arXiv reader

The derived category of a generalized complex torus

Patrick Antweiler, Vicente Cortés, Julian Holstein, Markus Röser and Aldo Witte

arXiv
2610.21960v1 · PDF · 15 October 2026
Categories
math.DG, math-ph, math.AG, math.MP, math.SG
Licence
CC BY 4.0 The site shows the text.

Abstract

For any translation invariant generalized complex structure J on a torus we construct the derived category D(J) generalizing the derived category of coherent sheaves on a complex torus and the affine/convergent Fukaya category of a symplectic torus. The category is defined in terms of a non-commutative curved deformation of the Dolbeault algebra of a complex torus. We show that this category is independent of the chosen deformation path, which we prove using non-commutative Fourier-Mukai duality. We also prove that the action of O(2n,2n; Z) on the space of generalized complex structures leaves the category D(J) invariant. In particular, T-duality acts trivially on D(J). We construct concrete objects associated to affine Lagrangians in the derived category associated to a symplectic torus and show that their spaces of morphisms are given by Floer cohomology. We show that our construction is consistent with the homological mirror symmetry conjecture (including its non-commutative version).

Items 129

Theorem

Theorem

WWGD

Let X be a compact complex manifold, and consider its Dolbeault algebra \mathbf{A} = (\Omega^{0,\bullet}(X),\bar{\partial},0). Then the homotopy category of cohesive modules {\mathcal{P}}_{\mathbf{A}} is equivalent to {\operatorname{D^b_{coh}}}(X).

Theorem

Theorem

RYPZ

\operatorname{Ho} {\mathcal{P}}_{\mathbf{A}(I,0,B)} is equivalent to the bounded derived category of complexes of B-twisted sheaves of {\mathcal{O}}_X-modules with coherent cohomology.

Remark

Remark

3IKE

Let {\mathcal{J}} be a translation invariant GC structure on V of even cotype. There exist B\in\scalebox{1.2}{}^2 V^*, \beta\in \scalebox{1.2}{}^2 V and J\in\mathcal C_V, such that

{\mathcal{J}} = e^{\beta } e^{B } {\mathcal{J}}_J e^{-B } e^{-\beta } .
Remark

Remark

RCPQ

Let {\mathcal{J}} be a translation-invariant GC structure on V/\Gamma of even cotype. Choose I, \beta and B such that {\mathcal{J}} = e^{\beta } e^{B } {\mathcal{J}}_I e^{-B} e^{-\beta}. The derived category of {\mathcal{J}} is defined as

{\mathcal{D}}({\mathcal{J}}) = {\mathcal{P}}_{\mathbf{A}(I,\beta,B)}.

If {\mathcal{J}} is of odd cotype we let {\mathcal{J}}' be the GC structure obtained by performing T-duality relative to any circle. We define {\mathcal{D}}({\mathcal{J}}) := {\mathcal{D}}({\mathcal{J}}').

Remark

Remark

XI6Q

Let \mathcal J be a translation-invariant GC structure of even cotype on a torus. Then the generalized derived category {\mathcal{D}}({\mathcal{J}}) is well-defined up to A_\infty-quasi-equivalence.

Remark

Remark

9V8I

The generalized derived category of a 2n-dimensional torus is well-defined also for translation invariant GC structures of odd cotype and satisfies integral invariance, i.e. it is invariant under the natural action of SO(2n,2n; \mathbb{Z}) on GC structures.

Remark

Remark

2XSH

Given a torus with translation invariant B-symplectic form \omega + \mathbf{i} B and associated GC structure {\mathcal{J}}_{\omega+\mathbf{i} B}= e^{B} {\mathcal{J}}_\omega e^{-B} its Fukaya category is given by the derived category of the mirror of {\mathcal{J}}_{\omega + \mathbf{i} B}.

Remark

Remark

S9HS

With notation as above there is an isomorphism:

H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}(\tilde N(L,\alpha),\tilde N(\tilde L, \tilde \alpha))) \cong HF^\bullet( ( L,\alpha), (\tilde L, \tilde \alpha) ).
Definition

Definition 2.1

KFM4

Let M be a smooth manifold. A Courant algebroid over M is a vector bundle E\to M together with a non-degenerate scalar product \langle\cdot,\cdot\rangle\in\Gamma(\operatorname{Sym}^2E^*), a bilinear bracket [\cdot,\cdot]\colon \Gamma(E)\times \Gamma(E)\to\Gamma(E) and a bundle map \rho\colon E\to TM, called the anchor, such that for all u,v,w\in\Gamma(E) the following holds:

  • [u,[v,w]] = [[u,v],w]+[v,[u,w]],

  • \rho(u)\langle v,w\rangle = \langle [u,v],w\rangle+\langle v,[u,w]\rangle,

  • \langle [u,v]+[v,u],w\rangle = \rho(w)\langle u,v\rangle.

Definition

Definition 2.2

0X93

The generalized tangent bundle is the Courant algebroid with vector bundle

E = \mathbb{T} M = TM\oplus T^*M

with anchor given by the natural projection \rho \colon\mathbb{T}M \rightarrow TM, scalar product given by the natural pairing

\langle v+\xi, v+\xi\rangle = \xi(v)

and bracket given by the Dorfman bracket

[X+\xi, Y+\eta] = \mathcal L_X(Y+\eta)-\iota_Yd\xi,

where X,Y\in \Gamma(TM), \xi,\eta\in\Gamma(T^*M) = \Omega^1(M).

Definition

Definition 2.3

6XOL

Let M be a smooth manifold. A generalized complex (GC) structure on M is an element {\mathcal{J}}\in \Gamma(\operatorname{End}(\mathbb{T} M)) such that

  1. {\mathcal{J}}^2 = -\textbf{\textit{I}}_{\mathbb{T} M},

  2. {\mathcal{J}} is skew-symmetric with respect to \langle\cdot,\cdot\rangle,

  3. {\mathcal{J}} is integrable in the sense that the +\mathbf{i}-eigenbundle L_{{\mathcal{J}}} = \mathrm{Eig}({\mathcal{J}},\mathbf{i}) \subset \mathbb{T} M _\mathbb{C} is involutive, i.e.

    [\Gamma(L_{{\mathcal{J}}}),\Gamma(L_{{\mathcal{J}}})]\subset \Gamma(L_{{\mathcal{J}}}). \hfill \qedhere
Remark

Remark 2.0

IP1M

If the manifold M admits a generalized (almost) complex structure the dimension of M must be even.

Definition

Definition 2.4

4JHI

Let {\mathcal{J}}\in\Gamma(\operatorname{End}(\mathbb TM)) be a GC structure. The type of {\mathcal{J}} at a point p\in M is the number

\mathrm{type}({\mathcal{J}})_p = {\operatorname{codim}}_{TM_\mathbb{C}} (\pi(L_{{\mathcal{J}}}))\in \{0,1,\dots,\frac{1}{2}\dim_{\mathbb{R}} M\}

We say {\mathcal{J}} is of type k\in \{0,\dots,\frac{1}{2}\dim_{\mathbb{R}} M\} if \mathrm{type}({\mathcal{J}})_p = k for all p\in M.

We also define the cotype of \mathcal J at p as

\mathrm{cotype}({\mathcal{J}})_p = \dim_{\mathbb C}(\pi(L_{\mathcal J})) - \frac{1}{2}\dim_{\mathbb{R}} M = \frac{1}{2}\dim_{\mathbb{R}} M - \mathrm{type}({\mathcal{J}})_p \in \{0,1,\dots,\frac{1}{2}\dim_{\mathbb{R}} M\} . \hfill \qedhere
Remark

Remark 2.0

GZ6M

The type of a GC structure is invariant under B-field transformations, since these are automorphisms of the Courant algebroid structure on \mathbb{T} M.

Example

Example 2.5

BQ9D

There are two important examples of GC structures of minimal, respectively maximal, possible type:

  1. Let J\in \Gamma(\operatorname{End}(TM)) be a complex structure on M (in the ordinary sense). Then we obtain a GC structure via

    {\mathcal{J}}_J = \begin{pmatrix} J & 0 \\ 0 & -J^* \end{pmatrix}.

    The corresponding maximal isotropic subbundle of \mathbb{T}M_\mathbb{C} is given by

    L_{{\mathcal{J}}_J} = \left(\textbf{\textit{I}}_{\mathbb TM}-\mathbf{i}{\mathcal{J}}_J\right)(\mathbb TM) = TM^{1,0} + T^*M^{0,1}\subset \mathbb{T} M_\mathbb{C}.

    We have

    \pi(L_{{\mathcal{J}}_J}) = TM^{1,0},

    which has codimension n= \dim_{\mathbb{C}}(M) = \frac{1}{2}\dim_{\mathbb{R}}M in TM_\mathcal{C}. Thus, {\mathcal{J}}_J is of type n = \dim_\mathbb{C} M and hence of cotype 0. One can show that any GC structure {\mathcal{J}} of this type is a B-field transformation of a complex structure, i.e. of the form

    {\mathcal{J}} = e^{B } {\mathcal{J}}_J e^{-B }

    with corresponding isotropic subbundle

    e^{B } L_{{\mathcal{J}}_J}.
  2. Let \omega\in\Omega^2(M) be a symplectic form, i.e. d\omega =0 and \omega is non-degenerate. Then

    {\mathcal{J}}_\omega = \begin{pmatrix} 0 & -\omega^{-1}\\ \omega & 0 \end{pmatrix}

    defines a GC structure with associated subbundle

    L_{{\mathcal{J}}_\omega} =\mathrm{graph}(-\mathbf{i}\omega ) = \{X-\mathbf{i}\omega (X)~:~ X\in TM_\mathbb{C}\}\subset \mathbb{T} M_\mathbb{C}.

    We have

    \pi(L_{{\mathcal{J}}_\omega}) = TM_{\mathbb{C}},

    thus {\mathcal{J}}_\omega is of type 0 (and thus of cotype n= \frac{1}{2}\dim_{\mathbb{R}}M). Any GC structure {\mathcal{J}} of type 0 is a B-field transform of a symplectic structure {\mathcal{J}}_\omega, i.e. is of the form

    {\mathcal{J}} = e^{B } {\mathcal{J}}_\omega e^{-B }

    with corresponding maximal isotropic subbundle

    L_{{\mathcal{J}}} = e^{B } L_{{\mathcal{J}}_\omega} = \mathrm{graph}(-\mathbf{i}(\omega+\mathbf{i} B)).

    Due to the last equality, we will sometimes use the notation {\mathcal{J}}_{\omega + \mathbf{i} B} = e^{B } {\mathcal{J}}_\omega e^{-B }.

Remark

Remark 2.0

UB4Z

We now explicitly describe certain kinds of symmetry transformations that act on \mathcal{GC}_V. Let

\mathcal{J} = \begin{pmatrix} A & \pi \\ \omega & -A^* \end{pmatrix}\in \mathcal{GC}_V,

where A\in\operatorname{End}(V), \omega \in \scalebox{1.2}{}^2V^*, \pi \in\scalebox{1.2}{}^2V.

