tau}$ as the dual space of $\cnum_t$ by the pairing $(\tau,t)\longmapsto\tau t$. Let $M$ be an algebraic $\nbigd_{\cnum_t\times X}$-module. We have its partial Fourier-Laplace transform $\FL_X(M)$ on $\cnum_{\tau}\times X$. As a transform from $\nbigd_{X}\otimes W_t$-modules to $\nbigd_{X}\otimes W_{\tau}$-modules, it is described as the quotient of $\del_t-\tau: M[\tau]\lrarr M[\tau]$. Here, the action of $\nbigd_X\otimes W_{\tau}$ on $M[\tau]$ is given as follows. We have $v(m\tau^j)=(vm)\tau^j$ for a vector field $v$ on $X$, and $\del_{\tau}(m\tau^j)= jm\tau^{j-1} -tm\tau^{j}$. By taking the restriction, we have the localized partial Fourier transform $\FL^{\loc}_X(M)$ on $\cnum_{\tau}^{\ast}\times X$. We set $\lambda:=\tau^{-1}$. We may naturally regard $\FL_X^{\loc}(M)$ as a $\nbigd_{X}\otimes W_{\lambda}^{\loc}$-module. Let us explicit it more explicitly. We have the action of $\nbigd_{\cnum_{\lambda}^{\ast}\times X}$ on $M[\lambda,\lambda^{-1}]$ given as follows. We have $v(m\lambda^j)=(vm)\lambda^j$ for a vector field $v$ on $X$, and $\del_{\lambda}(m\lambda^j)= tm\lambda^{j-2}+jm\lambda^{j-1}$. Then, the $\nbigd_{\cnum_{\lambda}^{\ast}\times X}$-module $\FL_X^{\loc}(M)$ is described as the quotient of \[ M[\lambda,\lambda^{-1}] \stackrel{\del_t-\lambda^{-1}}{\lrarr} M[\lambda,\lambda^{-1}]. \] The natural inclusion $M\lrarr M[\lambda,\lambda^{-1}]$ induces a $\nbigd_X$-homomorphism $\loc:M\lrarr \FL^{\loc}_X(M)$. Let $(M,F)$ be an algebraic filtered $\nbigd_{\cnum_t\times X}$-module, i.e., $M$ is an algebraic $\nbigd_X$-module, and $F$ is a good filtration of $M$. Following , we set \[ G_0\FL_X^{\loc}(M,F):= \sum_{j\in\seisuu} \lambda^j\loc(F_jM) \subset \FL_X^{\loc}(M) \] on $X$. It is naturally an algebraic $\nbigr_X$-module. It is also equipped with the action of $\lambda^{2}\del_{\lambda}$. Thus, we obtain an algebraic $\nbigrtilde_X$-module. Because $\del_{t}F_j(M)\subset F_{j+1}(M)$, we have $\loc(F_jM)\subset \lambda\loc(F_{j+1}M)$, and hence $G_0\FL_X^{\loc}(M,F) =\sum_{j\geq N} \lambda^j\loc(F_jM)$ for any $N$. We have the Rees module $R(M,F)=\sum F_jM\lambda^j$. We have the $\nbigr_X$-homomorphism $R(M,F)\rarr \FL_X^{\loc}(M,F)$, and the image is $G_0\FL_X^{\loc}(M,F)$. The quotient of the following is denoted by $G_0'\FL_X^{\loc}(M,F)$: \[ \lambda\,R(M,F) \stackrel{\del_t-\lambda^{-1}}\lrarr R(M,F) \] It is naturally an algebraic $\nbigrtilde_X$-module. \begin{lem} The image of the induced morphism $G'_0\FL_X^{\loc}(M,F) \lrarr \FL_X^{\loc}(M)$ is $G_0\FL_X^{\loc}(M,F)$. If the multiplication of $\lambda$ on $G_0'\FL_X^{\loc}(M,F)$ is injective, the natural morphism $G_0'\FL_X^{\loc}(M,F) \lrarr G_0\FL_X^{\loc}(M,F)$ is an isomorphism of $\nbigrtilde_X$-modules. \end{lem} \pf The first claim is clear by the construction. The morphism $G_0'\FL_X^{\loc}(M,F) \lrarr G_0\FL_X^{\loc}(M,F)$ is an isomorphism after the localization with respect to $\lambda$. Then, the second claim follows. \hfill\qed \vspace{.1in} We may interpret the construction $G_0'\FL^{\loc}_X$ in terms of $\nbigr$-modules. Let $p$ denote the projection of $\cnum_t\times X$ onto $X$, as above. We have the algebraic $\nbigr_{\cnum_t\times X}$-module $\nbigl(-t)=\bigl( \nbigo_{\cnum_{\lambda}\times \cnum_t\times X},\, d-d(\lambda^{-1}t) \bigr)$. Then, by definition, $G_0'\FL^{\loc}_X(M,F)$ is isomorphic to \[ \lambda\cdot p^0_{\dagger} \bigl( R(M,F)\otimes\nbigl(-t) \bigr) \simeq \lambda\cdot \hyperr^1p_{\ast}\Bigl( R(M,F)\otimes\nbigl(-t) \stackrel{b}{\lrarr} R(M,F)\otimes\nbigl(-t) \otimes \bigl( \lambda^{-1} \Omega^{1}_{\cnum_{\lambda}\times X\times\cnum_t /\cnum_{\lambda}\times X} \bigr) \Bigr) \] Here, $b$ is induced by the meromorphic flat connection of $R(M,F)\otimes\nbigl(-t)$. (See \S for example.) \begin{prop} Suppose that $(M,F)$ is a filtered $\nbigd_{\cnum_t\times X}$-module underlying an algebraic mixed Hodge module. Then, the natural morphism $G_0'\FL^{\loc}_X(M,F) \lrarr G_0\FL^{\loc}_X(M,F)$ is an isomorphism. \end{prop} \pf By the assumption, we have a mixed Hodge module on $\proj^1_t\times X$ whose underlying filtered $\nbigd$-module $(M',F')$ on $\proj^1_t\times X$ satisfies $(M',F')_{|\cnum_t\times X}=(M,F)$. As explained in \cite[\S13.5]{Mochizuki-MTM}, the analytification of $R(M',F')$ underlies an integrable mixed twistor $\nbigd$-module $(\nbigt,W)$ on $\proj^1_t\times X$. Let $\nbigv(-t)=\bigl(\nbigl(-t),\nbigl(-t),C(-t)\bigr)$ denote the $\nbigr_{\proj^1_t\times X}$-triple as in \S. As explained in \cite[\S11.3.2]{Mochizuki-MTM}, we have a mixed twistor $\nbigd$-module $\bigl((\nbigt,W)\otimes\nbigv(-t)\bigr)[\ast (t)_{\infty}]$ on $\proj^1_t\times X$. \begin{lem} The underlying $\nbigr_{\proj^1_t\times X}$-module of $\bigl((\nbigt,W)\otimes\nbigv(-t)\bigr)[\ast (t)_{\infty}]$ is the analytification of $R(M,F)\otimes \nbigl(-t)$. \end{lem} \pf The claim is clear on $\cnum_t\times X$. Let $Q$ be any point of $X$. It is enough to check the claim locally around $(\infty,Q)$. Let $\nbigm$ and $\nbigmtilde$ be the $\nbigr_{\proj^1_t\times X}$-modules underlying $\nbigt$ and $\bigl((\nbigt,W)\otimes\nbigv(-t)\bigr)[\ast (t)_{\infty}]$, respectively. We have $\nbigmtilde(\ast (t)_{\infty}) =\nbigm\otimes\nbigl(-t)$. Let $\kappa:=t^{-1}$. We consider $V\nbigr_{\cnum_{\kappa}\times X} \subset \nbigr_{\cnum_{\kappa}\times X}$ which is generated by $\lambda \Theta_{\cnum_{\kappa}\times X}(\log \kappa)$ over $\nbigo_{\cnum_{\lambda}\times\cnum_{\kappa}\times X}$. We have the $V$-filtration $V(\nbigm)$ of $\nbigm$ along $\kappa$. Then, $V_{c}(\nbigm)$ is $V\nbigr_{\cnum_{\kappa}\times X}$-coherent for any $c\in\real$. Let $q:\cnum_{\kappa}\times X\lrarr X$ be the projection. Note that $V_{c}(\nbigm)$ are $q^{\ast}\nbigr_X$-coherent because $\nbigm$ is regular. We take sections $s_1,\ldots,s_m$ of $V_c(\nbigm)$ which generates $V_c(\nbigm)$ over $q^{\ast}\nbigr_X$. Let $\upsilon$ be the generator of $\nbigl(-t)$ such that $\nabla\upsilon=\upsilon d(-\lambda^{-1}t)$. Let $\nbigm_1$ be the $\nbigr_{\cnum_{\kappa}\times X}$-submodule of $\nbigm\otimes\nbigl(-t)$ generated by $s_i\otimes\upsilon$ $(i=1,\ldots,m)$. Let us observe $\nbigm_1=\nbigm\otimes\nbigl(-t)$. It is enough to prove that $\kappa^{-n}(V_c\nbigm\otimes \upsilon) \subset \nbigm_1$ for any $n$. We have $V_c\nbigm\otimes \upsilon \subset \nbigm_1$ because $V_c(\nbigm)$ is generated by $s_1,\ldots,s_m$ over $q^{\ast}\nbigr_X$. Suppose $\kappa^{-n}(V_c\nbigm\otimes \upsilon) \subset \nbigm_1$. Let $f$ be any section of $V_c\nbigm$. We have \[ \kappa\del_{\kappa} \bigl( \kappa^{-n}f\otimes\upsilon \bigr) =-n\kappa^{-n}f\otimes\upsilon +\kappa^{-n}(\kappa\del_{\kappa}f)\otimes\upsilon -\kappa^{-n-1}f\otimes\upsilon. \] Hence, we obtain $\kappa^{-n-1}f\otimes\upsilon \in\nbigm_1$. Suppose $c<0$. For a sufficiently large $N$, $t^Ns_i\otimes \upsilon$ $(i=1,\ldots,m)$ are sections of $\nbigmtilde$. We also have that $t^Ns_i$ generates $V_{c-N}\nbigm$. Hence, we obtain that $t^Ns_i\otimes \upsilon$ $(i=1,\ldots,m)$ generates $\nbigm\otimes\nbigl(-t)$ over $\nbigr_X$. It implies that $\nbigmtilde=\nbigm\otimes\nbigl(-t)$. \hfill\qed \vspace{.1in} Hence, the analytification of $p^0_{\dagger} \bigl( R(M,F)\otimes\nbigl(-t) \bigr)$ underlies a mixed twistor $\nbigd$-module. It implies that the multiplication of $\lambda$ on $p^0_{\dagger} \bigl( R(M,F)\otimes\nbigl(-t) \bigr)$ is injective. Then, the claim follows from the Lemma . \hfill\qed \subsubsection{Comparison} Let $Y$ be a smooth projective variety. Let $X$ be any smooth algebraic variety. Let $\pi_1:Y\times \cnum_{t}\times X \lrarr \cnum_{t}\times X$ and $\pi_2:Y\times\cnum_{t}\times X \lrarr X$ be the projections. Let $(M,F)$ be a filtered $\nbigd$-module on $Y\times\cnum_{t}\times X$ which underlies a mixed Hodge module. We obtain filtered $\nbigd$-modules $\pi^j_{1\dagger}(M,F)$ on $\cnum_{t}\times X$, which underlie mixed Hodge modules. Applying the construction in \S, we obtain $\nbigr_X$-modules $G_0\FL^{\loc}_X\pi^j_{1\dagger}(M,F)$. \vspace{.1in} Let $R(M,F)$ be the Rees module of $(M,F)$. It is naturally an algebraic $\nbigrtilde_{Y\times\cnum_{t}\times X}$-module. We have the $\nbigr_{Y\times\cnum_t\times X}$-module $\nbigl(-t)$. We obtain an algebraic $\nbigrtilde_{Y\times\cnum_t\times X}$-module $R(M,F)\otimes \nbigl(-t)$. It underlies a mixed twistor $\nbigd$-module (Lemma ). We obtain $\nbigr_{X}$-modules $\pi^j_{2\dagger}\bigl( R(M,F)\otimes\nbigl(-t) \bigr)$. \begin{prop} We have natural isomorphisms $\lambda\cdot\pi^j_{2\dagger}\bigl( R(M,F)\otimes\nbigl(-t) \bigr) \simeq G_0\FL^{\loc}_X\pi^j_{1\dagger}(M,F)$. \end{prop} \pf Note that we have $\pi_{1\dagger}^j\bigl(R(M,F)\bigr) \simeq R\bigl(\pi_{1\dagger}^j(M,F)\bigr)$ by the theory of mixed Hodge modules. Let $p_X:\cnum_t\times X\lrarr X$ denote the projection. We have $p_X\circ\pi_1=\pi_2$. Let $L(-t)$ denote the $\nbigd$-module associated to $-t$ on $\proj^1_t\times X$ as in \S. Let us recall the following lemma. \begin{lem} For regular holonomic $\nbigd$-modules $M$ on $\cnum_t\times X$, we have $p_{X\dagger}^jM\otimes L(-t)=0$ for $j\neq 0$. \end{lem} \pf It is enough to consider the case $j=-1$. Let $s$ be any section of $M$ such that $\del_ts-s=0$, and let us check that $s=0$. We have the $V$-filtration of $V_{\bullet}M$ along $u=t^{-1}$. Suppose $s\neq 0$ around $(\infty,P)\in\proj^1\times X$. Because $M$ is regular along $u$, we have $a_0:=\min\{a\,|\,s\in V_{a}M\}$. Because $\del_ts\in V_{<a_0}$, we also have $s\in V_{<a_0}$ which contradicts with the choice of $a_0$. Hence, we have $s=0$ around $(\infty,P)$. Let $M_1\subset M$ be the $\nbigd$-submodule generated by $s$. We have $\Supp(M_1)\cap \{\infty\}\times X=\emptyset$. Let $P\in X$. By shrinking $X$ appropriately, we take an algebraic function $f$ on $X$ such that $f(P)=0$ and $df$ is nowhere vanishing. We have the $V$-filtrations $V_f$ of $M_1$ and $M_1\otimes L(-t)$ along $f$. It is easy to see that $\Gr^{V_f}(M\otimes L(-t)) \simeq \Gr^{V_f}(M)\otimes L(-t)$. If $s$ is non-zero around $(t,P)$, there exists $b$ such that $s$ gives a non-zero section $[s]$ of $\Gr_b^{V_f}(M)$ satisfying $\del_t[s]-[s]=0$. Hence, we can reduce the claim to the case $\dim X=0$, where it can be checked easily. Thus, we obtain Lemma . \hfill\qed \vspace{.1in} By the lemma, we have $p^0_{X\dagger} \pi_{1\dagger}^j\bigl( R(M,F)\otimes \nbigl(-t) \bigr) \simeq \pi_{2\dagger}^j\bigl( R(M,F)\otimes\nbigl(-t) \bigr)$. Then, the claim of Proposition follows from Proposition . \hfill\qed \subsection{$\nbigrtilde$-modules associated to subvarieties of $\proj^m$} \subsubsection{Setting} We fix a homogeneous coordinate system $[z_0:\cdots:z_m]$ of $\proj^m$. We set $V:=H^0\bigl(\proj^m,\nbigo(1)\bigr)$. We put $V_1:=\{\alpha_0 z_0\,|\,\alpha_0\in\cnum\}\subset V$ and $V_2:=\bigl\{\sum_{i=1}^m\alpha_iz_i\,\big|\, \alpha_i\in\cnum \bigr\} \subset V$. We may regard $\alpha_0$ and $(\alpha_1,\ldots,\alpha_m)$ as coordinate systems of $V_1$ and $V_2$, respectively. We set $H_0:=\{z_0=0\}$. Let $U$ be a smooth quasi-projective variety with an immersion $\iota_U:U\lrarr\proj^m\setminus H_0$. For simplicity, we assume that there exists a hypersurface $H\subset\proj^m$ for which $\iota_U(U)$ is a closed subset of $\proj^m\setminus(H_0\cup H)$. We do not assume that $H$ is smooth. We set $V_2^{\ast}:=V_2\setminus\{0\}$. We shall construct some $\nbigrtilde$-modules on $V_2^{\ast}$ associated to $U$. \subsubsection{A construction} We take a smooth projective variety $Y$ with an open immersion $\iota_1:U\lrarr Y$ and a morphism $\iota_2:Y\lrarr \proj^m$ such that (i) $D_Y:=Y\setminus U$ is normal crossing, (ii) $\iota_2\circ\iota_1=\iota_U$. We have $D_Y=\iota_{2}^{\ast}(H\cup H_0)$. We set $D_{Y,V_2^{\ast}}:=D_Y\times V_2^{\ast}$. We set $\iota_{2,V_2^{\ast}}:=\iota_2\times \id_{V_2^{\ast}}: Y\times V_2^{\ast}\lrarr \proj^m\times V_2^{\ast}$. We have the meromorphic function \[ F^u_{Y,V_2^{\ast}}:= \iota_{2,V_2^{\ast}}^{\ast} \Bigl( \sum_{i=1}^m \alpha_iz_i/z_0\Bigr) \] on $(Y,D_Y)\times V_2^{\ast}$. We obtain the integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$ $(\star=\ast,!)