  1. We have already encountered B-field transformations. To a 2-form B\in\scalebox{1.2}{}^2V^* we associate the element

    e^{B } = \begin{pmatrix} 1 & 0 \\ B & 1 \end{pmatrix}\in\mathrm{O}(\mathbb{V}),

    which acts on {\mathcal{J}} by

    e^{B } {\mathcal{J}} e^{-B } = \begin{pmatrix} A-\pi B & \pi \\ B A+A^*B +\omega -B \pi B & B \pi -A^* \end{pmatrix}.
  2. We shall also encounter \beta-transformations. To a bivector \beta\in\scalebox{1.2}{}^2V we associate the element

    e^{\beta } = \begin{pmatrix} 1 & \beta \\ 0 & 1 \end{pmatrix}\in\mathrm{O}(\mathbb{V}),

    which acts on {\mathcal{J}} by

    e^{\beta } {\mathcal{J}} e^{-\beta } = \begin{pmatrix} A+\beta \omega & \pi -A\beta -\beta A^*-\beta \omega \beta \\ \omega & -A^*-\omega \beta \end{pmatrix}. \hfill\qedhere
Used by Proof
Theorem

Theorem 2.9 (Generalized Darboux theorem, Section 4.7)

0KKR

Any GC structure {\mathcal{J}} on \mathbb{V} = V \oplus V^* is a B-field transformation of a direct sum {\mathcal{J}}_1 \oplus {\mathcal{J}}_2, where {\mathcal{J}}_1 = {\mathcal{J}}_J \in \mathrm{O}(V_1 \oplus V_1^*) and {\mathcal{J}}_2 = {\mathcal{J}}_\omega \in \mathrm{O}(V_2 \oplus V_2^*) are associated with a complex structure J on V_1 and a symplectic structure \omega on V_2 for some splitting V = V_1 \oplus V_2. The type of {\mathcal{J}} is \dim_\mathbb{C} (V_1,J).

Proposition

Proposition 2.7

7U6K

Let {\mathcal{J}}\in\mathcal{GC}_V be a GC structure of even cotype.

  1. There exist B\in\scalebox{1.2}{}^2V^*, \beta\in\scalebox{1.2}{}^2 V and J\in\mathcal C_V, such that

    {\mathcal{J}} = e^{\beta } e^{B } {\mathcal{J}}_J e^{-B } e^{-\beta } .
  2. There exist B\in\scalebox{1.2}{}^2V^*, \beta\in\scalebox{1.2}{}^2 V and J\in\mathcal C_V, such that

    {\mathcal{J}} = e^{B } e^{\beta } {\mathcal{J}}_J e^{-\beta } e^{-B } .
Proof of Proposition 2.7
  1. By the generalized Darboux theorem, we can find \tilde{B}\in\scalebox{1.2}{}^2V^* and a splitting V= V_1\oplus V_2, a complex structure J_1\in\mathcal C_{V_1} and a symplectic form \omega_2\in\scalebox{1.2}{}^2 V_2^* such that

    e^{\tilde{B} } {\mathcal{J}} e^{-\tilde{B} } = {\mathcal{J}}_{J_1}\oplus {\mathcal{J}}_{\omega_2}.

    Since the cotype of {\mathcal{J}} is even, it follows that the complex dimension of V_2 is even and therefore Proposition 2.10 allows us to find B_2\in\scalebox{1.2}{}^2 V_2^*, \beta_2\in \scalebox{1.2}{}^2 V_2 and J_2\in\mathcal C_{V_2} such that

    {\mathcal{J}}_{\omega_2}= e^{B _2} e^{\beta_2 } {\mathcal{J}}_{J_2} e^{-\beta_2 } e^{-B _2} \in\mathcal{GC}_{V_2}.

    Using the splitting V = V_1\oplus V_2 and extending by zero on V_2, respectively V_2^*, we may view B _2,\beta_2 as elements of \mathfrak{so}(\mathbb{V}). With this interpretation we can write

    {\mathcal{J}} = e^{-\tilde{B} +B _2} e^{\beta_2 } {\mathcal{J}}_{J_1\oplus J_2} e^{-\beta_2 } e^{\tilde{B} -B_2 } .
  2. Exchanging the role of V and V^* in the generalized Darboux theorem, we see that there also exists \tilde{\beta} \in\scalebox{1.2}{}^2V such that

    e^{\tilde{\beta} } {\mathcal{J}} e^{-\tilde{\beta} } = {\mathcal{J}}_{J_1}\oplus {\mathcal{J}}_{\omega_2}.

    Note that although \beta-transforms do not preserve the cotype in general, they still preserve the parity of the cotype (see [Gualtieri:2003dx]).

Proposition

Proposition 2.8

3KF0

Suppose J,J'\in \mathcal{C}_V are complex structures and \omega,\omega'\in\scalebox{1.2}{}^2V^* are symplectic forms. Then for any B\in \scalebox{1.2}{}^2V^*,\beta\in\scalebox{1.2}{}^2V we have

  1. e^{B } {\mathcal{J}}_\omega e^{-B } = {\mathcal{J}}_{\omega'} if and only if B=0 and \omega=\omega'.

  2. e^{B } {\mathcal{J}}_J e^{-B } = {\mathcal{J}}_{J'} if and only if J=J' and B\in\scalebox{1.2}{}^{1,1}_JV^*.

  3. e^{\beta } {\mathcal{J}}_J e^{-\beta } = {\mathcal{J}}_{J'} if and only if J=J' and \beta\in\scalebox{1.2}{}^{1,1}_JV.

  4. e^{\beta } {\mathcal{J}}_\omega e^{-\beta } = {\mathcal{J}}_{\omega'} if and only if \beta=0 and \omega=\omega'.

In particular: If the equation

e^{\beta } e^{B } {\mathcal{J}}_\omega e^{-B } e^{-\beta } = {\mathcal{J}}_J

admits a solution (B,\beta), then B is unique and \beta is unique up to addition of a bivector of type (1,1) and if the equation

e^{\beta } e^{B } {\mathcal{J}}_J e^{-B } e^{-\beta } = {\mathcal{J}}_\omega

admits a solution (B,\beta), then \beta is unique and B is unique up to addition of a form of type (1,1).

Proposition

Proposition 2.9

FX62

Let \omega\in\scalebox{1.2}{}^2V^* be a symplectic form and let J\in \mathcal{C}_V be a complex structure on V.

  1. The equation

    e^{\beta } e^{B } {\mathcal{J}}_\omega e^{-B } e^{-\beta } = {\mathcal{J}}_J

    admits a solution (B,\beta)\in \scalebox{1.2}{}^2V^*\times\scalebox{1.2}{}^2V if and only if \omega\in (\scalebox{1.2}{}^{2,0}_JV^*+\scalebox{1.2}{}^{0,2}_JV^*)\cap\scalebox{1.2}{}^2V^*. In this case B lies in the subspace (\scalebox{1.2}{}^{2,0}_JV^*+\scalebox{1.2}{}^{0,2}_JV^*)\cap\scalebox{1.2}{}^2V^* and is uniquely determined by

    B = \omega J.

    The general form of \beta is

    \beta = -\frac{1}{2} B^{-1} + \tau

    for an arbitrary \tau\in\scalebox{1.2}{}^{1,1}_JV.

  2. The equation

    e^{\beta } e^{B } {\mathcal{J}}_J e^{-B } e^{-\beta } = {\mathcal{J}}_\omega

    admits a solution (B,\beta)\in \scalebox{1.2}{}^2V^*\times\scalebox{1.2}{}^2V if and only if \omega\in (\scalebox{1.2}{}^{2,0}_JV^*+\scalebox{1.2}{}^{0,2}_JV^*)\cap\scalebox{1.2}{}^2V^*. In this case \beta lies in the subspace (\scalebox{1.2}{}^{2,0}_JV+\scalebox{1.2}{}^{0,2}_JV)\cap\scalebox{1.2}{}^2V and is uniquely determined by

    \beta = -J \omega^{-1} = (\omega J)^{-1}.

    The general form of B\in\scalebox{1.2}{}^2V^* is

    B = -\frac{1}{2} \beta^{-1}+ B

    for an arbitrary b\in\scalebox{1.2}{}^{1,1}_JV^*.

Used by Example 2.11
Proof of Proposition 2.9
  1. Remark 2.0 implies in particular for a symplectic form \omega\in \scalebox{1.2}{}^2V^*:

    \begin{aligned} & e^{\beta } e^{B } {\mathcal{J}}_\omega e^{-B } e^{-\beta } = e^{\beta } \begin{pmatrix} \omega^{-1} B & -\omega^{-1} \\ \omega +B \omega^{-1} B & -B \omega^{-1}\end{pmatrix} e^{-\beta } \\ &= \begin{pmatrix} \omega^{-1} B +\beta (\omega +B \omega^{-1} B ) & -\omega^{-1}-\omega^{-1} B \beta -\beta B \omega^{-1}-\beta (\omega +B \omega^{-1} B )\beta \\ \omega +B \omega^{-1} B & -B \omega^{-1}-(\omega +B \omega^{-1} B )\beta \end{pmatrix} \end{aligned}

    It follows that the equation

    e^{\beta } e^{B } {\mathcal{J}}_\omega e^{-B } e^{-\beta } = {\mathcal{J}}_{J}\quad \text{for some } J\in \mathcal{C}_V

    is equivalent to the following conditions:

    \begin{aligned} 0&= \omega +B \omega^{-1}B \\ 0&= \omega^{-1}+\omega^{-1} B\beta +\beta B\omega^{-1}+\beta (\omega+B\omega^{-1} B)\beta . \end{aligned}

    If these are satisfied, then

    e^{\beta} e^{B} {\mathcal{J}}_\omega e^{-B} e^{-\beta} = \begin{pmatrix} \omega^{-1} B & 0 \\ 0 & -B\omega^{-1} \end{pmatrix}.

    In particular

    J := \omega^{-1} B,

    is a complex structure on V and satisfies

    J^2 = -1,\quad J^* = B\omega^{-1}

    and

    B = \omega J.

    Note that for this to be a two-form, i.e. B^* = - B, we must require

    -\omega J = (\omega J)^* = J^*\omega^* = -J^*\omega ,

    i.e.

    J^*\omega J = -\omega ,

    which implies that also

    J^*B J=-B .

    In particular,

    B,\omega \in (\scalebox{1.2}{$\wedge$}^{2,0}_JV^*+\scalebox{1.2}{$\wedge$}^{0,2}_JV^*)\cap \scalebox{1.2}{$\wedge$}^2V^*.

    Note that if (?) holds, then (?) simplifies to

    0 = \omega^{-1}+\omega^{-1}B \beta +\beta B \omega^{-1}.

    With the above notation, (?) becomes

    \omega^{-1} = -J\beta -\beta J^*,

    which we may rewrite in the form

    J B^{-1} = -J\beta -\beta J^*,

    or, after multiplying both sides by -J = J^{-1} from the left,

    B^{-1} = -\beta +J\beta J^*.

    In other words, we see that B\in (\scalebox{1.2}{}^{2,0}_JV^*+\scalebox{1.2}{}^{0,2}_JV^*)\cap \scalebox{1.2}{}^2V^* is non-degenerate and - B^{-1} is given by (two times) the component of \beta in (\scalebox{1.2}{}^{2,0}_JV+\scalebox{1.2}{}^{0,2}_JV)\cap \scalebox{1.2}{}^2V , which is given by

    \beta^{(2,0)}+\beta^{(0,2)} = \frac{1}{2}(\beta -J\beta J^*).
  2. Remark 2.0 implies in particular

    \begin{aligned} & e^{\beta } e^{B } {\mathcal{J}}_J e^{-B } e^{-\beta } = e^{\beta } \begin{pmatrix} J & 0 \\ B J+J^*B & -J^*\end{pmatrix} e^{-\beta } \\ &= \begin{pmatrix} J + \beta(B J+J^*B ) & -J\beta-\beta J^*-\beta(B J+J^*B )\beta \\ B J+J^*B & -J^*-(B J+J^*B )\beta\end{pmatrix} \end{aligned}

    It follows that

    e^{\beta } e^{B } {\mathcal{J}}_{J} e^{-B } e^{-\beta } = {\mathcal{J}}_\omega

    for some nondegenerate \omega\in\scalebox{1.2}{}^2V^* is equivalent to the equation

    \begin{aligned} 0&= J + \beta (B J+J^*B ) \end{aligned}

    If this equation is satisfied, then

    e^{\beta } e^{B } {\mathcal{J}}_{J} e^{-B } e^{-\beta } = \begin{pmatrix} 0& -J\beta -\beta J^*-\beta (B J+J^*B )\beta \\ B J+J^*B & 0 \end{pmatrix}.

    Equating this with {\mathcal{J}}_\omega and looking at the lower left entry, we find

    \omega = B J+J^*B \in(\scalebox{1.2}{$\wedge$}^{2,0}_JV^*+\scalebox{1.2}{$\wedge$}^{0,2}_JV^*)\cap \scalebox{1.2}{$\wedge$}^2V^*.

    It follows that

    -B +J^*B J = \omega J.

    The left hand side of this equation is (minus two times) the projection of B onto the subspace (\scalebox{1.2}{}^{2,0}_JV^*+\scalebox{1.2}{}^{0,2}_JV^*)\cap \scalebox{1.2}{}^2V^*. And we see that we are free to add to B any two-form b\in\scalebox{1.2}{}^{1,1}_JV^*.

    Inserting (?) into (?), we get

    0 = J+\beta \omega ,

    i.e.

    \beta = -J\omega^{-1}. \hfill \qedhere
Proposition

Proposition 2.10

7ZOH

Let V be a real vector space of dimension 2n.

  1. Let \omega\in\scalebox{1.2}{}^2V^* be a symplectic form. Then there exists a complex structure J\in \mathcal{C}_V, a 2-form B\in\scalebox{1.2}{}^2V^* and a bivector \beta\in\scalebox{1.2}{}^2V such that

    e^{\beta } e^{B } {\mathcal{J}}_{\omega} e^{-B } e^{-\beta } = {\mathcal{J}}_J

    if and only if n is even, i.e. \dim V is divisible by 4.

  2. Let J\in\mathcal C_V be a complex structure. Then there exists a symplectic form \omega\in\scalebox{1.2}{}^2V^*, a 2-form B\in\scalebox{1.2}{}^2V^* and a bivector \beta\in\scalebox{1.2}{}^2V such that

    e^{\beta } e^{B } {\mathcal{J}}_{J} e^{-B } e^{-\beta } = {\mathcal{J}}_\omega

    if and only if n is even, i.e. \dim V is divisible by 4.

Used by Proof
Proof of Proposition 2.10

The discussion above shows that if J\in\mathcal C_V and a solution (B,\beta) exist, then B = \omega J and this is a two-form iff \omega\in (\scalebox{1.2}{}^{2,0}_JV^*+\scalebox{1.2}{}^{0,2}_JV^*)\cap V. Extend B,\omega complex bilinearly and consider

\Omega = B+\mathbf{i}\omega \in \scalebox{1.2}{$\wedge$}^2V_\mathbb{C}^*.

Let a,b\in V such that v = a+\mathbf{i} b\in V_{\mathbb{C}} lies in the kernel of \Omega, i.e. \omega (v) = 0. Then

0 = \omega (v) = B (a+\mathbf{i} b)+\mathbf{i}\omega (a+\mathbf{i} b) = B (a)-\omega (b) + \mathbf{i}(B (b)+\omega (a)).

This means that

\omega (b) = B (a) = \omega (Ja)

and

\omega (a) = -B (b) = -\omega (Jb).

By the nondegeneracy of \omega this is equivalent to

b = Ja.

Thus, we see that v = a+\mathbf{i} b\in \ker\Omega is equivalent to v\in V^{0,1} = \mathrm{Eig}(J,-\mathbf{i})\subset V_\mathbb{C}. It follows that \Omega\in\scalebox{1.2}{}^{2,0}V^* is a complex symplectic form, and thus the complex dimension of V must be even.

Conversely, if the dimension of V is divisible by 4 and \omega\in\scalebox{1.2}{}^2V^* is nondegenerate, we may pick a complex structure J and a complex symplectic form \Omega\in\scalebox{1.2}{}^{2,0}V^*, and decompose it into real and imaginary parts \Omega = \tilde{B}+\mathbf{i}\tilde\omega. Then the calculation in the first part of the proof shows that \tilde \omega J = \tilde{B}. Since any two symplectic forms on V are in the same \mathrm{GL}(V)-orbit, we may use the \mathrm{GL}(V) action to achieve \tilde\omega = \omega.

Example

Example 2.11

CKHX

Suppose \dim V = 4n for some n\in\mathbb N. Consider a (pseudo-)Hyperkähler structure (g,I,J,K) on V, i.e. g is a (possibly indefinite) inner product on V and I,J,K\in\mathcal{C}_V are skew-symmetric with respect to g and satisfy the quaternionic relations. Denote by

\omega _I = g I,\quad \omega_J = g J,\quad \omega_K = g K

the associated symplectic forms. The quaternionic relations imply \omega_J,\omega_K \in \scalebox{1.2}{}^{2,0}_IV^*+\scalebox{1.2}{}^{0,2}_IV^* and moreover

\omega_K I= g KI = g J = \omega_J .

It then follows from Proposition 2.9 that

{\mathcal{J}}_{\omega_J} = e^{-\omega_K^{-1}} e^{\frac{\omega _K}{2}} {\mathcal{J}}_I e^{\frac{-\omega _K}{2}} e^{\omega_K^{-1}} .\hfill\qedhere
Definition

Definition 3.1

3FUD

Let M and \tilde{M} be principal T^k-bundles with base B. Consider the fiber-product M\times_B \tilde{M}, and the following diagram:

The paper has a diagram here, which this page does not show.

We say that M and \tilde{M} are T-dual if there exists a closed T^{2k}-invariant two-form F \in \Omega^2_{T^{2k}}(M\times_B\tilde{M}) such that the restriction

F \colon \mathfrak{t}_M \times \mathfrak{t}_{\tilde{M}} \rightarrow \mathbb{R},

where \mathfrak{t}_M,\mathfrak{t}_{\tilde{M}} denotes the tangent space to the fibres of p, \tilde{p} respectively, is non-degenerate. We call M \times_B \tilde{M} the correspondence space.

Theorem

Theorem 3.2

SPWC

If M and \tilde{M} are T-dual, then there is an isomorphism of Courant algebroids

\varphi \colon (TM \oplus T^*M)/T^k \rightarrow (T\tilde{M} \oplus T^*\tilde{M})/T^k.

Moreover, if {\mathcal{J}} is a T^k-invariant GC structure on M it gets transformed to a T^k-invariant GC structure on \tilde{M} denoted \tilde{{\mathcal{J}}}.

Remark

Remark 3.0 (Type change)

ULT0

Given a GC structure {\mathcal{J}} on M there are several ways in which one my predict the type of the T-dual structure \tilde{{\mathcal{J}}}. This is carefully described in Example 4.3 in [CG11]. For us it’s important to remark that when k=1, one has the relation that

\operatorname{type}(\tilde{{\mathcal{J}}}) \pm 1

implying that the parity will always change when performing T-duality along the circle.

Lemma

Lemma 3.3

JDSS

Let

L_{{\mathcal{J}}} = e^{\beta } e^{B } L_{{\mathcal{J}}_I}

be a GC structure on V/\Gamma and

L_{{\mathcal{J}}'} = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I}}

a GC structure on V^* / \Gamma^*. Then \tilde{{\mathcal{J}}} = {\mathcal{J}}' if and only if

\begin{aligned} &\tilde I(1-\tilde{\beta} \beta ) = -(1-\tilde{\beta} \beta ) I^* \\ &(B I + I^* B ) (1 - \beta \tilde{\beta} ) = I^* \tilde{\beta} - \tilde{\beta} \tilde I^* \\ & (\tilde{B} \tilde I + \tilde I^* \tilde{B} ) (1-\tilde{\beta} \beta ) = \tilde I^* \beta - \beta I^* \end{aligned}

In particular, if \alpha^{-1} := (1-\tilde{\beta} \beta ) is invertible, then \tilde{{\mathcal{J}}} = {\mathcal{J}}' if and only if

\begin{aligned} &\tilde I = -\alpha^{-1} I^* \alpha \\ & B^{(0,2)_I} = (\tilde{\beta} \alpha^*)^{(0,2)_I} \\ & \tilde{B} ^{(0,2)_{\tilde I}} = ( \beta \alpha)^{(0,2)_{\tilde I}}. \end{aligned}
Proof.

We compute {\mathcal{J}}:

\begin{aligned} {\mathcal{J}} &= e^{\beta } e^{B } \begin{pmatrix} I & 0 \\ 0 &-I^* \end{pmatrix} e^{-B } e^{-\beta } \\&= \begin{pmatrix} I + \beta (B I+I^*B ) & -I\beta -\beta I^*-\beta (B I+I^*B )\beta \\ B I+I^*B & -I^*-(B I+I^*B )\beta \end{pmatrix} \end{aligned}

Hence

\begin{aligned} \tilde{{\mathcal{J}}} = \begin{pmatrix} -I^*-(B I+I^*B )\beta & B I+I^*B \\ -I\beta -\beta I^*-\beta (B I+I^*B )\beta & I + \beta (B I+I^*B ) \end{pmatrix} \end{aligned}

needs to be equal to

\begin{aligned} {\mathcal{J}}' = \begin{pmatrix} \tilde I + \tilde{\beta} (\tilde{B} \tilde I+\tilde I^* \tilde{B} ) & -\tilde I \tilde{\beta} -\tilde{\beta} \tilde I^*-\tilde{\beta} (\tilde{B} \tilde I+\tilde I^*\tilde{B} )\tilde{\beta} \\ \tilde{B} \tilde I+\tilde I^* \tilde{B} & -\tilde I^*-(\tilde{B} \tilde I+\tilde I^* \tilde{B} )\tilde{\beta} \end{pmatrix}. \end{aligned}

Let us assume that {\mathcal{J}}' = \tilde{{\mathcal{J}}}. We define

B' := B I + I^* B

and

\tilde{B}' := \tilde{B} \tilde I + \tilde I^* \tilde{B} .