$ and the underlying $\nbigrtilde_{Y\times V_2^{\ast}}$-modules $\nbigl_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$. Let $\pi_{Y,V_2^{\ast}}:Y\times V_2^{\ast}\lrarr V_2^{\ast}$ be the projection. We obtain the integrable mixed twistor $\nbigd$-modules $\pi_{Y,V_2^{\ast}\dagger}^0 \nbigt_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$ and the underlying $\nbigrtilde_{V_2^{\ast}}$-modules $\pi_{Y,V_2^{\ast}\dagger}^0 \nbigl_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$. \vspace{.1in} If $Y'$ and $\iota_{i}'$ be another choice, then we take $Y''$ with morphisms $\iota_1'':U\lrarr Y''$, $a:Y''\lrarr Y'$ and $b:Y''\lrarr Y$ such that $a\circ\iota_1''=\iota'_1$, $b\circ\iota_1''=\iota_1$, $\iota_2'\circ a=\iota_2\circ b=:\iota_{2}''$. Then, we have natural commutative diagram of isomorphisms: \[ \begin{array}{ccccc} \pi^0_{Y,V_2^{\ast}\dagger} \nbigt_{!}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) & \simeq & \pi^0_{Y'',V_2^{\ast}\dagger} \nbigt_{!}(F^u_{Y'',V_2^{\ast}},D_{Y'',V_2^{\ast}}) & \simeq & \pi^0_{Y',V_2^{\ast}\dagger} \nbigt_{!}(F^u_{Y',V_2^{\ast}},D_{Y',V_2^{\ast}})\\ \darr & & \darr & & \darr \\ \pi^0_{Y,V_2^{\ast}\dagger} \nbigt_{\ast}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) & \simeq & \pi^0_{Y'',V_2^{\ast}\dagger} \nbigt_{\ast}(F^u_{Y'',V_2^{\ast}},D_{Y'',V_2^{\ast}}) & \simeq & \pi^0_{Y',V_2^{\ast}\dagger} \nbigt_{\ast}(F^u_{Y',V_2^{\ast}},D_{Y',V_2^{\ast}})\\ \end{array} \] In this sense, $\pi^0_{Y,V_2^{\ast}\dagger} \nbigt_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$ and the underlying $\nbigrtilde_{V_2^{\ast}}$-modules $\pi^0_{Y,V_2^{\ast}\dagger} \nbigl_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$ are independent of the choice of $Y$. We shall observe that $\pi^0_{Y,V_2^{\ast}\dagger} \nbigl_{\star}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$ are isomorphic to the $\nbigrtilde_{V_2^{\ast}}$-modules given in . \subsubsection{Comparison with a construction of Reichelt-Sevenheck} Let $Z\subset\proj^m\times V$ be the $0$-set of the universal section $\sum \alpha_iz_i$ of $\nbigo_{\proj^m}(1)\boxtimes\nbigo_V$. The projections $q_1:Z\lrarr \proj^m$ and $q_2:Z\lrarr V$ are smooth. We set $Z_U:=Z\times_{\proj^m}U$. Let $\iota_{Z_U}:Z_U\lrarr \proj^m\times V$ denote the natural inclusion. Let $\pi_V:\proj^m\times V\lrarr V$ be the projection. We have the variation of pure Hodge structure $(\nbigo_{Z_U,}F)$ with the canonical real structure, where the Hodge filtration $F$ is given by $F_0=\nbigo_{Z_U}$ and $F_{-1}=0$. We have the associated pure Hodge module which is also denoted by $(\nbigo_{Z_U},F)$. We obtain the mixed Hodge modules $\iota_{Z_U\star}(\nbigo_{Z_U},F)$ $(\star=\ast,!)$ on $\proj^m\times V$, and then $\pi^0_{V\dagger}\iota_{Z_U\star}(\nbigo_{Z_U},F)$ on $V=V_1\times V_2$. By applying the procedure in \S, and by taking the restriction to $V_2^{\ast}\subset V_2$, we obtain the following $\nbigrtilde_{V_2^{\ast}}$-modules: \[ G_0\FL^{\loc}_{V_2} \pi^0_{V\dagger}\iota_{Z_U\star}(\nbigo_{Z_U},F)_{|V_2^{\ast}} \] \begin{prop} We have isomorphisms of $\nbigrtilde_X$-modules: \[ \lambda\cdot \pi^0_{Y,V_2^{\ast}\dagger} \nbigl_{\star}(F_Y^u,D_{Y,V_2^{\ast}}) \simeq G_0\FL^{\loc}_{V_2} \pi^0_{V\dagger}\iota_{Z_U\star}(\nbigo_{Z_U},F)_{|V_2^{\ast}} \] We also have the following commutative diagram: \[ \begin{CD} \lambda\cdot \pi^0_{Y,V_2^{\ast}\dagger} \nbigl_{!}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) @>{\simeq}>> G_0\FL^{\loc}_{V_2} \pi^0_{V\dagger}\iota_{Z_U!}(\nbigo_{Z_U},F)_{|V_2^{\ast}} \\ @VVV @VVV \\ \lambda\cdot \pi^0_{Y,V_2^{\ast}\dagger} \nbigl_{\ast}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) @>{\simeq}>> G_0\FL^{\loc}_{V_2} \pi^0_{V\dagger}\iota_{Z_U\ast}(\nbigo_{Z_U},F)_{|V_2^{\ast}} \end{CD} \] \end{prop} \pf We set $Z_0:= Z\cap\bigl(\proj^m\times V_1\times V_2^{\ast} \bigr)$. Let $\proj_{V_1}$ denote the projective completion of $V_1$, i.e., $\proj_{V_1}=\proj(V_1^{\lor}\oplus\cnum)$. Let $\Zbar_0$ be the closure of $Z_0$ in $\proj^m\times\proj_{V_1}\times V_2^{\ast}$. Let $\iota_{\Zbar_0}: \Zbar_0\lrarr \proj^m\times\proj_{V_1}\times V_2^{\ast}$ denote the natural inclusion. By the construction of $\Zbar_0$, we have the following equality of meromorphic functions on $\Zbar_0$: \[ \iota_{\Zbar_0}^{\ast}\alpha_0 =-\iota_{\Zbar_0}^{\ast}\Bigl( \sum_{i=1}^m\alpha_iz_i/z_0 \Bigr) \] Let $Z_{0U}:=Z_U\cap Z_0$, and let $\Zbar_{0U}$ denote the closure of $Z_{0U}$ in $\Zbar_0$. Let $q:\Zbar_{0U}\lrarr \proj^m\times V_2^{\ast}$ denote the naturally induced morphism. Note that $Z_{0U}$ is naturally isomorphic to $U\times V_2^{\ast}$, and that the restriction $q_{|Z_{0U}}$ is an immersion. We can take a smooth complex algebraic variety $B$ with projective morphisms $\varphi_1:B\lrarr \Zbar_{0U}$ and $\varphi_2:B\lrarr Y\times V_2^{\ast}$, and an open immersion $j:U\times V_2^{\ast}\subset B$ such that (i) $q\circ\varphi_1 =\iota_{Y,V_2^{\ast}}\circ\varphi_2$, (ii) $\varphi_a\circ j$ $(a=1,2)$ are the identity of $U\times V_2^{\ast}$, (iii) $D_B:=B\setminus j(U\times V_2^{\ast})$ is a normal crossing hypersurface. We set $G:=-\varphi_1^{\ast}\alpha_0 =\varphi_2^{\ast}F_Y^u$. We have the integrable mixed twistor $\nbigd$-modules $\nbigt_{\star}(G,D_B)$ on $B$. Let $\nbigl_{\star}(G,D_B)$ denote the underlying $\nbigrtilde_B$-modules. Then, we have the following natural isomorphisms and the commutative diagrams: \[ \begin{CD} \varphi_{2\dagger}\nbigl_{!}(G,D_B) @>{\simeq}>> \nbigl_{!}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})\\ @VVV @VVV \\ \varphi_{2\dagger}\nbigl_{\ast}(G,D_B) @>{\simeq}>> \nbigl_{\ast}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) \end{CD} \] We also have the following: \[ \begin{CD} \varphi_{1\dagger}\nbigl_{!}(G,D_B) @>{\simeq}>> R\bigl( \iota_{Z_U!}(\nbigo_{Z_U},F) \bigr) \otimes\nbigl(-\alpha_0) \\ @VVV @VVV \\ \varphi_{1\dagger}\nbigl_{\ast}(G,D_B) @>{\simeq}>> R\bigl( \iota_{Z_U\ast}(\nbigo_{Z_U},F) \bigr) \otimes\nbigl(-\alpha_0) \end{CD} \] Then, the claim of the theorem follows from Proposition . \hfill\qed \vspace{.1in} Reichelt and Sevenheck considered the image $(M,F)$ of $\iota_{Z_U!}(\nbigo_{Z_U},F) \lrarr \iota_{Z_U\ast}(\nbigo_{Z_U},F)$, that is the minimal extension of $Z_U$ in $\proj^m\times V$. They proved that $G_0\FL^{\loc}_{V_2}\bigl( \pi^0_{V\dagger}(M,F) \bigr)_{|V_2^{\ast}}$ is isomorphic to the image of the natural morphism $G_0\FL^{\loc}_{V_2} \pi^0_{V\dagger}\iota_{Z_U!} \bigl(\nbigo_{Z_U},F \bigr)_{|V_2^{\ast}} \lrarr G_0\FL^{\loc}_{V_2} \pi^0_{V\dagger}\iota_{Z_U\ast} \bigl( \nbigo_{Z_U},F \bigr)_{|V_2^{\ast}}$. \begin{cor} $G_0\FL^{\loc}_{V_2}\bigl( \pi_{V\dagger}(M,F) \bigr)_{|V_2^{\ast}}$ is isomorphic to the image of the natural morphism \[ \lambda\cdot \pi_{Y,V_2^{\ast}\dagger}^0 \nbigl_{!}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) \lrarr \lambda\cdot \pi_{Y,V_2^{\ast}\dagger}^0 \nbigl_{\ast}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}). \] In particular, $G_0\FL^{\loc}_{V_2}\bigl( \pi_{V\dagger}(M,F) \bigr)_{|V_2^{\ast}}$ underlies a polarizable integrable pure twistor $\nbigd$-module. \end{cor} \pf The first claim follows from Proposition and the result of Reichelt and Sevenheck mentioned above. Because the morphism of the mixed twistor $\nbigd$-modules $\pi_{Y,V_2^{\ast}\dagger}^0 \nbigt_{!}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) \lrarr \pi_{Y,V_2^{\ast}\dagger}^0 \nbigt_{\ast}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}})$ factors through the morphism of the pure twistor $\nbigd$-modules \[ \Gr^W_{\dim U+m}\pi_{Y,V_2^{\ast}\dagger}^0 \nbigt_{!}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}) \lrarr \Gr^W_{\dim U+m}\pi_{Y,V_2^{\ast}\dagger}^0 \nbigt_{\ast}(F^u_{Y,V_2^{\ast}},D_{Y,V_2^{\ast}}), \] the second claim follows. \hfill\qed \subsection{$\nbigd$-modules} \subsubsection{The $\nbigd$-modules associated to some better behaved GKZ-systems} We recall a special version of better behaved GKZ-systems introduced by L. Borisov and P. Horja . Let $\nbiga=\{\veca_1,\ldots,\veca_m\}\subset \seisuu^n$ be a finite subset generating $\seisuu^n$. Let $K_{\real}(\nbiga)\subset\real^n$ denote the cone $\bigl\{ \sum_{j=1}^m r_j\veca_j\,\big|\, r_j\geq 0 \bigr\}$ generated by $\nbiga$. We set $K(\nbiga):=K_{\real}(\nbiga)\cap\seisuu^n$. The semigroup $K(\nbiga)$ is the saturation of $\seisuu_{\geq 0}\nbiga =\bigl\{ \sum_{j=1}^m n_{j}\veca_j\,\big|\, n_{j}\in\seisuu_{\geq 0} \bigr\}$. We denote $\veca_j=(a_{j1},\ldots,a_{jn})$. Let $\Gamma\subset K(\nbiga)$ be any subset such that $\Gamma+\veca\subset \Gamma$ for any $\veca\in\nbiga$. Let $\vecbeta\in\cnum^n$. The following system of differential equations $\GKZ(\nbiga,\Gamma,\vecbeta)$ for tuples of holomorphic functions $(\Phi_{\vecc}\,|\,\vecc\in \Gamma)$ on any open subset $\cnum^m$ is called the better behaved GKZ-hypergeometric system associated to $(\nbiga,\Gamma,\vecbeta)$: \[ \del_{x_j}\Phi_{\vecc}=\Phi_{\vecc+\veca_j} \quad(\forall \vecc\in \Gamma,\,\forall j=1,\ldots,m) \] \[ \Bigl( \sum_{j}a_{ji}x_j\del_{x_j}+c_i-\beta_i \Bigr) \Phi_{\vecc}=0 \quad (\forall \vecc\in \Gamma,\,\forall i=1,\ldots,n) \] Here, $(x_1,\ldots,x_m)$ denotes the standard coordinate system of $\cnum^m$. Let us describe the $\nbigd_{\cnum^m}$-module $M^{\GKZ}(\nbiga,\Gamma,\vecbeta)$ corresponding to the system $\GKZ(\nbiga,\Gamma,\vecbeta)$. We introduce a free $\nbigo_{\cnum^m}$-module $Q(\nbiga,\Gamma)$ generated by $\Gamma$. Let $e(\vecc)$ denote the element corresponding to $\vecc\in\Gamma$. So, we have $Q(\nbiga,\Gamma)= \bigoplus_{\vecc\in\Gamma} \nbigo_{\cnum^m}\cdot e(\vecc)$. We introduce the action of $\nbigd_{\cnum^m}$ on $Q(\nbiga,\Gamma)$ by $\del_{x_j}e(\vecc)=e(\vecc+\veca_j)$. Let $J(\nbiga,\Gamma,\vecbeta)$ denote the $\nbigd_{\cnum^m}$-submodule of $Q(\nbiga,\Gamma)$ generated by $\bigl( \sum_{j}a_{ji}x_j\del_{x_j}+c_i-\beta_i \bigr)e(\vecc)$ for $\vecc\in \Gamma$ and $i=1,\ldots,n$. We set $M^{\GKZ}(\nbiga,\Gamma,\vecbeta):= Q(\nbiga,\Gamma)/J(\nbiga,\Gamma,\vecbeta)$. Let $U\subset \cnum^m$ be any open subset. An $\nbigo$-homomorphism $\Phi:M^{\GKZ}(\nbiga,\Gamma,\vecbeta)_{|U} \lrarr \nbigo_U$ is uniquely determined by $\Phi_{\vecc}:=\Phi(e(\vecc))$ $(\vecc\in \Gamma)$. Conversely, any tuple of holomorphic functions $(\Phi_{\vecc}\,|\,\vecc\in\Gamma)$ determines an $\nbigo$-morphism $M^{\GKZ}(\nbiga,\Gamma,\vecbeta)_{|U} \lrarr \nbigo_U$. Then, the following lemma is clear by the construction. \begin{lem} Let $U\subset \cnum^m$ be any open subset. The above correspondence induces a bijective correspondence between morphisms of $\nbigd$-modules $M^{\GKZ}(\nbiga,\Gamma,\vecbeta)_{|U} \lrarr \nbigo_U$ and solutions of $\GKZ(\nbiga,\Gamma,\vecbeta)$. \hfill\qed \end{lem} \begin{rem} In {\rm}, Borisov and Horja considered an $m$-tuple of a finitely generated abelian group, instead of a finite subset of $\seisuu^n$. In that sense, we consider the only special case. However, we do not impose the existence of an element $\vecalpha\in(\seisuu^n)^{\lor}$ such that $\vecalpha(\veca_j)=1$ $(j=1,\ldots,m)$. \hfill\qed \end{rem} \subsubsection{Twisted de Rham complexes} We recall some notation used in \S. Let $T^n:=(\cnum^{\ast})^n$. We consider the morphism $\psi^{\aff}_{\nbiga}:T^n\lrarr \cnum^m$ given by $\psi^{\aff}_{\nbiga}(t_1,\ldots,t_n) =(t^{\veca_1},\ldots,t^{\veca_m})$, where $t^{\veca_j}=\prod_{i=1}^nt_i^{a_{ji}}$. Let $X_{\nbiga}^{\aff}$ denote the closure of the image of $\psi^{\aff}_{\nbiga}$. Let $\check{X}^{\aff}_{\nbiga}\lrarr X^{\aff}_{\nbiga}$ be the normalization. Let $\check{D}^{\aff}_{\nbiga}$ denote the complement of $T^n$ in $\check{X}^{\aff}_{\nbiga}$. Note that $\check{X}^{\aff}_{\nbiga} =\Spec \cnum[K(\nbiga)]$ and $X^{\aff}_{\nbiga} =\Spec\cnum[\seisuu_{\geq 0}\nbiga]$. The family of Laurent polynomials $\sum_{j=1}^m x_jt^{\veca_j}$ induces a meromorphic function $F_{\nbiga}$ on $(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga})\times\cnum^m$. We also have the logarithmic closed one form $\kappa(\vecbeta)=\sum_{i=1}^n \beta_i dt_i/t_i$. We obtain the relative algebraic de Rham complexes: \[ \nbigc^{\bullet}(\nbiga,\vecbeta)_{\bullet}:= \Bigl( \Omega^{\bullet}_{(\check{X}^{\aff}_{\nbiga}\times \cnum^m)/\cnum^m} (\log\check{D}^{\aff}_{\nbiga}\times\cnum^m), d+dF_{\nbiga}-\kappa(\vecbeta) \Bigr) \] \[ \nbigc^{\bullet}(\nbiga,\vecbeta)_{\circ}:= \Bigl( \Omega^{\bullet}_{(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga}) \times \cnum^m/\cnum^m}, d+dF_{\nbiga}-\kappa(\vecbeta) \Bigr) \] Let $\check{\pi}_{\nbiga}: \check{X}^{\aff}_{\nbiga}\times\cnum^m \lrarr \cnum^m$ denote the projection. Each $\check{\pi}_{\nbiga\ast}\nbigc^{k}(\nbiga,\vecbeta)_{\star}$ is naturally a $\nbigd_{\cnum^m}$-module by $\del_{x_j}\bullet(g)=\del_{x_j}g+\del_{x_j}F_{\nbiga}\cdot g$. The differential $d+dF_{\nbiga}-\kappa(\vecbeta)$ of the complexes are compatible with the actions of $\nbigd_{\cnum^m}$. So, we obtain the following $\nbigd_{\cnum^m}$-modules \[ M_{\nbiga,\vecbeta,\star}:= \hyperr^n\check{\pi}_{\nbiga\ast} \nbigc^{\bullet}(\nbiga,\vecbeta)_{\star} \quad (\star=\bullet,\circ). \] \begin{rem} In the notation of {\rm \S} (see also Example {\rm}), we have the following commutative diagram: \[ \begin{CD} M_{\nbiga,0,\bullet} @>>> M_{\nbiga,0,\circ} \\ @V{\simeq}VV @V{\simeq}VV \\ \pi_{\dagger}L_{\ast}(\nbiga,\cnum^m,\id) @>>> \pi_{\dagger}L_{!}(\nbiga,\cnum^m,\id) \end{CD} \] \hfill\qed \end{rem} \subsubsection{Comparison} Let $K(\nbiga)^{\circ}$ denote the intersection of $\seisuu^n$ and the interior part of $K_{\real}(\nbiga)$. \begin{prop} We have the following natural isomorphisms of $\nbigd_{\cnum^m}$-modules: \[ M^{\GKZ}(\nbiga,K(\nbiga),\vecbeta) \simeq M_{\nbiga,\vecbeta,\bullet}, \quad\quad M^{\GKZ}(\nbiga,K(\nbiga)^{\circ},\vecbeta) \simeq M_{\nbiga,\vecbeta,\circ}. \] \end{prop} \pf Although the claim is rather obvious by definition, we give some more details. Let $\psi_{\Sigma_1}:X_{\Sigma_1}\lrarr \check{X}^{\aff}_{\nbiga}$ be any toric projective resolution. Set $D_{\Sigma_1}:=X_{\Sigma_1}\setminus T$. According to , we have $R\psi_{\Sigma_1\ast}\Omega^p_{X_{\Sigma_1}}(\log D_{\Sigma_1}) \simeq \Omega^p_{\check{X}^{\aff}_{\nbiga}}(\log \check{D}^{\aff}_{\nbiga})$. Let $\nbigi(p)$ denote the set of tuples of integers $I=(i_1,\ldots,i_p)$ with $1\leq i_1<\cdots <i_p\leq n$. For any $I\in\nbigi(p)$, we set $\tau_I:= \prod_{j=1}^pt_{i_j}^{-1} dt_{i_1}\wedge\cdots\wedge dt_{i_p}$. Hence, we have $\Omega_{\check{X}^{\aff}_{\nbiga}}^p (\log\check{D}^{\aff}_{\nbiga})$ is a free sheaf over $\nbigo_{\check{X}^{\aff}_{\nbiga}}$ with a basis $\tau_I$ $(I\in S(p))$. Hence, we have \[ \check{\pi}_{\nbiga\ast} \nbigc^p(\nbiga)_{\ast} \simeq \bigoplus_{I\in S(p)} \bigoplus_{\vecc\in K(\nbiga)} \nbigo_{\cnum^m} t^{\vecc}\tau_I. \] We set $\omega:=\tau_{\{1,\ldots,n\}}$. We have $\del_{x_j}(t^{\vecc}\omega) =t^{\vecc+\veca_j}\omega$. Hence, we have the isomorphism of $\nbigd_{\cnum^m}$-modules $Q(\nbiga,K(\nbiga))\simeq \check{\pi}_{\nbiga\ast} \nbigc^n(\nbiga)_{\bullet}$ given by $e(\vecc)\longmapsto t^{\vecc}\omega$. Let $\omega_i$ denote the inner product of $\omega$ and $t_i\del/\del t_i$. We have $\omega=(dt_i/t_i)\omega_i$. We have \[ \bigl(d+dF_{\nbiga}-\kappa(\vecbeta)\bigr) (t^{\vecc}\omega_i) =\Bigl( c_i-\beta_i+\sum_{j=1}^na_{ji}x_jt^{\veca_j} \Bigr)(t^{\vecc}\omega) =\Bigl( c_i-\beta_i+\sum a_{ji}x_j\del_{x_j} \Bigr)(t^{\vecc}\omega). \] Hence, the image of $\check{\pi}_{\nbiga\ast} \nbigc^{n-1}(\nbiga)_{\bullet} \lrarr \check{\pi}_{\nbiga\ast} \nbigc^{n}(\nbiga)_{\bullet}$ is identified with $J(\nbiga,K(\nbiga),\vecbeta)$. Hence, we obtain the desired isomorphism $M^{\GKZ}(\nbiga,K(\nbiga),\vecbeta)\simeq M_{\nbiga,\vecbeta,\bullet}$. \vspace{.1in} According to , the space of global sections of $\Omega^p_{(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga})}$ is $\bigoplus_{I\in S(p)} \bigoplus_{\vecc\in K(\nbiga)^{\circ}} \cnum\cdot t^{\vecc}\tau_I$. Hence, we have \[ \check{\pi}_{\nbiga\ast} \nbigc^p(\nbiga)_{\circ} =\bigoplus_{I\in S(p)} \bigoplus_{\vecc\in K(\nbiga)^{\circ}} \nbigo_{\cnum^m} t^{\vecc}\tau_I. \] Then, as in the case of $\bullet$, we obtain the desired isomorphism $M^{\GKZ}(\nbiga,K(\nbiga)^{\circ},\vecbeta) \simeq M_{\nbiga,\vecbeta,\circ}$. \hfill\qed \subsubsection{Special cases} For any $\vecp=(p_1,\ldots,p_m)\in \seisuu^m$, we put $\supp_+(\vecp):=\{j\,|\,p_j\geq 0\}$ and $\supp_-(\vecp):=\{j\,|\,p_j\leq 0\}$. We set \[ \square_{\vecp} =\prod_{j\in\supp_+(\vecp)}\del_{x_j}^{p_j} -\prod_{j\in\supp_-(\vecp)}\del_{x_j}^{-p_j}. \] We have the morphism $\seisuu^m\lrarr \seisuu^n$ given by $\vecp=(p_1,\ldots,p_m)\longmapsto \sum_{j=1}^m p_j\veca_j$. Let $L_{\nbiga}$ denote the kernel. Suppose that $\Gamma=\seisuu_{\geq 0}\nbiga+\vecc_0$ for an element $\vecc_0\in K(\nbiga)$. Then, as remarked in , $\GKZ(\nbiga,\Gamma,\vecbeta)$ is equivalent to the following ordinary GKZ-hypergeometric system $\GKZ^{\ord}(\nbiga,\vecbeta-\vecc_0)$ for holomorphic functions $\Phi_{\vecc_0}$ on any open subset of $\cnum^m$: \[ \square_{\vecp}\Phi_{\vecc_0}=0 \quad (\forall\vecp\in L_{\nbiga}) \] \[ \Bigl(c_{0i}-\beta_i+\sum_{j=1}^ma_{ji}x_j\del_{x_j}\Bigr) \Phi_{\vecc_0}=0 \quad (i=1,\ldots,n) \] For any $\vecgamma\in\cnum^n$, let $I(\nbiga,\vecgamma)$ denote the left ideal of $\nbigd_{\cnum^m}$ generated by $\square_{\vecp}$ $(\vecp\in L_{\nbiga})$ and $-\gamma_i+\sum_{j=1}^ma_{ji}x_j\del_{x_j}$ $(i=1,\ldots,n)$. Then, we have a natural isomorphism $M^{\GKZ}(\nbiga,\Gamma,\vecbeta) \simeq \nbigd_{\cnum^m}/I(\nbiga,\vecbeta-\vecc_0)$. Suppose that we are given $\Gamma_1\subset\Gamma_2\subset K(\nbiga)$ such that $\Gamma_i=\vecc_i+\seisuu_{\geq 0}\nbiga$. We have the natural morphism $M^{\GKZ}(\nbiga,\Gamma_1,\vecbeta) \lrarr M^{\GKZ}(\nbiga,\Gamma_2,\vecbeta)$ induced by the inclusion $Q(\nbiga,\Gamma_1)\subset Q(\nbiga,\Gamma_2)$. The above isomorphisms induce $g_{\vecc_2,\vecc_1}: \nbigd_{\cnum^m}/I(\nbiga,\vecc_1,\vecbeta) \lrarr \nbigd_{\cnum^m}/I(\nbiga,\vecc_2,\vecbeta)$. We have an expression $\vecc_1=\vecc_2+\sum_{j=1}^m b_j\veca_j$ $(b_j\in\seisuu_{\geq 0})$. We have the morphism of $\nbigd_{\cnum^m}$-modules $\htilde_{\vecc_2,\vecc_1}: \nbigd_{\cnum^m}\lrarr \nbigd_{\cnum^m}$ induced by the right multiplication of $B_{\vecc_2,\vecc_1}= \prod_{j=1}^m\del_j^{b_j}$. We have $\square_{\vecp}\cdot B_{\vecc_2,\vecc_1} =B_{\vecc_2,\vecc_1}\cdot \square_{\vecp}$. We also have \[ \Bigl( c_{1i}-\beta_i+\sum_{j=1}^ma_{ji}x_j\del_{x_j} \Bigr) \cdot\prod_{j=1}^m\del_j^{b_j} = \prod_{j=1}^m\del_j^{b_j}\cdot \Bigl( c_{2i}-\beta_i+\sum_{j=1}^ma_{ji}x_j\del_{x_j} \Bigr). \] Hence, we have the induced morphism $h_{\vecc_2,\vecc_1}: \nbigd_{\cnum^m}/I(\nbiga,\vecc_1,\vecbeta) \lrarr \nbigd_{\cnum^m}/I(\nbiga,\vecc_2,\vecbeta)$. It is easy to check $g_{\vecc_2,\vecc_1} =h_{\vecc_2,\vecc_1}$. \vspace{.1in} Let us consider some more special cases. Suppose that $K(\nbiga)=\seisuu_{\geq 0}\nbiga$, i.e., $X_{\nbiga}^{\aff} =\check{X}^{\aff}_{\nbiga}$. Then, the system $\GKZ(\nbiga,K(\nbiga),0)$ is equivalent to $\GKZ^{\ord}(\nbiga,0)$, and we have $M^{\GKZ}(\nbiga,K(\nbiga),0) \simeq \nbigd_{\cnum^m}/I(\nbiga,0)$. Suppose moreover that $K(\nbiga)^{\circ}=K(\nbiga)+\vecc_1$ for some $\vecc_1\in K(\nbiga)$. Then, $\GKZ(\nbiga,K(\nbiga)^{\circ},0)$ is equivalent to $\GKZ^{\ord}(\nbiga,-\vecc_1)$, and we have $M^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) \simeq \nbigd_{\cnum^m}/I(\nbiga,-\vecc_1)$. The morphism $M^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) \lrarr M^{\GKZ}(\nbiga,K(\nbiga),0)$ induces $\nbigd_{\cnum^m}/I(\nbiga,-\vecc_1) \lrarr \nbigd_{\cnum^m}/I(\nbiga,0)$. If $\vecc_1=\sum_{i=1}^{\ell} \veca_i$ for some $\ell$, it is equal to the morphism induced by the multiplication of $\prod_{j=1}^{\ell}\del_{x_j}$. \subsection{$\nbigrtilde$-modules} \subsubsection{Systems of differential equations and the associated $\nbigrtilde$-modules} We continue to use the notation in \S. Let $\Gamma\subset K(\nbiga)$ be any subset satisfying $\Gamma+\veca\subset \Gamma$ for any $\veca\in\nbiga$. We consider the following system of differential equations $\GKZ_{\nbigrtilde}(\nbiga,\Gamma,\vecbeta)$ for a tuple $\Phi_{\Gamma}=(\Phi_{\vecc}\,|\,\vecc\in \Gamma)$ of holomorphic functions on any open subset of $\cnum\times\cnum^m =\{(\lambda,x_1,\ldots,x_m)\}$: \[ \lambda\del_{x_j}\Phi_{\vecc} =\Phi_{\vecc+\veca_j}, \quad (\forall \vecc\in \Gamma,\,\,j=1,\ldots,m) \] \[ \Bigl( \lambda^2\del_{\lambda} +n\lambda +\sum_{j=1}^m\lambda x_j\del_{x_j} \Bigr) \Phi_{\vecc} =0, \quad (\forall \vecc\in \Gamma) \] \[ \Bigl( \lambda (c_i-\beta_i)+\sum_{j=1}^m a_{ji}\lambda x_j\del_{x_j} \Bigr) \Phi_{\vecc}=0, \quad (\forall\vecc\in \Gamma,\,\,\,i=1,\ldots,n) \] We describe the corresponding $\nbigrtilde_{\cnum^m}$-module. We introduce a free $\nbigo_{\cnum\times\cnum^m}$-module $\nbigq(\nbiga,\Gamma)$ generated by $\Gamma$. Let $\overline{e}(\vecc)$ denote the section corresponding to $\vecc\in\Gamma$. So, we have $\nbigq(\nbiga,\Gamma) =\bigoplus_{\vecc\in\Gamma} \nbigo_{\cnum\times\cnum^m}\overline{e}(\vecc)$. It is an $\nbigrtilde_{\cnum^m}$-module by the actions $\lambda\del_{x_j}\overline{e}(\vecc) =\overline{e}(\vecc+\veca_j)$ and \[ \lambda^2\del_{\lambda} \overline{e}(\vecc) =-n\lambda \overline{e}(\vecc) -\sum_{j=1}^mx_j\lambda\del_{x_j}\overline{e}(\vecc). \] Let $\nbigj(\nbiga,\Gamma,\vecbeta)$ denote the $\nbigrtilde$-submodule of $\nbigq(\nbiga,\Gamma)$ generated by $\bigl( \lambda (c_i-\beta_i)+\sum_{j=1}^m a_{ji}\lambda x_j\del_{x_j} \bigr) \overline{e}(\vecc)$ for $\vecc\in\Gamma$ and $i=1,\ldots,n$. We set $\nbigm^{\GKZ}(\nbiga,\Gamma,\vecbeta):= \nbigq(\nbiga,\Gamma)/\nbigj(\nbiga,\Gamma,\vecbeta)$. As in \S, we have the following. \begin{lem} Let $U$ be any open subset in $\cnum\times\cnum^m$. We have the natural bijective correspondence between the $\nbigrtilde$-homomorphisms $\nbigm^{\GKZ}(\nbiga,\Gamma,\vecbeta)_{|U} \lrarr \nbigo_{U}$ and the solutions of $\GKZ_{\nbigrtilde}(\nbiga,\Gamma,\vecbeta)$ on $U$. \end{lem} \subsubsection{Comparison with the twisted de Rham complexes} We use the notation in \S. Let $q:\cnum\times \check{X}_{\nbiga}^{\aff}\times\cnum^m \lrarr \check{X}_{\nbiga}^{\aff}$ denote the projection. We set \[ \overline{\nbigc}^{k}(\nbiga,\vecbeta)_{\bullet}:= \lambda^{-k} q^{\ast}\Omega^k_{\check{X}^{\aff}_{\nbiga}} (\log \check{D}^{\aff}_{\nbiga}), \quad\quad \overline{\nbigc}^k(\nbiga,\vecbeta)_{\circ}:= \lambda^{-k} q^{\ast}\Omega^k_{(\check{X}^{\aff}_{\nbiga},\check{D}^{\aff}_{\nbiga})}. \] With the differential induced by $d+d(\lambda^{-1}F_{\nbiga})-\kappa(\vecbeta)$, we obtain the complexes of sheaves $\overline{\nbigc}^{\bullet}(\nbiga,\vecbeta)_{\star}$ $(\star=\bullet,\circ)$. Let $\check{\pi}_{\nbiga}: \cnum\times\check{X}^{\aff}_{\nbiga}\times\cnum^m \lrarr \cnum\times \cnum^m$ denote the projection. Each $\check{\pi}_{\nbiga\ast} \overline{\nbigc}^{k}(\nbiga,\vecbeta)_{\star}$ is naturally an $\nbigrtilde_{\cnum^m}$-module by $\lambda\del_{x_j}\bullet(g)= \lambda\del_{x_j}g+\del_{x_j}F_{\nbiga}\cdot g$ $(j=1,\ldots,m)$ and $\lambda^2\del_{\lambda}\bullet(g) =\lambda^2\del_{\lambda}g -F_{\nbiga}\cdot g$. The differentials $d+d(\lambda^{-1}F_{\nbiga})-\kappa(\vecbeta)$ of the complexes are compatible with the actions of $\nbigrtilde_{\cnum^m}$. So, we obtain the following $\nbigrtilde$-modules \[ \nbigm_{\nbiga,\vecbeta,\star}:= \hyperr^n\check{\pi}_{\nbiga\ast} \overline{\nbigc}^{\bullet}(\nbiga,\vecbeta)_{\star} \quad (\star=\bullet,\circ). \] \begin{prop} We have the following natural isomorphisms of $\nbigrtilde_{\cnum^m}$-modules: \[ \nbigm^{\GKZ}(\nbiga,K(\nbiga),\vecbeta)\simeq \nbigm_{\nbiga,\vecbeta,\bullet}, \quad \nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},\vecbeta) \simeq \nbigm_{\nbiga,\vecbeta,\circ}. \] \end{prop} \pf The argument is similar to that in the proof of Proposition . We give just an indication. We have \[ \check{\pi}_{\nbiga\ast} \overline{\nbigc}^p(\nbiga,\vecbeta)_{\bullet} =\bigoplus_{I\in S(p)} \bigoplus_{\vecc\in K(\nbiga)} \nbigo_{\cnum\times\cnum^m} t^{\vecc}\lambda^{-p}\tau_I. \] We have the following equalities. \[ \lambda\del_{x_j}(t^{\vecc}\lambda^{-n}\omega) =t^{\vecc+\veca_j}\lambda^{-n}\omega \] \[ \Bigl( \lambda^2\del_{\lambda} +n\lambda+\sum_{j=1}^m \lambda x_j\del_{x_j} \Bigr) t^{\vecc}\lambda^{-n}\omega =0 \] We also have the following: \[ \Bigl(\lambda (c_i-\beta_i)+\sum_{j=1}^m a_{ji}\lambda x_j\del_{x_j}\Bigr) (t^{\vecc}\lambda^{-n}\omega) = \Bigl( d+d(\lambda^{-1}F_{\nbiga})-\kappa(\vecbeta) \Bigr) (t^{\vecc}\lambda^{-n+1}\omega_i) \] Then, we obtain the claim in the case for $\bullet$ as in the case of Proposition . The claim for $\circ$ is obtained in a similar way. \hfill\qed \begin{rem} As studied in {\rm\S}, we have the following commutative diagram: \[ \begin{CD} \pi_{\dagger}^0\nbigl_{!}(\nbiga,\cnum^m,\id) @>>> \pi_{\dagger}^0\nbigl_{\ast}(\nbiga,\cnum^m,\id)\\ @V{\simeq}VV @V{\simeq}VV \\ \check{\pi}_{\ast} \overline{\nbigc}^{\bullet}(\nbiga,0)_{\circ} @>>> \check{\pi}_{\ast} \overline{\nbigc}^{\bullet}(\nbiga,0)_{\bullet} \end{CD} \] Hence, Proposition {\rm} gives expressions of $\pi_{\dagger}^0\nbigl_{\star}(\nbiga,\cnum^m,\id)$ $(\star=\ast,!)$ as systems of differential equations. \hfill\qed \end{rem} \subsubsection{Special case} We use the notation in \S. For any $\vecp\in \seisuu^m$, we set \[ \overline{\square}_{\vecp}:= \prod_{j\in \supp_+(\vecp)} (\lambda\del_j)^{p_j} -\prod_{i\in\supp_-(\vecp)} (\lambda\del_j)^{-p_j}. \] Suppose that $\Gamma=\seisuu_{\geq 0}\nbiga+\vecc_0$ for an element $\vecc_0\in K(\nbiga)$. Then, $\GKZ_{\nbigrtilde}(\nbiga,\Gamma,\vecbeta)$ is equivalent to the following system $\GKZ_{\nbigrtilde}^{\ord}(\nbiga,\vecbeta-\vecc_0)$ for $\Phi_{\vecc_0}$: \[ \overline{\square}_{\vecp}\Phi_{\vecc_0}=0 \quad (\forall\vecp\in L_{\nbiga}) \] \[ \Bigl( \lambda^2\del_{\lambda} +n\lambda+\sum_{j=1}^mx_j\lambda\del_{x_j} \Bigr)\Phi_{\vecc_0}=0 \] \[ \Bigl( \lambda(c_{0i}-\beta_i) +\sum_{j=1}^ma_{ji}x_j\lambda\del_{x_j}\Bigr) \Phi_{\vecc_0}=0 \quad (i=1,\ldots,n) \] For any $\vecgamma\in\seisuu^m$, let $\nbigi(\nbiga,\vecgamma)$ denote the left ideal of $\nbigrtilde_{\cnum^m}$ generated by $\overline{\square}_{\vecp}$ $(\vecp\in L_{\nbiga})$, $\lambda^2\del_{\lambda}+n\lambda +\sum_{j=1}^m\lambda x_j\del_{x_j}$, and $-\lambda\gamma_i+\sum_{j=1}^ma_{ji}\lambda x_j\del_{x_j}$ $(i=1,\ldots,n)$. Then, we have a natural isomorphism \[ \nbigm^{\GKZ}(\nbiga,\Gamma,\vecbeta) \simeq \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,\vecbeta-\vecc_0). \] Suppose that we are given $\Gamma_1\subset\Gamma_2\subset K(\nbiga)$ such that $\Gamma_i=\vecc_i+\seisuu_{\geq 0}\nbiga$. We have the natural morphism $\nbigm^{\GKZ}(\nbiga,\Gamma_1,\beta) \lrarr \nbigm^{\GKZ}(\nbiga,\Gamma_2,\beta)$ induced by the inclusion $\nbigq(\nbiga,\Gamma_1)\subset \nbigq(\nbiga,\Gamma_2)$. The above isomorphisms induce $g_{\vecc_2,\vecc_1}: \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,\vecc_1,\beta) \lrarr \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,\vecc_2,\beta)$. We have an expression $\vecc_1=\vecc_2+\sum_{i=1}^m b_i\veca_i$ $(b_i\in\seisuu_{\geq 0})$. We have the morphism of $\nbigrtilde_{\cnum^m}$-modules $\htilde_{\vecc_2,\vecc_1}: \nbigrtilde_{\cnum^m} \lrarr \nbigrtilde_{\cnum^m}$ induced by the right multiplication of $\overline{B}_{\vecc_2,\vecc_1}:= \prod_{j=1}^m (\lambda\del_{x_j})^{b_j}$. We have $\square_{\vecp}\cdot \overline{B}_{\vecc_2,\vecc_1} =\overline{B}_{\vecc_2,\vecc_1}\cdot \square_{\vecp}$. We have \[ \Bigl( \lambda(c_{2i}-\beta_i) +\sum_{j=1}^ma_{ji}x_j\lambda\del_{x_j} \Bigr) \prod_{k=1}^m(\lambda\del_{x_k})^{b_k} = \prod_{k=1}^m(\lambda\del_{x_k})^{b_k} \Bigl( \lambda(c_{1i}-\beta_i) +\sum_{j=1}^ma_{ji}x_j\lambda\del_{x_j} \Bigr) \] We have $\bigl(\lambda^2\del_{\lambda}+x_j\lambda\del_{x_j}\bigr) \lambda \del_{x_j} =\lambda\del_{x_j} \bigl(\lambda^2\del_{\lambda}+x_j\lambda\del_{x_j}\bigr)$. Hence, \[ \Bigl( \lambda^2\del_{\lambda} +n\lambda +\sum_{j=1}^mx_j\lambda\del_{x_j} \Bigr)\cdot \prod_{k=1}^m(\lambda\del_k)^{b_k} = \prod_{k=1}^m(\lambda\del_k)^{b_k} \cdot \Bigl( \lambda^2\del_{\lambda} +n\lambda +\sum_{j=1}^mx_j\lambda\del_{x_j} \Bigr) \] Hence, we have the induced morphism $h_{\vecc_2,\vecc_1}: \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,\vecc_1,\vecbeta) \lrarr \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,\vecc_2,\vecbeta)$. It is easy to see that $g_{\vecc_2,\vecc_1}=h_{\vecc_2,\vecc_1}$. \vspace{.1in} Let us consider some more special cases. Suppose that $K(\nbiga)=\seisuu_{\geq 0}\nbiga$, i.e., $X_{\nbiga}^{\aff} =\check{X}^{\aff}_{\nbiga}$. Then, the system $\GKZ_{\nbigrtilde}\bigl(\nbiga,K(\nbiga),0\bigr)$ is equivalent to $\GKZ_{\nbigrtilde}^{\ord}(\nbiga,0)$, and we have $\nbigm^{\GKZ}(\nbiga,K(\nbiga),0) \simeq \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,0)$. Suppose moreover that $K(\nbiga)^{\circ}=K(\nbiga)+\vecc_1$ for some $\vecc_1\in K(\nbiga)$. Then, $\GKZ_{\nbigrtilde}(\nbiga,K(\nbiga)^{\circ},0)$ is equivalent to $\GKZ_{\nbigrtilde}^{\ord}(\nbiga,-\vecc_1)$, and we have $\nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) \simeq \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,-\vecc_1)$. The morphism $\nbigm^{\GKZ}(\nbiga,K(\nbiga)^{\circ},0) \lrarr \nbigm^{\GKZ}(\nbiga,K(\nbiga),0)$ induces $\nbigrtilde_{\cnum^m}/\nbigi(\nbiga,-\vecc_1) \lrarr \nbigrtilde_{\cnum^m}/\nbigi(\nbiga,0)$. If $\vecc_1=\sum_{j=1}^{\ell} \veca_j$ for some $\ell$, it is equal to the morphism induced by the multiplication of $\prod_{j=1}^{\ell}(\lambda\del_{x_j})$. \title{Twistor property of GKZ-hypergeometric systems} \author{Takuro Mochizuki} \date{} \maketitle \begin{abstract} We study the mixed twistor $\nbigd$-modules associated to meromorphic functions. In particular, we describe their push-forward and specialization under some situations. We apply the results to study the twistor property of a type of better behaved GKZ-hypergeometric systems, and to study their specializations. As a result, we obtain some isomorphisms of mixed TEP-structures in the local mirror symmetry. {\footnotesize \vspace{.1in} \noindent Keywords: mixed twistor $\nbigd$-module, generalized Hodge structure, polarization, GKZ-hypergeometric system, local mirror symmetry. \noindent MSC2010: 14F10 32C38 32S35 } \end{abstract} \section{Introduction} \input{1} \section{$\nbigd$-modules associated to meromorphic functions} \input{2} \section{Mixed twistor $\nbigd$-modules associated to meromorphic functions} \input{3} \section{Graded sesqui-linear dualities} \input{4} \section{Comparisons of polarizations} \input{5} \section{Mixed twistor $\nbigd$-modules and GKZ-hypergeometric systems} \input{6} \section{Quantum $\nbigd$-modules} \input{7} \appendix \section{Duality of meromorphic flat bundles} \input{A} \section{Pairings and their functoriality} \input{B} \section{Some functoriality of $\nbigrtilde$-modules} \input{C} \section{Comparison of