By comparing top-left and top-right entries, we obtain

\begin{aligned} B' &= -(\tilde I \tilde{\beta} +\tilde{\beta} \tilde I^* + \tilde{\beta} \tilde{B}' \tilde{\beta} ) \\ &= -(\tilde I + \tilde{\beta} \tilde{B}' ) \tilde{\beta} - \tilde{\beta} \tilde I^* \\ &= (I^* + B' \beta ) \tilde{\beta} - \tilde{\beta} \tilde I^*, \end{aligned}

which implies

B' (1 - \beta \tilde{\beta} ) = I^* \tilde{\beta} - \tilde{\beta} \tilde I^*.

Similarly, we find

\begin{aligned} \tilde{B}' (1-\tilde{\beta} \beta ) = \tilde I^* \beta - \beta I^*. \end{aligned}

Now compose top-left entries from the right with (1-\tilde{\beta} \beta ) to get

(\tilde I + \tilde{\beta} \tilde{B}' ) (1-\tilde{\beta} \beta ) = -(I^* + B' \beta ) (1-\tilde{\beta} \beta )

and use the identities (Proof) and (Proof) to obtain

\tilde I(1-\tilde{\beta} \beta ) = -(1-\tilde{\beta} \beta ) I^*.

The remaining statements follow similarly.

Used by Proof, Proof
Proof

Proof

41SX

We compute {\mathcal{J}}:

\begin{aligned} {\mathcal{J}} &= e^{\beta } e^{B } \begin{pmatrix} I & 0 \\ 0 &-I^* \end{pmatrix} e^{-B } e^{-\beta } \\&= \begin{pmatrix} I + \beta (B I+I^*B ) & -I\beta -\beta I^*-\beta (B I+I^*B )\beta \\ B I+I^*B & -I^*-(B I+I^*B )\beta \end{pmatrix} \end{aligned}

Hence

\begin{aligned} \tilde{{\mathcal{J}}} = \begin{pmatrix} -I^*-(B I+I^*B )\beta & B I+I^*B \\ -I\beta -\beta I^*-\beta (B I+I^*B )\beta & I + \beta (B I+I^*B ) \end{pmatrix} \end{aligned}

needs to be equal to

\begin{aligned} {\mathcal{J}}' = \begin{pmatrix} \tilde I + \tilde{\beta} (\tilde{B} \tilde I+\tilde I^* \tilde{B} ) & -\tilde I \tilde{\beta} -\tilde{\beta} \tilde I^*-\tilde{\beta} (\tilde{B} \tilde I+\tilde I^*\tilde{B} )\tilde{\beta} \\ \tilde{B} \tilde I+\tilde I^* \tilde{B} & -\tilde I^*-(\tilde{B} \tilde I+\tilde I^* \tilde{B} )\tilde{\beta} \end{pmatrix}. \end{aligned}

Let us assume that {\mathcal{J}}' = \tilde{{\mathcal{J}}}. We define

B' := B I + I^* B

and

\tilde{B}' := \tilde{B} \tilde I + \tilde I^* \tilde{B} .

By comparing top-left and top-right entries, we obtain

\begin{aligned} B' &= -(\tilde I \tilde{\beta} +\tilde{\beta} \tilde I^* + \tilde{\beta} \tilde{B}' \tilde{\beta} ) \\ &= -(\tilde I + \tilde{\beta} \tilde{B}' ) \tilde{\beta} - \tilde{\beta} \tilde I^* \\ &= (I^* + B' \beta ) \tilde{\beta} - \tilde{\beta} \tilde I^*, \end{aligned}

which implies

B' (1 - \beta \tilde{\beta} ) = I^* \tilde{\beta} - \tilde{\beta} \tilde I^*.

Similarly, we find

\begin{aligned} \tilde{B}' (1-\tilde{\beta} \beta ) = \tilde I^* \beta - \beta I^*. \end{aligned}

Now compose top-left entries from the right with (1-\tilde{\beta} \beta ) to get

(\tilde I + \tilde{\beta} \tilde{B}' ) (1-\tilde{\beta} \beta ) = -(I^* + B' \beta ) (1-\tilde{\beta} \beta )

and use the identities (Proof) and (Proof) to obtain

\tilde I(1-\tilde{\beta} \beta ) = -(1-\tilde{\beta} \beta ) I^*.

The remaining statements follow similarly.

Remark

Remark 3.0

H9AH

In the case where (1-\tilde{\beta} \beta ) is invertible, there is a Poisson geometric interpretation of the equations (3.3) and (3.3).

Given two Poisson structures \pi,\pi' on M, they are said to be gauge equivalent if there exists a closed two-form B such that (\mathrm{id}+\pi B ) is invertible and satisfies the equation:

\pi' = \pi (1+\pi B )^{-1}.

Given a Poisson structure \pi it’s graph L_{\pi} = {\rm graph }(\pi) \subset TM \oplus T^*M defines a Dirac structure. The term gauge equivalence comes from the fact the associated Dirac structures are related via L_{\pi'} = e^{B } L_{\pi}.

If we view B as a bivector on V^* and \beta as a two-form on V^*, then the equation (3.3) tells us precisely that \tilde{\beta} is the \beta-gauge transformation of B. Similarly, (3.3) tells us that \beta is the B-gauge of \tilde{\beta}.

Definition

Definition 3.4

PGVQ

Let V / \Gamma and \tilde V / \tilde \Gamma be tori. We say that V/\Gamma and \tilde V / \tilde \Gamma are affine integral T-dual if we have a diagram

\begin{tikzcd} & \Gamma \times_{\Gamma_B} \tilde \Gamma \arrow[ld] \arrow[rd] & \\ \Gamma \arrow[rd, "p"'] & & \tilde \Gamma \arrow[ld, "\tilde p"] \\ & \Gamma_B & \end{tikzcd}

together with a 2-form F \in \scalebox{1.2}{}^2 (\Gamma \times_{\Gamma_B} \tilde \Gamma)^*, such that

  1. p and \tilde p are \mathbb Z-linear and surjective,

  2. F is unimodular, i.e., letting \Gamma_f := \text{ker}(p) and \tilde \Gamma_f := \text{ker}(\tilde p), we have that F|_{\Gamma_f \times \tilde \Gamma_f} : \Gamma_f \to (\tilde \Gamma_f)^* is invertible.

Used by Lemma 3.5
Lemma

Lemma 3.5

DL6S

Let V/\Gamma and \tilde V/ \tilde \Gamma be affine integral T-dual. Then the induced equivalence of linear Courant algebroids \phi : V \oplus V^* \to \tilde V \oplus \tilde V^* is integral, i.e., \phi(\Gamma \times \Gamma^*) = \tilde \Gamma \times \tilde \Gamma^*. In particular, if we choose an isomorphism \Gamma \cong \tilde \Gamma, then the corresponding automorphism of V \oplus V^* is contained in O(\Gamma, \Gamma^*) \cong O(n,n;\mathbb Z).

Proof.

By Theorem 3.2, we obtain an isomorphism of Courant algebroids over the base B/\Gamma_B. In the case of affine T-duality, the Courant isomorphism is constant along the base B / \Gamma_B (it also follows from the inspection below), hence we can describe it by an isomorphism

\phi : V \oplus V^* \to \tilde V \oplus \tilde V^*

preserving the inner products. To show that \phi preserves lattices, we inspect the definition of \phi given in [CG11, Section 3]. Let X + \xi \in \Gamma \oplus \Gamma^*. Choose an auxiliary lift \widehat X_{\mathrm{aux}} of X along \Gamma \times_{\Gamma_B} \tilde \Gamma \to \Gamma. Any other lift in V \times_{B} \tilde V will differ by an element in \tilde V_f. Hence, according to [CG11, Section 3], we need to find the unique element \Delta X \in \tilde V_f such that

(p^* \xi - F(\widehat X_{\mathrm{aux}})) |_{\Gamma_f} = F(\Delta X)|_{\Gamma_f}.

By construction and as F is integral, the left hand side is contained in \Gamma_f^*. Moreover, by unimodularity, F|_{\Gamma_f \times \tilde \Gamma_f} : \Gamma_f \to (\tilde \Gamma_f)^* is invertible, so \Delta X \in \tilde \Gamma_f \subset \tilde V_f. Thus the correct lift \widehat X = \widehat X_{\mathrm{aux}} + \Delta X lies in \Gamma \times_{\Gamma_B} \tilde \Gamma. By construction, p^* \xi - F(\widehat X) is basic, and

\phi(X + \xi) = \tilde p_* \widehat X + (p^* \xi - F(\widehat X)) \in \tilde \Gamma \times \tilde \Gamma^*. \qedhere
Proof

Proof

1WTE

By Theorem 3.2, we obtain an isomorphism of Courant algebroids over the base B/\Gamma_B. In the case of affine T-duality, the Courant isomorphism is constant along the base B / \Gamma_B (it also follows from the inspection below), hence we can describe it by an isomorphism

\phi : V \oplus V^* \to \tilde V \oplus \tilde V^*

preserving the inner products. To show that \phi preserves lattices, we inspect the definition of \phi given in [CG11, Section 3]. Let X + \xi \in \Gamma \oplus \Gamma^*. Choose an auxiliary lift \widehat X_{\mathrm{aux}} of X along \Gamma \times_{\Gamma_B} \tilde \Gamma \to \Gamma. Any other lift in V \times_{B} \tilde V will differ by an element in \tilde V_f. Hence, according to [CG11, Section 3], we need to find the unique element \Delta X \in \tilde V_f such that

(p^* \xi - F(\widehat X_{\mathrm{aux}})) |_{\Gamma_f} = F(\Delta X)|_{\Gamma_f}.

By construction and as F is integral, the left hand side is contained in \Gamma_f^*. Moreover, by unimodularity, F|_{\Gamma_f \times \tilde \Gamma_f} : \Gamma_f \to (\tilde \Gamma_f)^* is invertible, so \Delta X \in \tilde \Gamma_f \subset \tilde V_f. Thus the correct lift \widehat X = \widehat X_{\mathrm{aux}} + \Delta X lies in \Gamma \times_{\Gamma_B} \tilde \Gamma. By construction, p^* \xi - F(\widehat X) is basic, and

\phi(X + \xi) = \tilde p_* \widehat X + (p^* \xi - F(\widehat X)) \in \tilde \Gamma \times \tilde \Gamma^*. \qedhere
Remark

Remark 3.0

26C6

T-duality relative to a circle as in Remark 3.0 corresponds to a matrix \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} \oplus \textbf{\textit{I}}_{2n-2} when we unravel definitions.

Definition

Definition 4.1

EGEY

A curved dg-algebra or cdg-algebra A = (A^\bullet,\cdot,d,B) consists of

  1. a \mathbb Z-graded algebra A^\# = (A^{\bullet}, \cdot),

  2. a map, called the differential, d \colon A^{\bullet} \to A^{\bullet+1} satisfying

    d(ab) = d(a)b + (-1)^{|a|}a d(b),
  3. and an element B \in A^2, called the curvature element, satisfying

    \begin{aligned} &d(B) = 0,\\ &d^2(a) = [B,a] = Ba - aB. \end{aligned}

A (left) cdg-module P = (P^\bullet, ., d) over A is a graded left module P^\# = (P^{\bullet}, .) over the graded algebra A^\# together with a map d_P of degree 1 satisfying the graded Leibniz rule

d_P(a.p) = d(a).p + (-1)^{|a|} a.d_P (p)

and

d_P^2(p) = B.p.

Given two cdg-modules P = (P^\bullet,.,d_P) and Q = (Q^\bullet,.,d_Q) we define

\operatorname{Hom}_{A}^\bullet(P,Q)

in degree n to consist of maps f \colon P^\bullet \to Q^{\bullet+n} commuting with the A^\#-action and endowed with the differential

d(f) = d_Q \circ f - (-1)^{|f|} f \circ d_P

where |f| = n; one checks that indeed d^2 = 0 even when B \neq 0. With these definitions, we obtain the dg-category A\mathrm{-Mod} of left cdg-modules over the cdg-algebra A.

Definition

Definition 4.2

7PGH

Let A be a cdg-algebra. A cohesive module over A is a left cdg-module P over A such that its underlying graded module P^\# over A^\# is finitely generated and projective. The full dg subcategory of A\mathrm{-Mod} consisting of cohesive modules over A is denoted by \mathcal P_{A}.

Definition

Definition 4.3

EQJC

Let \mathcal{C} be a differential graded category, the associated homotopy category \operatorname{Ho} \mathcal{C} is the category with the same objects as {\mathcal{C}} and whose morphisms are the 0th cohomology

\operatorname{Ho} {\mathcal{C}}(x,y) = H^0({\mathcal{C}}(x,y)). \hfill \qedhere
Remark

Remark 4.0

YN4C

The original definition of cohesive modules is due to Block [BlockNC1] and is easily seen to be equivalent to Definition 4.1 in the case that A is concentrated in non-negative degrees. Let us thus assume that A is concentrated in non-negative degrees. Then a cohesive module over A is equivalently specified by a bounded, graded A^0-module E^\# consisting of finitely generated projective A^0-modules together with an A-superconnection, i.e. a map

\mathbb E \colon A^\bullet \otimes_{A^0} E^\bullet \to A^\bullet \otimes_{A^0} E^\bullet

which fulfills the Leibniz rule and satisfies the curvature equation

\mathbb E^2(m) = B.m.

One can decompose \mathbb E = \mathbb E^0 + \mathbb E^1 + ... with

\mathbb E^i \colon E^\bullet \to A^i \otimes_{A^0} E^{\bullet+1-i}.

The Leibniz rule is then equivalent to \mathbb E^i being A^0-linear if i \neq 1 and fulfilling the Leibniz rule for i = 1. The curvature equation decomposes into

\sum_{a+b = k} \mathbb E^a \mathbb E^b = 0 \quad \text{ for } k \neq 2

and

\mathbb E^0 \mathbb E^2 + \mathbb E^2 \mathbb E^0 + (\mathbb E^1)^2 = B.

In particular, (E^\bullet, \mathbb E^0) is a complex of finitely generated projective module A^0-modules. It turns out that the cohesive module (A^\bullet \otimes_{A^0} E^\bullet, \mathbb E) is isomorphic to the zero-object in \operatorname{Ho} (\mathcal P_A) if and only if the complex (E^\bullet, \mathbb E^0) is acyclic.

Remark

Remark 4.0

E4WP

The homotopy category of cohesive modules over a non-negatively graded curved algebra A is equivalent to the compact objects in the so-called twisted derived category of A and conversely the twisted derived category can be recovered as the derived category of \mathcal P_A, see [HH2] and references therein. If A is a dg-algebra the twisted derived category is a refinement of the derived category.

Theorem

Theorem 4.6

NTPG

Let X be a compact complex manifold, and consider its Dolbeault algebra \mathbf A(X). Then the homotopy category of cohesive modules \operatorname{Ho} {\mathcal{P}}_{\mathbf{A}(X)} is equivalent to {\operatorname{D^b_{coh}}}(X).

Example

Example 4.5

T7PN

The equivalence between {\rm Ho} {\mathcal{P}}_{\mathbf{A}} and {\operatorname{D^b_{coh}}}(X) is a generalization of the classical Koszul-Malgrange theorem. It states that a holomorphic vector E bundle is equivalent to a smooth vector bundle together with a partial flat connection on it. In our terms, the corresponding cohesive module is of the form \Gamma^\infty(E \otimes_{\mathcal O_X} (\Omega^{0,\bullet}, \bar \partial)) =(\Omega^{0,\bullet}(X,E), \bar \partial_{E}) where \bar \partial_{E} is the unique extension of \bar \partial_E \colon \Gamma^{\infty}(E) \to \Omega^{0,1}(M,E) compatible with the Leibniz rule.

Theorem

Theorem 5.1

D95H

Let X = (M,I) be a complex manifold. There is an L_\infty-quasi-isomorphism

(\Gamma(\scalebox{1.2}{$\wedge$}^{0,\bullet} T^*X \otimes \scalebox{1.2}{$\wedge$}^{\bullet+1,0} TX), \bar \partial, [\cdot, \cdot]_{GH}) \xrightarrow{\sim} HC_{\mathrm{cont}}(\mathbf A(X)).

In particular,

HH^\bullet_{\mathrm{cont}}(\mathbf A(X)) = H^\bullet(X, \scalebox{1.2}{$\wedge$}^\bullet T_{\mathrm{hol}} X) = HH^\bullet(X).
Example

Example 5.2

76YS

In the case of a complex torus V / \Gamma with constant complex structure I \colon V \to V the space of first order deformations is equal to

\begin{aligned} HH^2(V/\Gamma,I) = \scalebox{1.2}{$\wedge$}^{0,2} V_{\mathbb C}^* \oplus (V_{\mathbb C}^*)^{0,1} \otimes V_{\mathbb C}^{1,0} \oplus \scalebox{1.2}{$\wedge$}^{2,0} V_{\mathbb C}.\hfill\qedhere \end{aligned}
Definition

Definition 5.3

0Q3B

An \check{H}-twisted sheaf M over \mathcal O_X^\beta is a collection \{M_i\}_{i \in I} where each M_i is a sheaf over \mathcal O_{U_i}^\beta and a collection of isomorphisms \phi_{ij} \colon M_{i}|_{U_{ij}} \to M_{j}|_{U_{ij}} which are linear over \mathcal O_{U_{ij}}^\beta and satisfy the twisted cocycle condition

\phi_{jk} \phi_{ij} = \check{H}_{ijk} \phi_{ik}.

The notion of maps between \check{H}-twisted sheaves is defined in the same fashion. For consistency, it is important to note that \check{H} is central. Denote by D_{\mathrm{perf}}(\mathcal O_X^\beta, \check{H}) the subcategory of the derived category of all \check{H}-twisted sheaves which are locally quasi-isomorphic to perfect complexes, which is independent of the choices involved.

Proposition

Proposition 5.4

TYEB

There is an equivalence of \mathbb C_\epsilon-linear categories

\operatorname{Ho} (\mathcal P_{\mathbf A(X,\epsilon \beta, \epsilon B)}) \to D_{\mathrm{perf}}(\mathcal O_X^\beta, \check{H}).
Proof (Proof sketch).

The full proof is essentially an application of the techniques and results in [RHMC] and [Antweiler24, Section 7]. We thus focus on defining the functor on objects. Fix the following data:

f_{ij} \in C^\infty(U_{ij}), \quad \omega_{i} \in \Omega^{0,1}(U_i)

such that

\check B = \check d(f), \quad \bar \partial f = \check d(\omega),\quad \bar \partial(\omega_i) = B|_{U_i}

where \check d denotes the Čech differential.

Given a cohesive module P \in \mathcal P_{\mathcal A(X,\epsilon \beta, \epsilon B)} we can make it into a sheaf by

\underline{P} = \underline{\mathbf A(X,\epsilon \beta, \epsilon B)} \otimes_{\mathcal A(X,\epsilon \beta, \epsilon B)} P.

This tensor-product will automatically result in a sheaf because of the softness of C^\infty(X) and (graded) projectivity of P. Define a complex of twisted sheaves linear over \mathcal O_X^\beta as follows. For i \in I, set

P_i := (\underline{P}|_{U_i}, d - \epsilon 2\pi \mathbf{i} \omega_i)

and define isomorphisms

\phi_{ij} = 1 + \epsilon 2 \pi \mathbf{i} f_{ij} \colon P_i|_{U_{ij}} \to P_j|_{U_{ij}}.

Then by construction

\check d \phi = \check H

and P_i is cohesive over \underline{\mathbf A(U_i,\epsilon \beta, 0)} and hence also defines a module over \mathcal O_{U_i}^\beta. One can show that P_i is locally quasi-isomorphic to a perfect complex of \mathcal O_{U_i}^\beta-modules.

Proof

Proof (Proof sketch)

VU2F

The full proof is essentially an application of the techniques and results in [RHMC] and [Antweiler24, Section 7]. We thus focus on defining the functor on objects. Fix the following data:

f_{ij} \in C^\infty(U_{ij}), \quad \omega_{i} \in \Omega^{0,1}(U_i)

such that

\check B = \check d(f), \quad \bar \partial f = \check d(\omega),\quad \bar \partial(\omega_i) = B|_{U_i}

where \check d denotes the Čech differential.

Given a cohesive module P \in \mathcal P_{\mathcal A(X,\epsilon \beta, \epsilon B)} we can make it into a sheaf by

\underline{P} = \underline{\mathbf A(X,\epsilon \beta, \epsilon B)} \otimes_{\mathcal A(X,\epsilon \beta, \epsilon B)} P.

This tensor-product will automatically result in a sheaf because of the softness of C^\infty(X) and (graded) projectivity of P. Define a complex of twisted sheaves linear over \mathcal O_X^\beta as follows. For i \in I, set

P_i := (\underline{P}|_{U_i}, d - \epsilon 2\pi \mathbf{i} \omega_i)

and define isomorphisms

\phi_{ij} = 1 + \epsilon 2 \pi \mathbf{i} f_{ij} \colon P_i|_{U_{ij}} \to P_j|_{U_{ij}}.

Then by construction

\check d \phi = \check H

and P_i is cohesive over \underline{\mathbf A(U_i,\epsilon \beta, 0)} and hence also defines a module over \mathcal O_{U_i}^\beta. One can show that P_i is locally quasi-isomorphic to a perfect complex of \mathcal O_{U_i}^\beta-modules.

Definition

Definition 5.5

BLHS

Let B^{0,2} \in \Omega^{0,2}(X) be a \bar \partial-closed (0,2)-form. We define the generalized Dolbeault algebra of the GC structure {\mathcal{J}}_{I}^{B} as

\mathbf A(I,0,B) := (\Omega^{0,\bullet}(X),\wedge,\bar \partial, 2\pi \mathbf{i} {B^{0,2}}). \hfill\qedhere
Theorem

Theorem 5.6

7UUO

The homotopy category {\rm Ho} {\mathcal{P}}_{\mathbf{A}(I,0,B)} is equivalent to the bounded derived category {\operatorname{D^b_{coh}}}(X,\tilde{B}) of complexes of \exp(2\pi \mathbf{i} B^{0,2})-twisted sheaves of {\mathcal{O}}_X-modules with coherent cohomology.