some $\nbigrtilde$-modules} \input{D} \section{Better behaved GKZ-systems and de Rham complexes} \input{E} \input{reference} \vspace{.1in} \noindent {\em Address\\ Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-8502, Japan\\ takuro@kurims.kyoto-u.ac.jp } \newcounter{goodsection} \newtheorem{thm}{Theorem}[section] \newtheorem{cor}[thm]{Corollary} \newtheorem{conj}[thm]{Conjecture} \newtheorem{rem}[thm]{Remark} \newtheorem{lem}[thm]{Lemma} \newtheorem{prop}[thm]{Proposition} \newtheorem{df}[thm]{Definition} \newtheorem{prob}[thm]{Problem} \newtheorem{example}[thm]{Example} \newtheorem{condition}[thm]{Condition} \newtheorem{assumption}[thm]{Assumption} \newtheorem{principle}[thm]{Principle} \newtheorem{notation}[thm]{Notation} \newtheorem{convention}[thm]{Convention} \newcommand{\nbiga}{\mathcal{A}} \newcommand{\nbigb}{\mathcal{B}} \newcommand{\nbigc}{\mathcal{C}} \newcommand{\nbigd}{\mathcal{D}} \newcommand{\nbige}{\mathcal{E}} \newcommand{\nbigf}{\mathcal{F}} \newcommand{\nbigg}{\mathcal{G}} \newcommand{\nbigh}{\mathcal{H}} \newcommand{\nbigi}{\mathcal{I}} \newcommand{\nbigj}{\mathcal{J}} \newcommand{\nbigk}{\mathcal{K}} \newcommand{\nbigl}{\mathcal{L}} \newcommand{\nbigm}{\mathcal{M}} \newcommand{\nbign}{\mathcal{N}} \newcommand{\nbigo}{\mathcal{O}} \newcommand{\nbigp}{\mathcal{P}} \newcommand{\nbigq}{\mathcal{Q}} \newcommand{\nbigr}{\mathcal{R}} \newcommand{\nbigs}{\mathcal{S}} \newcommand{\nbigt}{\mathcal{T}} \newcommand{\nbigu}{\mathcal{U}} \newcommand{\nbigv}{\mathcal{V}} \newcommand{\nbigw}{\mathcal{W}} \newcommand{\nbigx}{\mathcal{X}} \newcommand{\nbigy}{\mathcal{Y}} \newcommand{\nbigz}{\mathcal{Z}} \newcommand{\nbiggr}{\mathcal{GR}} \newcommand{\Epsilon}{\nbige} \newcommand{\pb}[1]{{#1}^{\ast}} \newcommand{\proj}{\mathbb{P}} \newcommand{\seisuu}{{\mathbb Z}} \newcommand{\rnum}{{\mathbb Q}} \newcommand{\nnum}{{\mathbb N}} \newcommand{\cnum}{{\mathbb C}} \newcommand{\real}{{\mathbb R}} \newcommand{\hyperh}{\mathbb{H}} \newcommand{\hyperr}{\mathbb{R}} \newcommand{\iti}{\mathbb{I}} \newcommand{\gm}{\mathbb{G}_m} \newcommand{\Tate}{\mathbb{T}} \newcommand{\newTate}{\pmb{T}} \newcommand{\DD}{\mathbb{D}} \newcommand{\EE}{\mathbb{E}} \newcommand{\FF}{\mathbb{F}} \newcommand{\GG}{\mathbb{G}} \newcommand{\gbiga}{\mathfrak A} \newcommand{\gbigb}{\mathfrak B} \newcommand{\gbigc}{\mathfrak C} \newcommand{\gbigd}{\mathfrak D} \newcommand{\gbige}{\mathfrak E} \newcommand{\gbigf}{\mathfrak F} \newcommand{\gbigg}{\mathfrak G} \newcommand{\gbigh}{\mathfrak H} \newcommand{\gbigi}{\mathfrak I} \newcommand{\gbigj}{\mathfrak J} \newcommand{\gbigk}{\mathfrak K} \newcommand{\gbigl}{\mathfrak L} \newcommand{\gbigm}{\mathfrak M} \newcommand{\gbign}{\mathfrak N} \newcommand{\gbigo}{\mathfrak O} \newcommand{\gbigp}{\mathfrak P} \newcommand{\gbigq}{\mathfrak Q} \newcommand{\gbigr}{\mathfrak R} \newcommand{\gbigs}{\mathfrak S} \newcommand{\gbigt}{\mathfrak T} \newcommand{\gbigu}{\mathfrak U} \newcommand{\gbigv}{\mathfrak V} \newcommand{\gbigw}{\mathfrak W} \newcommand{\gbigx}{\mathfrak X} \newcommand{\gbigy}{\mathfrak Y} \newcommand{\gbigz}{\mathfrak Z} \newcommand{\gminia}{\mathfrak a} \newcommand{\gminib}{\mathfrak b} \newcommand{\gminic}{\mathfrak c} \newcommand{\gminid}{\mathfrak d} \newcommand{\gminie}{\mathfrak e} \newcommand{\gminif}{\mathfrak f} \newcommand{\gminig}{\mathfrak g} \newcommand{\gminih}{\mathfrak h} \newcommand{\gminii}{\mathfrak i} \newcommand{\gminij}{\mathfrak j} \newcommand{\gminik}{\mathfrak k} \newcommand{\gminil}{\mathfrak l} \newcommand{\gminim}{\mathfrak m} \newcommand{\gminin}{\mathfrak n} \newcommand{\gminio}{\mathfrak o} \newcommand{\gminip}{\mathfrak p} \newcommand{\gminiq}{\mathfrak q} \newcommand{\gminir}{\mathfrak r} \newcommand{\gminis}{\mathfrak s} \newcommand{\gminit}{\mathfrak t} \newcommand{\gminiu}{\mathfrak u} \newcommand{\gminiv}{\mathfrak v} \newcommand{\gminiw}{\mathfrak w} \newcommand{\gminix}{\mathfrak x} \newcommand{\gminiy}{\mathfrak y} \newcommand{\gminiz}{\mathfrak z} \newcommand{\vexi}{{\boldsymbol \xi}} \newcommand{\veeta}{{\boldsymbol \eta}} \newcommand{\vecxi}{{\boldsymbol \xi}} \newcommand{\veceta}{{\boldsymbol \eta}} \newcommand{\vecrho}{{\boldsymbol \rho}} \newcommand{\vece}{{\boldsymbol e}} \newcommand{\vecr}{{\boldsymbol r}} \newcommand{\vecv}{{\boldsymbol v}} \newcommand{\vecu}{{\boldsymbol u}} \newcommand{\vecw}{{\boldsymbol w}} \newcommand{\vecgamma}{{\boldsymbol \gamma}} \newcommand{\vecl}{{\boldsymbol l}} \newcommand{\veczero}{{\boldsymbol 0}} \newcommand{\vecone}{{\boldsymbol 1}} \newcommand{\vecalpha}{{\boldsymbol \alpha}} \newcommand{\veca}{{\boldsymbol a}} \newcommand{\vecb}{{\boldsymbol b}} \newcommand{\vecbeta}{{\boldsymbol \beta}} \newcommand{\vecdelta}{{\boldsymbol \delta}} \newcommand{\vecs}{{\boldsymbol s}} \newcommand{\vect}{{\boldsymbol t}} \newcommand{\vecc}{{\boldsymbol c}} \newcommand{\vecd}{{\boldsymbol d}} \newcommand{\vech}{{\boldsymbol h}} \newcommand{\veck}{{\boldsymbol k}} \newcommand{\vecm}{{\boldsymbol m}} \newcommand{\vecM}{{\boldsymbol M}} \newcommand{\vecN}{{\boldsymbol N}} \newcommand{\vecI}{{\boldsymbol I}} \newcommand{\vecomega}{{\boldsymbol \omega}} \newcommand{\vecx}{{\boldsymbol x}} \newcommand{\vecf}{{\boldsymbol f}} \newcommand{\vecepsilon}{{\boldsymbol \epsilon}} \newcommand{\vecF}{{\boldsymbol F}} \newcommand{\vecE}{{\boldsymbol E}} \newcommand{\vecW}{{\boldsymbol W}} \newcommand{\vecn}{{\boldsymbol n}} \newcommand{\vecp}{{\boldsymbol p}} \newcommand{\veczeta}{{\boldsymbol \zeta}} \newcommand{\vecz}{{\boldsymbol z}} \newcommand{\vecC}{{\boldsymbol C}} \newcommand{\vecU}{{\boldsymbol U}} \newcommand{\vecV}{{\boldsymbol V}} \newcommand{\vecA}{{\boldsymbol A}} \newcommand{\vecB}{{\boldsymbol B}} \newcommand{\vecS}{{\boldsymbol S}} \newcommand{\dual}{\lor} \newcommand{\larr}{\leftarrow} \newcommand{\llarr}{\longleftarrow} \newcommand{\uarr}{\uparrow} \newcommand{\Larr}{\Longleftarrow} \newcommand{\Llarr}{\Longleftarrow} \newcommand{\Uarr}{\Uparrow} \newcommand{\rarr}{\rightarrow} \newcommand{\lrarr}{\longrightarrow} \newcommand{\darr}{\downarrow} \newcommand{\Rarr}{\Rightarrow} \newcommand{\Lrarr}{\Longrightarrow} \newcommand{\Darr}{\Downarrow} \newcommand{\barr}{\leftrightarrow} \newcommand{\lbarr}{\longleftrightarrow} \newcommand{\udarr}{\updownarrow} \newcommand{\Barr}{\Leftrightarrow} \newcommand{\Lbarr}{\Longleftrightarrow} \newcommand{\Udarr}{\Updownarrow} \newcommand{\lmapsto}{\longmapsto} \newcommand{\hlarr}{\hookleftarrow} \newcommand{\hrarr}{\hookrightarrow} \newcommand{\lhrarr}{ \mbox{ \begin{picture}(2,8)(0,-2) \put(2,2){\oval(4,4)[l]} \end{picture}\rule[.65mm]{6mm}{.1mm}\raisebox{-.1mm}{$\rightarrow$}}} \newcommand{\hdarr} {\mbox{$ \begin{array}{r} \raisebox{-.6ex} {\begin{picture}(6,2)(-4,0) \put(-2,0){\oval(4,4)[t]} \put(0,0){\line(0,-1){8}} \end{picture}}\hspace{.23mm}\\ \rule[-1.5mm]{.1mm}{3mm}\hspace{.9mm}\\ \raisebox{.6ex}{$\downarrow$} \end{array}$}} \newcommand{\pf}{{\bf Proof}\hspace{.1in}} \newcommand{\qed}{\mbox{\rule{1.2mm}{3mm}}} \def\Hom{\mathop{\rm Hom}\nolimits} \def\poly{\mathop{\rm poly}\nolimits} \def\End{\mathop{\rm End}\nolimits} \def\Ext{\mathop{\rm Ext}\nolimits} \def\Extbar{\underline{\mathop{\rm Ext}\nolimits}} \def\Cok{\mathop{\rm Cok}\nolimits} \def\cok{\mathop{\rm Cok}\nolimits} \def\Image{\mathop{\rm Im}\nolimits} \def\Realpart{\mathop{\rm Re}\nolimits} \def\realpart{\mathop{\rm Re}\nolimits} \def\Re{\mathop{\rm Re}\nolimits} \def\CH{\mathop{\rm CH}\nolimits} \def\Gr{\mathop{\rm Gr}\nolimits} \def\gr{\mathop{\rm Gr}\nolimits} \def\GL{\mathop{\rm GL}\nolimits} \def\SL{\mathop{\rm SL}\nolimits} \def\Tot{\mathop{\rm Tot}\nolimits} \def\Cone{\mathop{\rm Cone}\nolimits} \def\rank{\mathop{\rm rank}\nolimits} \def\Spec{\mathop{\rm Spec}\nolimits} \def\Nu{{\mathcal V}} \def\RHom{\mathop{\rm RHom}\nolimits} \def\Rhombar{\underline{\mathop{\rm RHom}\nolimits}} \def\Hombar{\underline{\mathop{\rm Hom}\nolimits}} \def\trdeg{\mathop{\rm tr.deg}\nolimits} \def\Ker{\mathop{\rm Ker}\nolimits} \def\ker{\mathop{\rm Ker}\nolimits} \def\modulo{\mathop{\rm modulo}\nolimits} \def\length{\mathop{\rm length}\nolimits} \def\Gr{\mathop{\rm Gr}\nolimits} \def\Sym{\mathop{\rm Sym}\nolimits} \def\sym{\mathop{\rm sym}\nolimits} \def\ad{\mathop{\rm ad}\nolimits} \def\Res{\mathop{\rm Res}\nolimits} \def\Rep{\mathop{\rm Rep}\nolimits} \def\ord{\mathop{\rm ord}\nolimits} \def\degpar{\mathop{\rm par\textrm{-}deg}\nolimits} \def\parch{\mathop{\rm par\textrm{-}ch}\nolimits} \def\ch{\mathop{ch}\nolimits} \def\parchern{\mathop{\rm par\textrm{-}c}\nolimits} \def\degnil{\mathop{\rm nil\textrm{-}deg}\nolimits} \def\Ric{\mathop{\rm Ric}\nolimits} \def\tr{\mathop{\rm tr}\nolimits} \def\Tr{\mathop{\rm Tr}\nolimits} \def\vol{\mathop{\rm dvol}\nolimits} \def\dvol{\mathop{\rm dvol}\nolimits} \def\Diff{\mathop{\rm Diff}\nolimits} \def\vecdeg{\mathop{\rm \bf deg}\nolimits} \def\vecord{\mathop{\rm \bf ord}\nolimits} \def\can{\mathop{\rm can}\nolimits} \def\var{\mathop{\rm var}\nolimits} \def\id{\mathop{\rm id}\nolimits} \def\Pat{\mathop{\rm Pat}\nolimits} \def\Prim{\mathop{\rm Prim}\nolimits} \def\Red{\mathop{\rm Red}\nolimits} \def\gcd{\mathop{\rm g.c.d.}\nolimits} \def\codim{\mathop{\rm codim}\nolimits} \def\gap{\mathop{\rm gap}\nolimits} \def\reduced{\mathop{\rm red}\nolimits} \def\filt{\mathop{\rm Filt}\nolimits} \def\equi{\mathop{\rm Equi}\nolimits} \def\Supp{\mathop{\rm Supp}\nolimits} \def\VPTgen{\mathop{\rm VPT_{gen}}\nolimits} \def\VPTgenwild{\mathop{\rm VPT_{gen}^{\wild}}\nolimits} \def\MPT{\mathop{\rm MPT}\nolimits} \def\RHD{\mathop{\rm RHD}\nolimits} \def\ch{\mathop{\rm ch}\nolimits} \def\Irrep{\mathop{\rm Irrep}\nolimits} \def\Irr{\mathop{\rm Irr}\nolimits} \def\Dom{\mathop{\rm Dom}\nolimits} \def\Zar{\mathop{\rm Zar}\nolimits} \def\wt{\mathop{\rm wt}\nolimits} \def\diag{\mathop{\rm diag}\nolimits} \def\Lef{\mathop{\rm Lef}\nolimits} \def\Obv{\mathop{\rm Obv}\nolimits} \newcommand{\Cpoly}[1]{\nbigc^{poly}_{#1}} \newcommand{\hh}[1]{{\rm H}^{#1}} \newcommand{\nbighh}[1]{\nbigh ^{#1}} \newcommand{\hhbar}[1]{\underbar{{\rm H}}^{#1}} \newcommand{\tentimes}{\otimes \cdots \otimes} \newcommand{\tenwedge}{\wedge \cdots \wedge} \newcommand{\del}{\partial} \newcommand{\delbar}{\overline{\del}} \newcommand{\bardel}{\overline{\del}} \newcommand{\delbartilde}{\widetilde{\delbar}} \newcommand{\deltilde}{\widetilde{\del}} \newcommand{\dtilde}{\widetilde{d}} \newcommand{\deltatilde}{\widetilde{\delta}} \newcommand{\deltabar}{(\Delta ^{l\,\circ})_{-}} \newcommand{\Deltabar}{\overline{\Delta}} \newcommand{\Deltabarast}{\Deltabar^{\ast}} \newcommand{\pardeg}{\degpar} \newcommand{\nildeg}{\degnil} \newcommand{\nhom}{{\mathcal Hom}} \newcommand{\ntor}{{\mathcal Tor}} \newcommand{\nend}{{\mathcal End}} \newcommand{\next}{{\mathcal Ext}} \newcommand{\hsdgla}[1]{\Omega (\gminig_{#1} ^{\cdot})^{\cdot}} \newcommand{\hsdglai}[2]{\Omega (\gminig_{#1} ^{\cdot})^{#2}} \newcommand{\vecP}{\bf P} \newcommand{\barh}{\hbar} \newcommand{\tpoly}{T_{poly}} \newcommand{\dpoly}{D_{poly}} \newcommand{\tpolybar}{\overline{T}_{poly}} \newcommand{\dpolybar}{\overline{D}_{poly}} \newcommand{\gaisekig}{\bigwedge^{\cdot}\gminig} \newcommand{\gaisekigg}{\bigwedge^{\cdot}\gminig^{\lor}} \newcommand{\gaisekih}{\bigwedge^{\cdot}\gminih} \newcommand{\gaisekihh}{\bigwedge^{\cdot}\gminih^{\bot}} \newcommand{\iremonot}{T_{poly}(U)\otimes\gaisekigg} \newcommand{\iremonod}{D_{poly}(U)\otimes\gaisekigg} \newcommand{\aaa}{\alpha} \newcommand{\ttt}{\theta} \newcommand{\TTT}{\Theta} \newcommand{\nbar}{\underline{n}} \newcommand{\dbar}{\underline{d}} \newcommand{\jbar}{\underline{j}} \newcommand{\mbar}{\underline{m}} \newcommand{\kbar}{\underline{k}} \newcommand{\jitibar}{\underline{j+1}} \newcommand{\jminusitibar}{\underline{j-1}} \newcommand{\mitibar}{\underline{m+1}} \newcommand{\mzerobar}{\underline{m_0}} \newcommand{\mzeroijbar}{\underline{m_0(i,j)}} \newcommand{\mitiitibar}{\underline{m_1+1}} \newcommand{\mbariti}{\underline{m_1}} \newcommand{\mminusitibar}{\underline{m-1}} \newcommand{\lbariti}{\underline{l_1}} \newcommand{\litibar}{\underline{l-1}} \newcommand{\ibar}{\underline{i}} \newcommand{\iitibar}{\underline{i+1}} \newcommand{\nminusitibar}{\underline{n-1}} \newcommand{\nminusnibar}{\underline{n-2}} \newcommand{\lminusitibar}{\underline{l-1}} \newcommand{\lbar}{\underline{l}} \newcommand{\pbar}{\underline{p}} \newcommand{\itibar}{\underline{1}} \newcommand{\nibar}{\underline{2}} \newcommand{\zerobar}{\underline{0}} \newcommand{\alphasitabar}{\underline{\alpha}} \newcommand{\shikaku}{\sharp} \newcommand{\sankaku}{\triangle} \newcommand{\twoprime}{\prime\prime} \newcommand{\harmonicbundle}{(E,\delbar_E,\theta,h)} \newcommand{\harmonicbundledual}{(E^{\lor},\delbar_{E^{\lor}},\theta^{\lor},h^{\lor})} \newcommand{\conjugate}{\clubsuit} \newcommand{\barz}{\overline{z}} \newcommand{\zbar}{\barz} \newcommand{\zetabar}{\overline{\zeta}} \newcommand{\baralpha}{\overline{\alpha}} \newcommand{\alphabar}{\baralpha} \newcommand{\barlambda}{\overline{\lambda}} \newcommand{\lambdabar}{\barlambda} \newcommand{\fbar}{\overline{f}} \newcommand{\varphibar}{\overline{\varphi}} \newcommand{\etabar}{\overline{\eta}} \newcommand{\nablabar}{\overline{\nabla}} \newcommand{\tbar}{\overline{t}} \newcommand{\xbar}{\overline{x}} \newcommand{\sbar}{\overline{s}} \newcommand{\Hbar}{\overline{H}} \newcommand{\Hbarepsilon}{\overline{H}_{\epsilon}} \newcommand{\Abar}{\overline{A}} \newcommand{\modelbundle}[2]{E({#1},{#2})} \newcommand{\modeldeform}{{\mathcal Mod}} \newcommand{\transmatrix}{B} \newcommand{\volume}{\Omega} \newcommand{\dotvecw}{\dot{\vecw}} \newcommand{\vecwdot}{\dotvecw} \newcommand{\dotF}{\dot{F}} \newcommand{\dotvece}{\dot{\vece}} \newcommand{\vecedot}{\dotvece} \newcommand{\dote}{\dot{e}} \newcommand{\dotw}{\dot{w}} \newcommand{\dotPhi}{\dot{\Phi}} \newcommand{\doth}{\dot{h}} \newcommand{\dottheta}{\dot{\theta}} \newcommand{\dotn}{\dot{n}} \newcommand{\dotdelbar}{\dot{\delbar}} \newcommand{\dotvecv}{\dot{\vecv}} \newcommand{\dotv}{\dot{v}} \newcommand{\soeji}{I} \newcommand{\Poin}{{\bf p}} \newcommand{\poin}{\Poin} \newcommand{\prolong}[1]{{}^{\diamond}{#1}} \newcommand{\prolongg}[2]{{}_{#1}{#2}} \newcommand{\leftupdown}[2]{{}^{#1}_{#2}} \newcommand{\resddlambda}{\Res(\DD^{\lambda})} \newcommand{\residuei}{\Res_{\nbigd_i}(\DD)} \newcommand{\residueidagger}{\Res_{\nbigd_i^{\dagger}}(\DD^{\dagger})} \newcommand{\residueiflat}{\Res_{\nbigd_i}(\DD^f)} \newcommand{\residueiflatdagger}{\Res_{\nbigd^{\dagger}_i}(\DD^{\dagger\,f})} \newcommand{\DDlambda}{\DD^{\lambda}} \newcommand{\DDlambdastar}{\DD^{\lambda\,\star}} \newcommand{\DDlambdahat}{\widehat{\DD}^{\lambda}} \newcommand{\DDhatlambda}{\DDlambdahat} \newcommand{\DDlambdaf}{\DD^{\lambda\,f}} \newcommand{\DDlambdaast}{\DD^{\lambda\,\ast}} \newcommand{\weightfilt}{W} \newcommand{\laplacian}{\Delta''} \newcommand{\Deltalambda}{\Delta^{\lambda}} \newcommand{\doublelangle}{\langle\langle} \newcommand{\doublerangle}{\rangle\rangle} \newcommand{\doublebiglangle}{\big\langle\big\langle} \newcommand{\doublebigrangle}{\big\rangle\big\rangle} \newcommand{\graded}{{\mathcal Gr}} \newcommand{\deformoverorigin}{\prolong{\nbige}_{|\cnum_{\lambda}\times O}} \newcommand{\deformoverorigindagger}{\prolong{\nbige^{\dagger}}_{|\cnum_{\mu}\times O}} \newcommand{\nbigelambda}{\nbige^{\lambda}} \newcommand{\nbigelambdazero}{\nbige^{\lambda_0}} \newcommand{\nbigelor}{\nbige^{\lor}} \newcommand{\tiisai}{\nbigu} \newcommand{\blowup}{\widetilde} \newcommand{\projection}{\gminiq} \newcommand{\directsummand}{\nbigu} \newcommand{\thetainfty}{\theta^{(\infty)}} \newcommand{\hinfty}{h^{(\infty)}} \newcommand{\superinfty}{^{(\infty)}} \newcommand{\partialkoszul}{\Pi(N_1,\ldots,N_n)} \newcommand{\partialkoszulinfty}{\Pi(N_1\superinfty,\ldots,N_n\superinfty)} \newcommand{\gradedrestrictedtoni}{\graded_{b|\nbigd_{\nibar}}^{(1)}} \newcommand{\nbigxlambda}{\nbigx^{\lambda}} \newcommand{\nbigxlambdazero}{\nbigx^{\lambda_0}} \newcommand{\nbigdlambda}{\nbigd^{\lambda}} \newcommand{\KMS}{{\mathcal{KMS}}} \newcommand{\EKMS}{{\mathcal{EKMS}}} \newcommand{\KMSoverline}{\overline{\mathcal{KMS}}} \newcommand{\KMSE}[1]{\KMS(\nbige^{#1})} \newcommand{\KMSEoverline}[1]{\overline{\KMS}(\nbige^{#1})} \newcommand{\KMSEprolongg}[2]{\KMS(\prolongg{#1}{\nbige^{#2}})} \newcommand{\Par}{{\mathcal Par}} \newcommand{\Sp}{{\mathcal Sp}} \newcommand{\ParE}[1]{{\mathcal Par}(\nbige^{#1})} \newcommand{\SpE}[1]{{\mathcal Sp}(\nbige^{#1})} \newcommand{\ParEprolongg}[2]{{\mathcal Par}(\prolongg{#1}{\nbige}^{#2})} \newcommand{\SpEprolongg}[2]{{\mathcal Sp}(\prolongg{#1}{\nbige}^{#2})} \newcommand{\kmsmap}{\gminik} \newcommand{\paramap}{\gminip} \newcommand{\eigenmap}{\gminie} \newcommand{\multiplicity}{\gminim} \newcommand{\twistmap}{\gminit} \newcommand{\lefttop}[1]{{}^{#1}\!} \newcommand{\leftbottom}[1]{{}_{#1}\!} \def\Filt{\mathop{\rm Filt}\nolimits} \def\Equi{\mathop{\rm Equi}\nolimits} \def\Bifilt{\mathop{\rm Bifilt}\nolimits} \def\irr{\mathop{\rm irr}\nolimits} \def\reg{\mathop{\rm reg}\nolimits} \def\Def{\mathop{\rm Def}\nolimits} \def\Harm{\mathop{\rm Harm}\nolimits} \def\nil{\mathop{\rm nil}\nolimits} \def\Nil{\mathop{\rm Nil}\nolimits} \newcommand{\FEzero}{(F^{(\lambda_0)},\EE^{(\lambda_0)})} \newcommand{\Fzero}{F^{(\lambda_0)}} \newcommand{\EEzero}{\EE^{(\lambda_0)}} \newcommand{\FFzero}{\FF^{(\lambda_0)}} \newcommand{\GGzero}{\GG^{(\lambda_0)}} \newcommand{\Vzero}{V^{(\lambda_0)}} \newcommand{\Vzeroprime}{V^{\prime\,(\lambda_0)}} \newcommand{\psizero}{\psi^{(\lambda_0)}} \newcommand{\tildepsizero}{\widetilde{\psi}^{(\lambda_0)}} \newcommand{\tildepsi}{\widetilde{\psi}} \newcommand{\psitilde}{\tildepsi} \newcommand{\nbigvzero}{\nbigv^{(\lambda_0)}} \newcommand{\nbigfzero}{\nbigf^{(\lambda_0)}} \newcommand{\nbigqzero}{\nbigq^{(\lambda_0)}} \newcommand{\Ulambdazero}{U^{(\lambda_0)}} \newcommand{\Uzero}{U^{(\lambda_0)}} \newcommand{\Psizero}{\Psi^{(\lambda_0)}} \newcommand{\deldel}{\eth} \newcommand{\deldelbar}{\overline{\deldel}} \newcommand{\lamda}{\lambda} \newcommand{\lambdazero}{\lambda_{0}} \newcommand{\lamdazero}{\lamda_0} \newcommand{\kataim}{\gbigm} \newcommand{\AAA}{{\boldsymbol A}} \newcommand{\distribution}{\gbigd\gminib} \newcommand{\naiveprolong}[1]{\lefttop{\square}{#1}} \newcommand{\naiveprolongg}[2]{{}^{\square}_{#1}{#2}} \newcommand{\rmoduleprolong}[1]{\mathfrak #1} \newcommand{\vecnbign}{{\mathcal {\boldsymbol {\tilde{N}}}}} \newcommand{\nbigvecn}{\boldsymbol{\mathcal N}} \newcommand{\nbigvecs}{\boldsymbol{\mathcal S}} \newcommand{\vecnbigs}{\boldsymbol{\mathcal S}} \newcommand{\vecnbigp}{\boldsymbol{\mathcal P}} \newcommand{\supp}{\gminis} \newcommand{\kakkolambdazero}{^{(\lambda_0)}} \newcommand{\tildenbigm}{\widetilde{\nbigm}} \newcommand{\tildenbigt}{\widetilde{\nbigt}} \newcommand{\tildegbige}{\widetilde{\gbige}} \newcommand{\tildenbige}{\widetilde{\nbige}} \newcommand{\dualhenomap}{\clubsuit} \newcommand{\rtriplecat}{\nbigr\textrm{-Triples}} \newcommand{\closedopen}[2]{[#1,#2[} \newcommand{\openclosed}[2]{]#1,#2]} \newcommand{\openopen}[2]{]#1,#2[} \newcommand{\closedclosed}[2]{[#1,#2]} \newcommand{\PH}{{\mathcal P}{\mathcal H}} \newcommand{\harmonicbundleretu}{\bigl(E^{(i)},\delbar^{(i)},\theta^{(i)},h^{(i)}\bigr)} \newcommand{\harmonicbundlebubunretu}{\bigl(E^{(i')},\delbar^{(i')},\theta^{(i')},h^{(i')}\bigr)} \newcommand{\harmonicbundleinfty}{\bigl(E^{(\infty)},\delbar^{(\infty)},\theta^{(\infty)},h^{(\infty)}\bigr)} \newcommand{\regularfilteredHiggsbundleretu}{\bigl(E^{(i)},\delbar^{(i)},\vecF^{(i)},\theta^{(i)}\bigr)} \newcommand{\regularfilteredHiggsbundlebubunretu}{\bigl(E^{(i')},\delbar^{(i')},\vecF^{(i')},\theta^{(i')}\bigr)} \newcommand{\regularfilteredHiggsbundleinfty}{\bigl(E^{(\infty)},\delbar^{(\infty)},\vecF^{(\infty)},\theta^{(\infty)}\bigr)} \newcommand{\nbigeilambda}{\nbige^{(i)\,\lambda}} \newcommand{\nbigeinftylambda}{\nbige^{(\infty)\,\lambda}} \newcommand{\DDilambda}{\DD^{(i)\,\lambda}} \newcommand{\DDinftylambda}{\DD^{(\infty)\,\lambda}} \newcommand{\nbigeizero}{\nbige^{(i)\,0}} \newcommand{\nbigeinftyzero}{\nbige^{(\infty)\,0}} \newcommand{\unitary}{\gminiu} \newcommand{\gl}{\gminig\gminil} \newcommand{\minisl}{\gminis\gminil} \newcommand{\delbarhat}{\widehat{\delbar}} \newcommand{\delhat}{\widehat{\del}} \newcommand{\omegahat}{\widehat{\omega}} \newcommand{\omegatilde}{\widetilde{\omega}} \newcommand{\vecEhat}{\widehat{\vecE}} \newcommand{\thetahat}{\widehat{\theta}} \newcommand{\Ehat}{\widehat{E}} \newcommand{\hhat}{\widehat{h}} \newcommand{\vecvhat}{\widehat{\vecv}} \newcommand{\vhat}{\widehat{v}} \newcommand{\Ahat}{\widehat{A}} \newcommand{\Gtilde}{\widetilde{G}} \newcommand{\rhotilde}{\widetilde{\rho}} \newcommand{\gun}{G} \newcommand{\xtilde}{\widetilde{x}} \newcommand{\Etilde}{\widetilde{E}} \newcommand{\vecEtilde}{\widetilde{\vecE}} \newcommand{\thetatilde}{\widetilde{\theta}} \newcommand{\Vhat}{\widehat{V}} \newcommand{\nablahat}{\widehat{\nabla}} \newcommand{\Vtilde}{\widetilde{V}} \newcommand{\nablatilde}{\widetilde{\nabla}} \newcommand{\vecvtilde}{\widetilde{\vecv}} \newcommand{\vecwtilde}{\widetilde{\vecw}} \newcommand{\vecutilde}{\widetilde{\vecu}} \newcommand{\zerohat}{\widehat{0}} \newcommand{\vecvbar}{\overline{\vecv}} \newcommand{\fhat}{\widehat{f}} \newcommand{\vecVhat}{\widehat{\vecV}} \newcommand{\vecwbar}{\overline{\vecw}} \newcommand{\wbar}{\overline{w}} \newcommand{\gbar}{\overline{g}} \newcommand{\Chat}{\widehat{C}} \newcommand{\kappatilde}{\widetilde{\kappa}} \newcommand{\mutilde}{\widetilde{\mu}} \newcommand{\nbigetilde}{\widetilde{\nbige}} \newcommand{\DDlambdatilde}{\widetilde{\DD}^{\lambda}} \newcommand{\htilde}{\widetilde{h}} \newcommand{\nbigelambdatilde}{\widetilde{\nbige}^{\lambda}} \newcommand{\ptilde}{\widetilde{p}} \newcommand{\pitilde}{\widetilde{\pi}} \newcommand{\Phitilde}{\widetilde{\Phi}} \newcommand{\vtilde}{\widetilde{v}} \newcommand{\phat}{\widehat{p}} \newcommand{\nbigbhat}{\widehat{\nbigb}} \newcommand{\ftilde}{\widetilde{f}} \newcommand{\utilde}{\widetilde{u}} \newcommand{\Ftilde}{\widetilde{F}} \newcommand{\Phibar}{\overline{\Phi}} \newcommand{\gminiabar}{\overline{\gminia}} \newcommand{\atilde}{\widetilde{a}} \newcommand{\stilde}{\widetilde{s}} \newcommand{\mubar}{\overline{\mu}} \newcommand{\Psitilde}{\widetilde{\Psi}} \newcommand{\nbiglhat}{\widehat{\nbigl}} \newcommand{\sigmatilde}{\widetilde{\sigma}} \newcommand{\nbigltilde}{\widetilde{\nbigl}} \newcommand{\nbigdhat}{\widehat{\nbigd}} \newcommand{\DDtilde}{\widetilde{\DD}} \newcommand{\nbigxhat}{\widehat{\nbigx}} \newcommand{\nbigkhat}{\widehat{\nbigk}} \newcommand{\Rtilde}{\widetilde{R}} \newcommand{\Dhat}{\widehat{D}} \newcommand{\vecgammabar}{\overline{\vecgamma}} \newcommand{\Itilde}{\widetilde{I}} \newcommand{\nbigbtilde}{\widetilde{\nbigb}} \newcommand{\vecwhat}{\widehat{\vecw}} \newcommand{\pihat}{\widehat{\pi}} \newcommand{\Omegabar}{\overline{\Omega}} \newcommand{\Fhat}{\widehat{F}} \newcommand{\nbigfhat}{\widehat{\nbigf}} \newcommand{\Ghat}{\widehat{G}} \newcommand{\puebar}{\overline{p}} \newcommand{\Rhat}{\widehat{R}} \newcommand{\vecuhat}{\widehat{\vecu}} \newcommand{\Ohat}{\widehat{O}} \newcommand{\Sbar}{\overline{S}} \newcommand{\nbigehat}{\widehat{\nbige}} \newcommand{\Phihat}{\widehat{\Phi}} \newcommand{\nbigyhat}{\widehat{\nbigy}} \newcommand{\nbigvtilde}{\widetilde{\nbigv}} \newcommand{\nbigvhat}{\widehat{\nbigv}} \newcommand{\ellsitabar}{\underline{\ell}} \newcommand{\Utilde}{\widetilde{U}} \newcommand{\Dtilde}{\widetilde{D}} \newcommand{\Xtilde}{\widetilde{X}} \newcommand{\vecy}{\boldsymbol y} \newcommand{\Zhat}{\widehat{Z}} \newcommand{\What}{\widehat{W}} \newcommand{\DDhat}{\widehat{\DD}} \newcommand{\nbigrtilde}{\widetilde{\nbigr}} \newcommand{\nbigghat}{\widehat{\nbigg}} \newcommand{\Ltilde}{\widetilde{L}} \newcommand{\nbigmtilde}{\widetilde{\nbigm}} \def\Der{\mathop{\rm Der}\nolimits} \def\Stab{\mathop{\rm Stab}\nolimits} \def\ord{\mathop{\rm ord}\nolimits} \def\Herm{\mathop{\rm Herm}\nolimits} \def\full{\mathop{\rm full}\nolimits} \def\MT{\mathop{\rm MT}\nolimits} \def\MTW{\mathop{\rm MTW}\nolimits} \def\MTN{\mathop{\rm MTN}\nolimits} \def\wild{\mathop{\rm wild}\nolimits} \def\Gal{\mathop{\rm Gal}\nolimits} \def\strict{\mathop{\rm strict}\nolimits} \def\pol{\mathop{\rm pol}\nolimits} \def\Can{\mathop{\rm Can}\nolimits} \def\Var{\mathop{\rm Var}\nolimits} \def\moderate{\mathop{\rm mod}\nolimits} \def\op{\mathop{\rm op}\nolimits} \def\Mod{\mathop{\rm Mod}\nolimits} \def\ramification{\mathop{\rm rami}\nolimits} \def\pureimaginary{\mathop{\rm pi}\nolimits} \def\Hol{\mathop{\rm Hol}\nolimits} \def\Sec{\mathop{\rm Sec}\nolimits} \def\Sep{\mathop{\rm Sep}\nolimits} \def\sing{\mathop{\rm sing}\nolimits} \def\Gys{\mathop{\rm Gys}\nolimits} \def\exchange{\mathop{\rm exchange}\nolimits} \def\HS{\mathop{\rm HS}\nolimits} \def\TNIL{\mathop{\rm TNIL}\nolimits} \def\Glue{\mathop{\rm Glue}\nolimits} \def\Seg{\mathop{\rm Seg}\nolimits} \def\El{\mathop{\rm El}\nolimits} \def\Sing{\mathop{\rm Sing}\nolimits} \def\Ray{\mathop{\rm Ray}\nolimits} \def\Arc{\mathop{\rm Arc}\nolimits} \def\Lift{\mathop{\rm Lift}\nolimits} \def\slope{\mathop{\rm slope}\nolimits} \def\dec{\mathop{\rm dec}\nolimits} \def\local{\mathop{\rm local}\nolimits} \def\DR{\mathop{\rm DR}\nolimits} \def\ob{\mathop{\rm ob}\nolimits} \def\Mor{\mathop{\rm Mor}\nolimits} \def\Ch{\mathop{\rm Ch}\nolimits} \def\good{\mathop{\rm good}\nolimits} \def\Map{\mathop{\rm Map}\nolimits} \def\hol{\mathop{\rm hol}\nolimits} \def\Loc{\mathop{\rm Loc}\nolimits} \def\Tor{\mathop{\rm Tor}\nolimits} \def\Cat{\mathop{\rm Cat}\nolimits} \def\rapid{\mathop{\rm rap}\nolimits} \def\Conv{\mathop{\rm Conv}\nolimits} \def\FL{\mathop{\rm FL}\nolimits} \def\loc{\mathop{\rm loc}\nolimits} \def\aff{\mathop{\rm aff}\nolimits} \def\Rad{\mathop{\rm Rad}\nolimits} \def\MTM{\mathop{\rm MTM}\nolimits} \def\Eff{\mathop{\rm Eff}\nolimits} \def\QDM{\mathop{\rm QDM}\nolimits} \def\vir{\mathop{\rm vir}\nolimits} \def\ev{\mathop{\rm ev}\nolimits} \def\Proj{\mathop{\rm Proj}\nolimits} \def\sp{\mathop{\rm sp}\nolimits} \def\integral{\mathop{\rm int}\nolimits} \def\alg{\mathop{\rm alg}\nolimits} \def\nor{\mathop{\rm nor}\nolimits} \def\GKZ{\mathop{\rm GKZ}\nolimits} \def\MHM{\mathop{\rm MHM}\nolimits} \newcommand{\Ubar}{\overline{U}} \newcommand{\Dbar}{\overline{D}} \newcommand{\Pbar}{\overline{P}} \newcommand{\Ibar}{\overline{I}} \newcommand{\Wtilde}{\widetilde{W}} \newcommand{\Ptilde}{\widetilde{P}} \newcommand{\psihat}{\widehat{\psi}} \newcommand{\iotahat}{\widehat{\iota}} \newcommand{\vecR}{{\boldsymbol R}} \newcommand{\vbar}{\overline{v}} \newcommand{\ubar}{\overline{u}} \newcommand{\Tbar}{\overline{T}} \newcommand{\gminibbar}{\overline{\gminib}} \newcommand{\vecj}{{\boldsymbol j}} \newcommand{\vecD}{{\boldsymbol D}} \newcommand{\nbigehatlambda}{\widehat{\nbige}^{\lambda}} \newcommand{\nbigelambdahat}{\nbigehatlambda} \newcommand{\nbigutilde}{\widetilde{\nbigu}} \newcommand{\nbigstilde}{\widetilde{\nbigs}} \newcommand{\tauhat}{\widehat{\tau}} \newcommand{\That}{\widehat{T}} \newcommand{\Phat}{\widehat{P}} \newcommand{\nbigmhat}{\widehat{\nbigm}} \newcommand{\nbigzhat}{\widehat{\nbigz}} \newcommand{\Nhat}{\widehat{N}} \newcommand{\gbigehat}{\widehat{\gbige}} \newcommand{\gtilde}{\widetilde{g}} \newcommand{\Ztilde}{\widetilde{Z}} \newcommand{\nbigttilde}{\widetilde{\nbigt}} \newcommand{\gbigstilde}{\widetilde{\gbigs}} \newcommand{\varphitilde}{\widetilde{\varphi}} \newcommand{\Ctilde}{\widetilde{C}} \newcommand{\nbigntilde}{\widetilde{\nbign}} \newcommand{\gbigetilde}{\widetilde{\gbige}} \newcommand{\gbigctilde}{\widetilde{\gbigc}} \newcommand{\Tzero}{T^{(\lambda_0)}} \newcommand{\Ttildezero}{\widetilde{T}^{(\lambda_0)}} \newcommand{\Stilde}{\widetilde{S}} \newcommand{\tautilde}{\widetilde{\tau}} \newcommand{\gbigelambda}{\gbige^{\lambda}} \newcommand{\ybar}{\overline{y}} \newcommand{\epsilonbar}{\overline{\epsilon}} \newcommand{\Atilde}{\widetilde{A}} \newcommand{\Otilde}{\widetilde{O}} \newcommand{\gbigrtilde}{\widetilde{\gbigr}} \newcommand{\gbigmtilde}{\widetilde{\gbigm}} \newcommand{\gbigltilde}{\widetilde{\gbigl}} \newcommand{\Deltatilde}{\widetilde{\Delta}} \newcommand{\nbigftilde}{\widetilde{\nbigf}} \newcommand{\nbigtzero}{\nbigt^{(\lambda_0)}} \newcommand{\nbigpzero}{\nbigp^{(\lambda_0)}} \newcommand{\Mtilde}{\widetilde{M}} \newcommand{\Ytilde}{\widetilde{Y}} \newcommand{\Ybar}{\overline{Y}} \newcommand{\wtilde}{\widetilde{w}} \newcommand{\ztilde}{\widetilde{z}} \newcommand{\nbiglbar}{\overline{\nbigl}} \newcommand{\nbigkbar}{\overline{\nbigk}} \newcommand{\nbigrmod}{\nbigr-\Mod} \newcommand{\itilde}{\widetilde{i}} \newcommand{\nbigxtilde}{\widetilde{\nbigx}} \newcommand{\gbigttilde}{\widetilde{\gbigt}} \newcommand{\nbigvlambda}{\nbigv^{\lambda}} \newcommand{\nbigztilde}{\widetilde{\nbigz}} \newcommand{\nbigdtilde}{\widetilde{\nbigd}} \newcommand{\Qtilde}{\widetilde{Q}} \newcommand{\Ttilde}{\widetilde{T}} \newcommand{\gbigvtilde}{\widetilde{\gbigv}} \newcommand{\gbigntilde}{\widetilde{\gbign}} \newcommand{\what}{\widehat{w}} \newcommand{\uhat}{\widehat{u}} \newcommand{\nbigsbar}{\overline{\nbigs}} \newcommand{\betabar}{\overline{\beta}} \newcommand{\vecatilde}{\widetilde{\veca}} \newcommand{\gminicbar}{\overline{\gminic}} \newcommand{\forget}{for} \newcommand{\nbigitilde}{\widetilde{\nbigi}} \newcommand{\Secbar}{\overline{\Sec}} \newcommand{\Multisector}{\nbigm\nbigs} \newcommand{\SD}{\nbigs\nbigd} \newcommand{\PSD}{\nbigp\nbigs\nbigd} \newcommand{\Irrbar}{\overline{\Irr}} \newcommand{\psitabar}{\underline{p}} \newcommand{\nbigktilde}{\widetilde{\nbigk}} \newcommand{\Ssitabar}{\underline{S}} \newcommand{\vecctilde}{\widetilde{\vecc}} \newcommand{\vecnbigi}{{\boldsymbol \nbigi}} \newcommand{\vectheta}{{\boldsymbol \theta}} \newcommand{\DDprimelambda}{\DD^{\prime\lambda}} \newcommand{\Lambdatilde}{\widetilde{\Lambda}} \newcommand{\nbigmlambda}{\nbigm^{\lambda}} \newcommand{\nrhom}{R{\mathcal Hom}} \newcommand{\DDD}{\boldsymbol D} \newcommand{\jtilde}{\widetilde{j}} \newcommand{\vecH}{{\boldsymbol H}} \newcommand{\nbiggzero}{\nbigg^{(\lambda_0)}} \newcommand{\Dtildehat}{\widehat{\Dtilde}} \newcommand{\ctilde}{\widetilde{c}} \newcommand{\taubar}{\overline{\tau}} \newcommand{\Kbar}{\overline{K}} \newcommand{\Ktilde}{\widetilde{K}} \newcommand{\veckappa}{{\boldsymbol \kappa}} \newcommand{\TTtilde}{{T}\widetilde{T}} \newcommand{\Hhat}{\widehat{H}} \newcommand{\Htilde}{\widetilde{H}} \newcommand{\Omegatilde}{\widetilde{\Omega}} \newcommand{\Ntilde}{\widetilde{N}} \newcommand{\nbigxzero}{\nbigx^{(\lambda_0)}} \newcommand{\nbigdzero}{\nbigd^{(\lambda_0)}} \newcommand{\nutilde}{\widetilde{\nu}} \newcommand{\vecmtilde}{\widetilde{\vecm}} \newcommand{\gminiatilde}{\widetilde{\gminia}} \newcommand{\gminibtilde}{\widetilde{\gminib}} \newcommand{\cnumtilde}{\widetilde{\cnum}} \newcommand{\phibar}{\overline{\phi}} \newcommand{\HSbar}{\overline{\HS}} \newcommand{\phitilde}{\widetilde{\phi}} \newcommand{\nbightilde}{\widetilde{\nbigh}} \newcommand{\Vbar}{\overline{V}} \newcommand{\Wbar}{\overline{W}} \newcommand{\Nbar}{\overline{N}} \newcommand{\DDbar}{\overline{\DD}} \newcommand{\Ebar}{\overline{E}} \newcommand{\Gbar}{\overline{G}} \newcommand{\Fourier}{\mathfrak{Four}} \newcommand{\Xbar}{\overline{X}} \newcommand{\nbigwhat}{\widehat{\nbigw}} \newcommand{\rhohat}{\widehat{\rho}} \newcommand{\varphihat}{\widehat{\varphi}} \newcommand{\gammatilde}{\widetilde{\gamma}} \newcommand{\Gammatilde}{\widetilde{\Gamma}} \newcommand{\Shat}{\widehat{S}} \newcommand{\Xitilde}{\widetilde{\Xi}} \newcommand{\gminiahat}{\widehat{\gminia}} \newcommand{\etahat}{\widehat{\eta}} \newcommand{\Cbar}{\overline{C}} \newcommand{\vecUtilde}{\widetilde{\vecU}} \newcommand{\vecUpsilon}{{\boldsymbol \Upsilon}} \newcommand{\vecUbar}{\overline{\vecU}} \newcommand{\Xibar}{\overline{\Xi}} \newcommand{\gminichat}{\widehat{\gminic}} \newcommand{\piinverseDhat}{\widehat{\pi^{-1}(D)}} \newcommand{\Nilhat}{\widehat{\Nil}} \newcommand{\qtilde}{\widetilde{q}} \newcommand{\veci}{\boldsymbol i} \newcommand{\vecell}{\boldsymbol \ell} \newcommand{\vecq}{\boldsymbol