Definition

Definition 5.7

48FQ

We define the \beta-deformed Dolbeault algebra as

\mathbf{A}(I, \beta, 0) = (C^\infty(V / \Gamma) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} V_{\mathbb C}^*, \star^{\beta}, \bar \partial_{I}, 0),

where

(u_{\gamma_1} \otimes \omega_1) \star^{\beta} (u_{\gamma_2} \otimes \omega_2) := e^{\pi \mathbf{i} \beta(\gamma_1,\gamma_2)} u_{\gamma_1 + \gamma_2} \otimes \omega_1 \wedge \omega_2

for u_{\gamma}(x) := e^{2 \pi \mathbf{i} \gamma(x)}, \gamma \in \Gamma^* and \omega_i \in \scalebox{1.2}{}^{(0,\bullet)} V_{\mathbb C}^*.

Lemma

Lemma 5.8

2PPS

Up to isomorphism of algebras, C^\infty(V/\Gamma,\beta) and \mathbf{A}(I, \beta, 0) only depend on [\beta] \in \scalebox{1.2}{}^2 V / \scalebox{1.2}{}^2 \Gamma.

Proof.

Let \beta \in \scalebox{1.2}{}^2 V and \sigma \in \scalebox{1.2}{}^2 \Gamma. To construct an isomorphism

\Phi \colon C^\infty(V/\Gamma,\beta) \to C^\infty(V/\Gamma, {\beta + \sigma})

we choose a \mathbb Z-basis of \Gamma^*, denoted \alpha_1,...,\alpha_n. Then there is a unique continuous algebra map \Phi satisfying

\Phi(u_{\alpha_i}) = u_{\alpha_i} \quad \forall i = 1,...,n.

More specifically it is given by

\Phi(u_{\gamma}) = \exp\left(\pi \mathbf{i} \sum_{i < j} \sigma(\gamma_i \alpha_i, \gamma_j \alpha_j)\right) u_{\gamma}

where \gamma = \sum_{i} \gamma_i \alpha_i.

Used by Proof
Proof

Proof

IFOG

Let \beta \in \scalebox{1.2}{}^2 V and \sigma \in \scalebox{1.2}{}^2 \Gamma. To construct an isomorphism

\Phi \colon C^\infty(V/\Gamma,\beta) \to C^\infty(V/\Gamma, {\beta + \sigma})

we choose a \mathbb Z-basis of \Gamma^*, denoted \alpha_1,...,\alpha_n. Then there is a unique continuous algebra map \Phi satisfying

\Phi(u_{\alpha_i}) = u_{\alpha_i} \quad \forall i = 1,...,n.

More specifically it is given by

\Phi(u_{\gamma}) = \exp\left(\pi \mathbf{i} \sum_{i < j} \sigma(\gamma_i \alpha_i, \gamma_j \alpha_j)\right) u_{\gamma}

where \gamma = \sum_{i} \gamma_i \alpha_i.

Definition

Definition 5.9

FDW9

Let V be a real vector space together with a linear complex structure I \colon V \to V and let \Gamma \subset V be a lattice. Let \beta \in \scalebox{1.2}{}^2 V be a real bivector and let B \in \scalebox{1.2}{}^2 V^* be a real 2-form. Then the (I, \beta, B)-generalized Dolbeault algebra is defined to be

\mathbf{A}(I, \beta, B) = (C^\infty(V/\Gamma) \otimes \scalebox{1.2}{$\wedge$}^\bullet (V_{\mathbb C}^*)^{0,1}, \star^\beta, \bar \partial_{I}, 2 \pi \mathbf{i} B^{0,2}). \hfill\qedhere
Theorem

Theorem 5.10

BM57

Let A be a Fréchet-Fredholm cdg-algebra. Then the morphism complexes in \mathcal P_{A} have finite-dimensional cohomology. More generally, let A, B, C be Fréchet-Fredholm cdg-algebras. Then the tensor product over B induces an A_\infty-functor

\mathcal P_{A \widehat \otimes B^{op}} \times \mathcal P_{B \widehat \otimes C^{op}} \to \mathcal P_{A \widehat \otimes C^{op}}.
Proposition

Proposition 5.12

A52R

The cdg-algebra \mathbf{A}(I, \beta, B) is Fréchet twist-smooth. In particular, there exists a canonical cohesive module

\Delta \in \mathcal P_{\mathbf{A}(I, \beta, B) \widehat \otimes \mathbf{A}(I, \beta, B)^{op}}

representing the identity functor

\mathcal P_{\mathbf{A}(I, \beta, B)} \to \mathcal P_{\mathbf{A}(I, \beta, B)}.
Definition

Definition 5.13

Y8S8

Let A and B be Fréchet-Fredholm cdg-algebras which are also Fréchet twist-smooth. We call A and B cohesive Morita equivalent if there exist cohesive bimodules R \in \mathcal P_{A \widehat \otimes B^{op}},S \in \mathcal P_{B \widehat \otimes A^{op}} such that

R \widehat \otimes_B S \simeq \Delta_A

and

S \widehat \otimes_A R \simeq \Delta_B,

where \simeq denotes continuous homotopy equivalence.

Remark

Remark 5.0

4OFW

For the reader uncomfortable with A_\infty-functors we mention that it is possible to induce an equivalence of categories by a dg functor at the price of enlarging the category of cohesive modules. Any bimodule induces a natural functor between the larger contraderived categories of A and B as in the corollary. Under the assumptions of the corollary it restricts to the subcategories of modules weakly equivalent to cohesive modules. We have to replace it by an A_\infty-functor to consider a functor between the categories of honest cohesive modules.

Example

Example 5.15

W15X

Let \mathbf A(X), \mathbf A(Y) be the Dolbeault algebras of compact complex manifolds X and Y. Then we may translate from cohesive modules to elements of the bounded derived category. \mathbf A(X) and \mathbf A(Y) are cohesive Morita equivalent if and only if they are related by invertible Fourier-Mukai transforms, i.e. if there exist

R \in {\operatorname{D^b_{coh}}}(X \times Y),\quad S \in {\operatorname{D^b_{coh}}}(Y \times X)

such that

\Delta_* (\mathcal O_X) \simeq R(p_{X \times X})_*( p_{X \times Y}^* R \otimes^L p_{Y \times X}^* S) \in {\operatorname{D^b_{coh}}}(X \times X),

where

\begin{aligned} p_{X \times X}\colon X \times Y \times X &\to X \times X\\ p_{X \times Y}\colon X \times Y \times X &\to X \times Y\\ p_{Y \times X}\colon X \times Y \times X &\to Y \times X. \end{aligned}

are the projections and

\Delta_* (\mathcal O_Y) \simeq R(q_{Y \times Y})_*( q_{Y \times X}^* S \otimes^L q_{X \times Y}^* R) \in {\operatorname{D^b_{coh}}}(Y \times Y).

Here the projections q_{(-)} are defined dually.

Definition

Definition 6.1

B3N8

Define an action of the non-commutative torus C^{\infty}(V/\Gamma,\sigma) on \mathcal S(L) by setting

(u_{\gamma} . g)(x) = e^{\gamma}g(x) = e^{\pi \mathbf{i} \alpha(2x-l)} g(x-l) ,

for L^* \oplus L \ni (\alpha,l ) = \gamma \in \Gamma^* and u_{\gamma}(x) := e^{2\pi \mathbf{i} \gamma(x)}. See [NCGrauert, Remark 4.2.1] for an interpretation of the notation e^{\gamma}g(x) = e^{2\pi \mathbf{i} \alpha(2x-v)} g(x-v). It is not difficult to check that the formula is well-defined (and continuous). The resulting module is finitely generated projective over C^{\infty}(V/\Gamma,\sigma) and we denote it by H(\sigma).

Definition

Definition 6.2

J5NX

The (\sigma,B) deformed de-Rham algebra is the curved dg-algebra

\Omega(V/\Gamma,\sigma,B) := (C^{\infty}(V/\Gamma) \otimes \scalebox{1.2}{$\wedge$}^\bullet V_{\mathbb C}^*, \star^{\sigma}, d_{DR}, 2\pi \mathbf{i} B). \qedhere
Lemma

Lemma 6.3

4C7C

The Heisenberg module H(\sigma) carries a canonical projectively flat connection \nabla with curvature

\nabla^2 = -2\pi \mathbf{i} \omega

and hence (\Omega^\bullet(V/\Gamma,\sigma) \widehat \otimes_{C^\infty(V/\Gamma,\sigma)} H(\sigma), \nabla) is a cohesive module over the cdg-algebra \Omega(V/\Gamma,\sigma,-\omega).

Proof.

We use the notation \nabla_v(g) := \iota_v ( \nabla(g) ) for v \in V. Then \nabla is uniquely determined by defining

\nabla_v(g) := \omega (v).g.

We verify the Leibniz rule using (?):

\begin{aligned} \nabla_v(u_{\gamma}.g) &= \omega(v).(e^{\gamma}g) \\ &= e^{\gamma} \omega(v).g + [r(\omega(v)),r(\gamma)]. e^{\gamma}g \\ &= u_\gamma.(\nabla_v(g)) + 2\pi \mathbf{i} \sigma( \omega(v), \gamma) u_\gamma.g\\ &= u_\gamma.(\nabla_v(g)) + 2\pi \mathbf{i} \gamma(v) u_\gamma.g \\ &= u_\gamma.(\nabla_v(g)) + d_v(u_\gamma).g. \end{aligned}

To get from the first to the second line one expands \omega(v).(e^{\gamma}g) and uses the commutation relations between r(\omega(v)) and r(\gamma).

Now we compute the curvature:

\begin{aligned} [\nabla_v, \nabla_w](g) &= \omega (v).\omega (w).g - \omega (w).\omega (v).g \\ &= [r(\omega (v)),r(\omega (w))].g\\ &= 2\pi \mathbf{i} \sigma( \omega (v), \omega (w) ).g\\ &= -2\pi \mathbf{i} \omega( v, w ).g. \qedhere \end{aligned}
Proof

Proof

QIEP

We use the notation \nabla_v(g) := \iota_v ( \nabla(g) ) for v \in V. Then \nabla is uniquely determined by defining

\nabla_v(g) := \omega (v).g.

We verify the Leibniz rule using (?):

\begin{aligned} \nabla_v(u_{\gamma}.g) &= \omega(v).(e^{\gamma}g) \\ &= e^{\gamma} \omega(v).g + [r(\omega(v)),r(\gamma)]. e^{\gamma}g \\ &= u_\gamma.(\nabla_v(g)) + 2\pi \mathbf{i} \sigma( \omega(v), \gamma) u_\gamma.g\\ &= u_\gamma.(\nabla_v(g)) + 2\pi \mathbf{i} \gamma(v) u_\gamma.g \\ &= u_\gamma.(\nabla_v(g)) + d_v(u_\gamma).g. \end{aligned}

To get from the first to the second line one expands \omega(v).(e^{\gamma}g) and uses the commutation relations between r(\omega(v)) and r(\gamma).

Now we compute the curvature:

\begin{aligned} [\nabla_v, \nabla_w](g) &= \omega (v).\omega (w).g - \omega (w).\omega (v).g \\ &= [r(\omega (v)),r(\omega (w))].g\\ &= 2\pi \mathbf{i} \sigma( \omega (v), \omega (w) ).g\\ &= -2\pi \mathbf{i} \omega( v, w ).g. \qedhere \end{aligned}
Example

Example 6.4

5CXA

Let V/\Gamma be a torus and V^* / \Gamma^* its dual torus. Smooth complex line bundles on V/\Gamma \times V^* / \Gamma^* are parametrized by

H^2(V/\Gamma \times V^* / \Gamma^*, \mathbb Z) = \scalebox{1.2}{$\wedge$}^2 (\Gamma^* \oplus \Gamma) \cong (\Gamma^* \otimes \Gamma) \oplus \scalebox{1.2}{$\wedge$}^2 \Gamma \oplus \scalebox{1.2}{$\wedge$}^2 \Gamma^*.

The canonical pairing between \Gamma^* and \Gamma yields an element

-\omega \in \Gamma^* \otimes \Gamma

so that -\iota_\gamma(\omega) = \gamma for \gamma \in \Gamma. The corresponding line bundle with curvature -2 \pi \mathbf{i} \omega is known as the Poincaré line bundle. As a map \Gamma \oplus \Gamma^* \to (\Gamma \oplus \Gamma^*)^* = \Gamma^* \oplus \Gamma, \omega has the form

-\omega = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}

and so the corresponding bivector \sigma := \omega^{-1} is of the form

\sigma = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix},

which lies on the lattice \scalebox{1.2}{}^2 (\Gamma \oplus \Gamma^*), hence C^{\infty}(V/\Gamma \times V^* / \Gamma^*,\sigma) \cong C^{\infty}(V/\Gamma \times V^* / \Gamma^*) and the Heisenberg module H(\sigma) is isomorphic to (the global sections of) the Poincaré line bundle.

Example

Example 6.5

HDE7

Let us explain the role of the de-Rham algebra in the untwisted, commutative case. For any smooth manifold, the Higher Riemann-Hilbert correspondence [RHMC] establishes an equivalence

D_{\mathrm{fclc}}(\underline{\mathbb C}) \cong H^0( \mathcal P_{\Omega(M)} )

where D_{\mathrm{fclc}}(\underline{\mathbb C}) denotes the subcategory of the derived category of complexes of sheaves of vector spaces whose cohomology sheaves are locally finitely generated and locally constant. If X = (M,I) is a complex manifold then the unit \underline{\mathbb C} \to \mathcal O_X induces a functor D_{\mathrm{fclc}}(\underline{\mathbb C}) \to {\operatorname{D^b_{coh}}}(\mathcal O_X) which agrees with the functor on the level of cohesive modules induced by the map of dg-algebras \Omega(M) \to \mathbf A(X). In the curved and non-commutative case, \mathcal P_{\Omega(V/\Gamma,\sigma,B)} should thus be thought of as a higher-categorical version of twisted non-commutative locally constant sheaves.

Proposition

Proposition 6.6

GVNC

Let \beta \in \scalebox{1.2}{}^2 V, \tilde{\beta} \in \scalebox{1.2}{}^2 V^* such that 1 - \beta \tilde{\beta} \colon V \to V is invertible. Let I\colon V\to V be a complex structure. Then there exists a cohesive Heisenberg bimodule P, which we call the non-commutative Poincaré line bundle, over

\mathbf{A}(I, \beta, B) \widehat \otimes \mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B})^{op}

if

e^{\beta } e^{B } L_{{\mathcal{J}}_I}

is T-dual to

e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I^*}}

via \varphi (see ?), i.e., if

\varphi\left( e^{\beta } e^{B } L_{{\mathcal{J}}_I} \right) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I^*}}

holds.

Proof.

From the computation of the curvature, we see that P defines a cohesive module over
\mathbf{A}(I, \beta, B) \widehat \otimes \mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B})^{op} if the equations

\begin{aligned} &\tilde{I} = -(1-\tilde{\beta} \beta ) I^* (1-\tilde{\beta} \beta )^{-1} \\ &B^{(0,2)_I} = (\tilde{\beta} (1-\beta \tilde{\beta} )^{-1})^{(0,2)_I} \\ &\tilde{B}^{(0,2)_{\tilde{I}}} = (\beta (1-\tilde{\beta} \beta )^{-1})^{(0,2)_{\tilde{I}}} \end{aligned}

hold. These are equivalent to

\varphi\left( e^{\beta } e^{B } L_{{\mathcal{J}}_I} \right) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I^*}}

by Lemma 3.3.

Uses Definition 3.1 · Used by Proof
Proof

Proof

XCTX

From the computation of the curvature, we see that P defines a cohesive module over
\mathbf{A}(I, \beta, B) \widehat \otimes \mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B})^{op} if the equations

\begin{aligned} &\tilde{I} = -(1-\tilde{\beta} \beta ) I^* (1-\tilde{\beta} \beta )^{-1} \\ &B^{(0,2)_I} = (\tilde{\beta} (1-\beta \tilde{\beta} )^{-1})^{(0,2)_I} \\ &\tilde{B}^{(0,2)_{\tilde{I}}} = (\beta (1-\tilde{\beta} \beta )^{-1})^{(0,2)_{\tilde{I}}} \end{aligned}

hold. These are equivalent to

\varphi\left( e^{\beta } e^{B } L_{{\mathcal{J}}_I} \right) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I^*}}

by Lemma 3.3.

Theorem

Theorem 6.7

N9CO

Let (V/\Gamma, I, \beta, B) and (V^*/\Gamma^*, \tilde I, \tilde{\beta}, \tilde{B}) be such that

e^{\beta } e^{B } L_{{\mathcal{J}}_I}

and

e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I}}

are T-dual via \varphi. Moreover, suppose that

1-\tilde{\beta} \beta \colon V^* \to V^*

is invertible. Then \mathbf{A}(I, \beta, B) is cohesive Morita equivalent to \mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B}). In particular, the non-commutative Poincaré line bundle induces an A_\infty-quasi-equivalence

\mathcal P_{\mathbf{A}(I, \beta, B)} \xrightarrow{\sim} \mathcal P_{\mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B})} .
Proof.

In Proposition 6.6, we showed that the existence of the non-commutative Poincaré line bundles is equivalent to T-duality

\varphi\left( e^{\beta } e^{B } L_{{\mathcal{J}}_I} \right) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I^*}}.

In [NCGrauert, Theorem 4.3.11] it is proved that the non-commutative Poincaré line bundles induce equivalences.

Proof

Proof

B6ZB

In Proposition 6.6, we showed that the existence of the non-commutative Poincaré line bundles is equivalent to T-duality

\varphi\left( e^{\beta } e^{B } L_{{\mathcal{J}}_I} \right) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I^*}}.

In [NCGrauert, Theorem 4.3.11] it is proved that the non-commutative Poincaré line bundles induce equivalences.

Remark

Remark 6.0

54E3

As a consequence of Theorem 7.7, we will see that the statement of the theorem also holds if 1-\beta \tilde{\beta} is singular but the equivalence bimodule may be more complicated.

Corollary

Corollary 6.8

O7N9

Let (V/\Gamma,I) be a complex torus and let \sigma \in \scalebox{1.2}{}^{2,0} V. Choose a real bivector \beta \in \scalebox{1.2}{}^2 V such that (\beta)^{2,0} = \sigma. Then \mathcal P_{\mathbf A(I,\beta,0)} does not depend on the choice of \beta up to A_\infty-quasi-equivalence.

Definition

Definition 7.1

CSCF

Let {\mathcal{J}} be a translation-invariant GC structure on V/\Gamma of even cotype. Choose I, \beta and B such that

L_{{\mathcal{J}}} = e^{\beta } e^{B } L_{{\mathcal{J}}_I}.

The derived category of {\mathcal{J}} is defined as

{\mathcal{D}}({\mathcal{J}}) = {\mathcal{P}}_{\mathbf{A}(I,\beta,B)}.

If {\mathcal{J}} is of odd cotype we let {\mathcal{J}}' be the GC structure obtained by performing integral T-duality relative to any circle and define

{\mathcal{D}}({\mathcal{J}}) := {\mathcal{D}}({\mathcal{J}}'). \qedhere
Remark

Remark 7.0

0DKV

By construction {\mathcal{D}}({\mathcal{J}}) is a pre-triangulated dg-category, or equivalently a stable \infty-category linear over \mathbb{C}. We may disregard some structure and consider its homotopy category as a triangulated category.

Example

Example 7.2

F59V

Let {\mathcal{J}} be the GC structure associated to the complex torus (X,J). Then

\operatorname{Ho} {\mathcal{D}}({\mathcal{J}}) = \operatorname{Ho} \mathcal P_{\mathbf A(J)} \cong {\operatorname{D^b_{coh}}}(X)

by Theorem 5.6.

Uses Theorem 5.6 · Used by Proof
Example

Example 7.3

OWMJ

Let {\mathcal{J}} be the GC structure associated to the complex torus (X,J) equipped with a gerbe B. Then

\operatorname{Ho} {\mathcal{D}}({\mathcal{J}}) = \operatorname{Ho} \mathcal P_{\mathbf A(J,0,B)} \cong {\operatorname{D^b_{coh}}}(X, B)

by Theorem 5.6.

Example

Example 7.4

99RA

Let {\mathcal{J}}_\omega be the GC structure associated to a symplectic manifold (M,\omega). As in Example 2.11 we choose an auxiliary Hyperkähler structure and find

{\mathcal{J}}_{\omega_J} = e^{-\omega_K^{-1}} e^{\frac{\omega _K}{2}} {\mathcal{J}}_I e^{\frac{-\omega _K}{2}} e^{\omega_K^{-1}} .\hfill

Thus the generalized Dolbeault algebra of {\mathcal{J}}_\omega is

\mathbf{A}(\omega) := \mathbf{A}\left(J,-\omega_K^{-1}, \frac{1}{2}\omega_K\right)

We will show in section ? that as a consequence of homological mirror symmetry {\mathcal{D}}({\mathcal{J}}_\omega) is equivalent to a suitable Fukaya category of (M, \omega).

Remark

Remark 7.0

4L6V

By Proposition 5.11, {\mathcal{D}}({\mathcal{J}}) is a locally proper category but it does in general not have a compact generator. This is already evident for the derived category of a generic complex torus.

Theorem

Theorem 7.7

Z9M6

If

L_{{\mathcal{J}}} = e^{\beta_1 } e^{B_1 } L_{{\mathcal{J}}_{I_1}} = e^{\beta_2 } e^{B_2 } L_{{\mathcal{J}}_{I_2}}

then the algebras \mathbf{A}(I_1,\beta_1,B_1) and \mathbf{A}(I_2,\beta_2,B_2) are cohesive Morita equivalent. In particular, {\mathcal{D}}({\mathcal{J}}) is well-defined up to A_\infty-quasi-equivalence if {\mathcal{J}} is of even cotype.