q} \newcommand{\nbigatilde}{\widetilde{\nbiga}} \newcommand{\starbar}{\overline{\star}} \newcommand{\mnuleq}{\prec} \newcommand{\mnugeq}{\succ} \newcommand{\nbigctilde}{\widetilde{\nbigc}} \newcommand{\iotatilde}{\widetilde{\iota}} \newcommand{\nbigqtilde}{\widetilde{\nbigq}} \newcommand{\ktilde}{\widetilde{k}} \newcommand{\nbigptilde}{\widetilde{\nbigp}} \newcommand{\Zbar}{\overline{Z}} \newcommand{\Fbar}{\overline{F}} \newcommand{\nbigubar}{\overline{\nbigu}} \newcommand{\nbigvbar}{\overline{\nbigv}} \newcommand{\QDMbar}{\overline{\QDM}} \newcommand{\Sigmatilde}{\widetilde{\Sigma}} \newcommand{\Btilde}{\widetilde{B}} \newcommand{\gbigptilde}{\widetilde{\gbigp}} \newcommand{\chitilde}{\widetilde{\chi}} \newcommand{\gbigbtilde}{\widetilde{\gbigb}} \newcommand{\Bbar}{\overline{B}} \newcommand{\nbignhat}{\widehat{\nbign}} \newcommand{\distributionbar}{\overline{\distribution}} \newcommand{\elltilde}{\widetilde{\ell}} \newcommand{\distributiontilde}{\widetilde{\distribution}} \newcommand{\Thetatilde}{\widetilde{\Theta}} \newcommand{\DDDtilde}{\widetilde{\DDD}} \newcommand{\Thetabar}{\overline{\Theta}} \newcommand{\nbigcbar}{\overline{\nbigc}} \newcommand{\trtilde}{\widetilde{\tr}} \newcommand{\scrD}{\mathscr{D}} \newcommand{\MTMint}{\MTM^{\rm int}} \begin{thebibliography}{99} \bibitem{Adolphson} A. Adolphson, {\em Hypergeometric functions and rings generated by monomials}, Duke Math. J. {\bf 73}, (1994), 269--290 \bibitem{Adolphson-Sperber} A. Adolphson, S. Sperber, {\em A-hypergeometric systems that come from geometry}, Proc. Amer. Math. Soc. {\bf 140}, (2012), 2033--2042 \bibitem{Batyrev-VMHS} V. V. Batyrev, {\em Variations of the mixed Hodge structure of affine hypersurfaces in algebraic tori}, Duke Math. J. {\bf 69}, (1993), 349--409. \bibitem{beilinson2} A. Beilinson, {\em How to glue perverse sheaves}, in; {\em $K$-theory, arithmetic and geometry (Moscow, 1984--1986)}, Lecture Notes in Math., {\bf 1289}, Springer, Berlin, (1987), 42--51. \bibitem{Borisov-Horja} L. Borisov, P. Horja, {\em On the better behaved version of the GKZ hypergeometric system}, Math. Ann. {\bf 357} (2013), 585--603. \bibitem{Chiang-Klemm-Yau-Zaslow} T.-M. Chiang, A. Klemm, S.-T. Yau, E. Zaslow, {\em Local mirror symmetry: calculations and interpretations}, Adv. Theor. Math. Phys. {\bf 3}, (1999), 495--565. \bibitem{Cox-et-al-toric} D. A. Cox, J. B. Little, H. K. Schenck, {\em Toric varieties}, Graduate Studies in Mathematics, {\bf 124}, American Mathematical Society, Providence, RI, 2011. \bibitem{Danilov-de-Rham} V. I. Danilov, {\em de Rham complex on toroidal variety}, Algebraic geometry (Chicago, IL, 1989), Lecture Notes in Math., {\bf 1479}, Springer, Berlin, (1991), 26--38. \bibitem{Douai-Sabbah1} A. Douai, C. Sabbah, {\em Gauss-Manin systems, Brieskorn lattices and Frobenius structures. I.}, Ann. Inst. Fourier (Grenoble), {\bf 53}, (2003), 1055--1116. \bibitem{Douai-Sabbah2} A. Douai, C. Sabbah, {\em Gauss-Manin systems, Brieskorn lattices and Frobenius structures. II.}, Frobenius manifolds, Aspects Math., E36, Vieweg, Wiesbaden, (2004), 1--18. \bibitem{Esnault-Sabbah-Yu} H. Esnault, C. Sabbah, J.-D. Yu, (with an appendix by M. Saito), {\em $E_1$-degeneration of the irregular Hodge filtration}, arXiv:1302.4537 \bibitem{Fan} H. Fan, {\em Schr\"{o}dinger equations, deformation theory and $tt^*$-geometry}, arXiv:1107.1290 \bibitem{Fulton-toric} W. Fulton, {\em Introduction to toric varieties}, Annals of Mathematics Studies, {\bf 131}, Princeton University Press, Princeton, NJ, 1993. \bibitem{Hertling} C. Hertling, {\em $tt^{\ast}$geometry, Frobenius manifolds, their connections, and the construction for singularities}, J. Reine Angew. Math. {\bf 555}, (2003), 77--161. \bibitem{GKZ} I. M. Gelʹfand, M. M. Kapranov, A. V. Zelevinsky, {\em Generalized Euler integrals and A-hypergeometric functions.} Adv. Math. {\bf 84}, (1990), 255--271. \bibitem{GKZ-book} I. M. Gelʹfand, M. M. Kapranov, A. V. Zelevinsky, {\em Discriminants, resultants, and multidimensional determinants}, Birkh\"{a}user Boston, Inc., Boston, MA, 1994. \bibitem{Givental} A. Givental, {\em A mirror theorem for toric complete intersections}, In {\em Topological field theory, primitive forms and related topics (Kyoto, $1996$)}, Progr. Math., {\bf 160}, Birkh\"auser Boston, Boston, MA, (1998) 141--175. \bibitem{Hertling-Manin} C. Hertling, Y. Manin, {\em Unfoldings of meromorphic connections and a construction of Frobenius manifolds}, in {\em Frobenius manifolds}, Aspects Math., {\bf E36}, Vieweg, Wiesbaden, (2004), 113--144. \bibitem{Hotta-Takeuchi-Tanisaki} R. Hotta, K. Takeuchi, T. Tanisaki, {\em D-modules, perverse sheaves, and representation theory}, Progress in Mathematics, {\bf 236}, Birkh\"{a}user Boston, Inc., Boston, MA, 2008. \bibitem{Iritani-convergence} H. Iritani, {\em Convergence of quantum cohomology by quantum Lefschetz}, J. Reine Angew. Math. {\bf 610}, (2007), 29--69. \bibitem{Iritani} H. Iritani, {\em An integral structure in quantum cohomology and mirror symmetry for toric orbifolds}, Adv. Math. {\bf 222}, (2009), 1016--1079. \bibitem{Iritani-quantum-cohomology-period} H. Iritani, {\em Quantum cohomology and periods}, Ann. Inst. Fourier (Grenoble) {\bf 61} (2011), 2909--2958. \bibitem{Iritani-shift-operators} H. Iritani, {\em Shift operators and toric mirror theorem}, arXiv:1411.6840 \bibitem{Iritani-Mann-Mignon} H. Iritani, E. Mann, T. Mignon, {\em Quantum Serre in terms of quantum D-modules}, arXiv:1412.4523 \bibitem{kashiwara-mixed-Hodge} M. Kashiwara, {\em A study of variation of mixed Hodge structure}, Publ. Res. Inst. Math. Sci. {\bf 22}, (1986) 991--1024. \bibitem{Kashiwara-Schapira} M. Kashiwara and P. Schapira, {\em Sheaves on manifolds}, Springer-Verlag, Berlin, 1990 \bibitem{kashiwara_text} M. Kashiwara, {\em $D$-modules and microlocal calculus,} Translations of Mathematical Monographs, {\bf 217}, American Mathematical Society, Providence, (2003). \bibitem{Katzarkov-Kontsevich-Pantev} L. Katzarkov, M. Kontsevich, T. Pantev, {\em Hodge theoretic aspects of mirror symmetry}, In {\em From Hodge theory to integrability and TQFT $tt^{\ast}$-geometry} Proc. Sympos. Pure Math., {\bf 78}, Amer. Math. Soc., Providence, RI, (2008) 87--174. \bibitem{Konishi-Minabe-cubic} Y. Konishi and S. Minabe, {\em Local Gromov-Witten invariants of cubic surfaces via nef toric degeneration}, Ark. Mat. {\bf 47}, (2009), 345--360. \bibitem{Konishi-Minabe} Y. Konishi, S. Minabe, {\em Local B-model and mixed Hodge structure}, Adv. Theor. Math. Phys. {\bf 14}, (2010), 1089--1145. \bibitem{Konishi-Minabe2} Y. Konishi, S. Minabe, {\em Mixed Frobenius Structure and Local A-model}, arXiv:1209.5550 \bibitem{Konishi-Minabe3} Y. Konishi, S. Minabe, {\em Local Quantum Cohomology and Mixed Frobenius Structure}, arXiv:1405.7476 \bibitem{Kouchnirenko} A. G. Kouchnirenko, {Poly\`{e}dres de Newton et nombres de Milnor}, Invent. Math. {\bf 32}, (1976), 1--31. \bibitem{Mignon-Mann} E. Mann, T. Mignon, {\em Quantum D-modules for toric nef complete intersections}, arXiv:1112.1552 \bibitem{Mochizuki-tame} T. Mochizuki, {\em Asymptotic behaviour of tame harmonic bundles and an application to pure twistor $D$-modules I, II}, Mem. AMS. {\bf 185}, (2007) \bibitem{mochi8} T. Mochizuki, {\em Asymptotic behaviour of variation of pure polarized TERP structures}, Publ. RIMS. {\bf 47} (2011), 419--534. \bibitem{Mochizuki-wild} T. Mochizuki, {\em Wild harmonic bundles and wild pure twistor $D$-modules}, Ast\'{e}risque {\bf 340}, Soci\'{e}t\'{e} Math\'{e}matique de France, Paris, 2011. \bibitem{Mochizuki-Toda-lattice} T. Mochizuki, {\em Harmonic bundles and Toda lattices with opposite sign II}, Comm. Math. Phys. {\bf 328}, 1159--1198, DOI:10.1007/s00220-014-1994-0, the second part of arXiv:1301.1718 \bibitem{Mochizuki-Betti} T. Mochizuki, {\em Holonomic D-modules with Betti structure}, arXiv:1001.2336, to apper in {\em M\'{e}m. Soc. Math. France} \bibitem{Mochizuki-MTM} T. Mochizuki, {\em Mixed twistor $D$-modules}, Springer, 2015. \bibitem{Mochizuki-K-complex} T. Mochizuki, {\em A twistor approach to the Kontsevich complexes}, arXiv:1501.0415 \bibitem{peters-steenbrink} C. Peters and J. Steenbrink, {\em Mixed Hodge structure}, Springer-Verlag, Berlin, 2008. \bibitem{Reichelt} T. Reichelt, {\em Laurent polynomials, GKZ-hypergeometric systems and mixed Hodge modules}, Compos. Math. {\bf 150}, (2014), 911--941 \bibitem{Reichelt-Sevenheck1} T. Reichelt, C. Sevenheck, {\em Logarithmic Frobenius manifolds, hypergeometric systems and quantum D-modules}, arXiv:1010.2118, to appear in J. Alg. Geom. \bibitem{Reichelt-Sevenheck2} T. Reichelt, C. Sevenheck, {\em Non-affine Landau-Ginzburg models and intersection cohomology}, \\ arXiv:1210.6527 \bibitem{Reichelt-Sevenheck3} T. Reichelt, C. Sevenheck, {\em Hypergeometric Hodge modules}, arXiv:1503.01004 \bibitem{sabbah2} C. Sabbah, {\em Polarizable twistor $D$-modules} Ast\'{e}risque, {\bf 300}, (2005) \bibitem{Sabbah-tame-polynomial} C. Sabbah, {\em Hypergeometric periods for a tame polynomial}, Port. Math. (N.S.) {\bf 63}, (2006), 173--226. \bibitem{sabbah-Fourier-Laplace} C. Sabbah, {\em Fourier-Laplace transform of a variation of polarized complex Hodge structure}, J. Reine Angew. Math. {\bf 621}, (2008), 123--158 \bibitem{sabbah5} C. Sabbah, {\em Wild twistor $D$-modules}, in {\em Algebraic analysis and around}, Adv. Stud. Pure Math., {\bf 54}, Math. Soc. Japan, Tokyo, (2009), 293--353. \bibitem{Sabbah-Yu} C. Sabbah, J.-D. Yu. {\em On the irregular Hodge filtration of exponentially twisted mixed Hodge modules}, arXiv:1406.1339 \bibitem{Saito-Takahashi-survey} K. Saito, A. Takahashi, {\em From primitive forms to Frobenius manifolds}, in {\em From Hodge theory to integrability and TQFT tt*-geometry,} 31--48, Proc. Sympos. Pure Math., {\bf 78}, Amer. Math. Soc., Providence, RI, 2008. \bibitem{saito1} M. Saito, {\em Modules de Hodge polarisables}, Publ. RIMS., {\bf 24}, (1988), 849--995. \bibitem{saito-Brieskorn} M. Saito, {\em On the structure of Brieskorn lattice}, Ann. Inst. Fourier (Grenoble), {\bf 39}, (1989), 27--72. \bibitem{saito3} M. Saito, {\em Duality for vanishing cycle functors}, Publ. Res. Inst. Math. Sci. {\bf 25}, (1989), 889--921. \bibitem{saito4} M. Saito, {\em Induced $D$-modules and differential complexes}, Bull. Soc. Math. France {\bf 117}, (1989), 361--387. \bibitem{saito2} M. Saito, {\em Mixed Hodge modules}, Publ. RIMS., {\bf 26}, (1990), 221--333. \bibitem{Schulze-Walther} M. Schulze and U. Walther, {\em Hypergeometric D-modules and twisted Gauss–Manin systems}, J. Algebra {\bf 322}, (2009), 3392--3409. \bibitem{s3} C. Simpson, {\it Mixed twistor structures}, math.AG/9705006. \bibitem{steenbrink-zucker} J. Steenbrink and S. Zucker, {\em Variation of mixed Hodge structure. I}, Invent. Math. {\bf 80}, (1985), 489--542. \bibitem{Stienstra} J. Stienstra, {\em Resonant hypergeometric systems and mirror symmetry}, Integrable systems and algebraic geometry (Kobe/Kyoto, 1997), World Sci. Publ., River Edge, NJ, (1998), 412--452. \bibitem{Yu} J.-D. Yu, {\em Irregular Hodge filtration on twisted de Rham cohomology}, Manuscripta Math. {\bf 144}, (2014), 99--133 \end{thebibliography} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.