Proof.

Suppose that

L_{{\mathcal{J}}} = e^{\beta_1 } e^{B_1 } L_{{\mathcal{J}}_{I_1}} = e^{\beta_2 } e^{B_2 } L_{{\mathcal{J}}_{I_2}}.

We want to use Theorem 6.7 to compare with the T-dual picture, but as we need invertiblity assumptions we may have to perturb the presentation. So choose \Delta as in Lemma 7.6 below. Since \Delta is of type (1,1) with respect to I_1, we have

L_{{\mathcal{J}}} = e^{\beta_1 } e^{B_1 + \Delta} L_{{\mathcal{J}}_{I_1}}

and

\mathbf{A}(I_1,\beta_1,B_1) = \mathbf{A}(I_1,\beta_1,B_1 + \Delta).

We write down a presentation of the T-dual of L_{{\mathcal{J}}} on V^* / \Gamma^*

\varphi(L_{{\mathcal{J}}}) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I}},

and by Lemma 3.3 we have

\begin{aligned} &\tilde{\beta} = (1+(B_1 + \Delta ) \beta_1 )^{-1} (B_1 + \Delta) \\ &\tilde{B} = \beta_1 (1+(B_1 + \Delta ) \beta_1 )^{-1}\\ &\tilde I = -(1+(B_1 + \Delta ) \beta_1 )^{-1} I^* (1+(B_1 + \Delta ) \beta_1 ). \end{aligned}

These equations are a particular instance of the equations in Lemma 3.3 applied to (I,\beta,B) := (I_1,\beta_1,B_1+\Delta) noting (1+(B_1+\Delta)\beta_1)^{-1} = 1 - \tilde \beta \beta_1 with invertibility provided by Lemma 7.6. Moreover, as \Delta is of type (1,1), the same equations hold for (I_1,\beta_1,B_1), and thus they hold for the second presentation (I_2,\beta_2,B_2). Since \Delta was chosen exactly so that 1 - \beta_2 \tilde{\beta} and 1-\beta_1 \tilde{\beta} are invertible, Theorem 6.7 yields the A_\infty-quasi-equivalences

\mathcal P_{\mathbf{A}(I_1,\beta_1,B_1+\Delta)} \simeq \mathcal P_{\mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B})} \simeq \mathcal P_{\mathbf{A}(I_2,\beta_2, B_2)}. \qedhere
Proof

Proof

TQ6J

Suppose that

L_{{\mathcal{J}}} = e^{\beta_1 } e^{B_1 } L_{{\mathcal{J}}_{I_1}} = e^{\beta_2 } e^{B_2 } L_{{\mathcal{J}}_{I_2}}.

We want to use Theorem 6.7 to compare with the T-dual picture, but as we need invertiblity assumptions we may have to perturb the presentation. So choose \Delta as in Lemma 7.6 below. Since \Delta is of type (1,1) with respect to I_1, we have

L_{{\mathcal{J}}} = e^{\beta_1 } e^{B_1 + \Delta} L_{{\mathcal{J}}_{I_1}}

and

\mathbf{A}(I_1,\beta_1,B_1) = \mathbf{A}(I_1,\beta_1,B_1 + \Delta).

We write down a presentation of the T-dual of L_{{\mathcal{J}}} on V^* / \Gamma^*

\varphi(L_{{\mathcal{J}}}) = e^{\tilde{\beta} } e^{\tilde{B} } L_{{\mathcal{J}}_{\tilde I}},

and by Lemma 3.3 we have

\begin{aligned} &\tilde{\beta} = (1+(B_1 + \Delta ) \beta_1 )^{-1} (B_1 + \Delta) \\ &\tilde{B} = \beta_1 (1+(B_1 + \Delta ) \beta_1 )^{-1}\\ &\tilde I = -(1+(B_1 + \Delta ) \beta_1 )^{-1} I^* (1+(B_1 + \Delta ) \beta_1 ). \end{aligned}

These equations are a particular instance of the equations in Lemma 3.3 applied to (I,\beta,B) := (I_1,\beta_1,B_1+\Delta) noting (1+(B_1+\Delta)\beta_1)^{-1} = 1 - \tilde \beta \beta_1 with invertibility provided by Lemma 7.6. Moreover, as \Delta is of type (1,1), the same equations hold for (I_1,\beta_1,B_1), and thus they hold for the second presentation (I_2,\beta_2,B_2). Since \Delta was chosen exactly so that 1 - \beta_2 \tilde{\beta} and 1-\beta_1 \tilde{\beta} are invertible, Theorem 6.7 yields the A_\infty-quasi-equivalences

\mathcal P_{\mathbf{A}(I_1,\beta_1,B_1+\Delta)} \simeq \mathcal P_{\mathbf{A}(\tilde I, \tilde{\beta}, \tilde{B})} \simeq \mathcal P_{\mathbf{A}(I_2,\beta_2, B_2)}. \qedhere
Lemma

Lemma 7.6

UKB5

Let B_1,B_2 be two forms on V, \beta_1, \beta_2 bivectors on V and I_1,I_2 complex structures such that

L_{{\mathcal{J}}} = e^{\beta_1 } e^{B_1 } L_{{\mathcal{J}}_{I_1}} = e^{\beta_2 } e^{B_2 } L_{{\mathcal{J}}_{I_2}}.

Then there is a real two form \Delta \in \scalebox{1.2}{}^{(1,1)_{I_1}} V^*_{\mathbb C} \cap \scalebox{1.2}{}^2 V^* of type (1,1) for I_1, such that the following endomorphisms are invertible

\begin{aligned} &\alpha := 1 + \beta_1 (B_1 +\Delta ) \colon V \to V\\ &1 - \beta_2 (B_1 +\Delta ) \alpha^{-1} \colon V \to V. \end{aligned}
Uses Definition 2.4 · Used by Proof
Proof of Lemma 7.6

The invertibility of these endomorphisms is equivalent to the invertibility of \alpha and

\begin{aligned} &Q:= 1 + (\beta_1 - \beta_2 )(B_1 +\Delta ) \colon V \to V. \end{aligned}

Let us focus on the invertibility of \alpha first. If \beta_1=0 there is nothing to prove, so suppose \beta_1\neq 0 and define the real subspace E := \mathrm{Im}(\beta_1 ). Since E is even-dimensional there exists \omega \in \scalebox{1.2}{}^{(1,1)_{I_1}} V^*_{\mathbb C} \cap \scalebox{1.2}{}^2 V^* such that \omega|_{E \times E} is non-degenerate by Lemma 7.7. Consider the polynomial

p(t) := \det(1 + \beta_1 (B_1 + t \omega )) = \det(1_E + \beta_1 (B_1 + t \omega )|_{E}).

The leading coefficient of p(t) in t is

\det(\beta_1 ( \omega )|_{E}) \neq 0

by construction of \omega. Hence, p(t) does not vanish identically. In fact \{ t \in \mathbb R, p(t) \neq 0\} \subset \mathbb R is cofinite. Choose t_0 \in \mathbb R with p(t_0) \neq 0 so that

1+\beta_1 (B_1 + t_0 \omega )

is invertible. Analogously, we can then solve for t_1 and \omega_1 such that

1 + (\beta_1 - \beta_2 )(B_1 + t_0 \omega + t_1 \omega_1 )

is invertible. As t_1 can be chosen arbitrarily close to 0, we can arrange that

1 + \beta_1 (B_1 +t_0 \omega + t_1 \omega_1 )

is also invertible. Hence, \Delta :=t_0 \omega + t_1 \omega_1 is the desired two form.

Lemma

Lemma 7.7

X0NB

Let V be a real vector space with a complex structure J \colon V \to V and let E \subset V be a real even-dimensional subspace of V. Then there exists a real 2-form of type (1,1), \omega \in \scalebox{1.2}{}^{(1,1)} V^*_{\mathbb C} \cap \scalebox{1.2}{}^2 V^* such that \omega|_{E \times E} is non-degenerate.

Proof.

Let

E_J := E \cap JE

be the maximal J-invariant subspace of E. Choose a complement C \subset E, so that E_J \oplus C = E. Then JC \cap E = 0 and C is even-dimensional. Thus,

V = E_J \oplus C \oplus JC \oplus T

where T is a complex subspace. Choose a non-degenerate real 2-form \omega_{E_J} on E_J of type (1,1) and a real non-degenerate 2-form \omega_C on C. Then for (e,c,Jc',t), (\tilde e, \tilde c,J \tilde c',\tilde t) \in V we define

\omega((e,c,Jc',t), (\tilde e, \tilde c,J \tilde c',\tilde t)) := \omega_E(e,\tilde e) + \omega_C(c,\tilde c) + \omega_C(c',\tilde c').

\omega is real of type (1,1) and its restriction to E = E_J \oplus C is non-degenerate.

Uses Definition 2.4 · Used by Proof
Proof

Proof

XR8Z

Let

E_J := E \cap JE

be the maximal J-invariant subspace of E. Choose a complement C \subset E, so that E_J \oplus C = E. Then JC \cap E = 0 and C is even-dimensional. Thus,

V = E_J \oplus C \oplus JC \oplus T

where T is a complex subspace. Choose a non-degenerate real 2-form \omega_{E_J} on E_J of type (1,1) and a real non-degenerate 2-form \omega_C on C. Then for (e,c,Jc',t), (\tilde e, \tilde c,J \tilde c',\tilde t) \in V we define

\omega((e,c,Jc',t), (\tilde e, \tilde c,J \tilde c',\tilde t)) := \omega_E(e,\tilde e) + \omega_C(c,\tilde c) + \omega_C(c',\tilde c').

\omega is real of type (1,1) and its restriction to E = E_J \oplus C is non-degenerate.

Theorem

Theorem 7.10

K5US

Let {\mathcal{J}} be a translation invariant GC structure on T^{2n} of even cotype. Then for any g\in SO(2n, 2n; \mathbb{Z}) there is an A_\infty-quasi-equivalence

{\mathcal{D}}({\mathcal{J}}) \simeq {\mathcal{D}}(g{\mathcal{J}} g^{-1}).
Proof of Theorem 7.10

We first note that SO(2n, 2n; \mathbb{Z}) is generated by the matrices of the following form:

\begin{pmatrix} P & 0 \\ 0 & P^{-T} \end{pmatrix}, \begin{pmatrix} \textbf{\textit{I}} & \beta \\ 0 & \textbf{\textit{I}} \end{pmatrix}, \begin{pmatrix} 0 & \textbf{\textit{I}} \\ \textbf{\textit{I}} & 0 \end{pmatrix},\quad

where P \in GL(2n, \mathbb{Z}) and \beta is an antisymmetric integral 2n\times 2n-matrix. This follows for example from [hahn2013classical, Theorem 9.2.8]: These matrices generate all elementary matrices, thus generating the spinoral kernel, and

P = \operatorname{diag} (-1, 1, \dots, 1)

provides an element with non-trivial spinor norm.

Thus the theorem follows in three steps. The first set of generators of SO(2n,2n;\mathbb{Z}) correspond to base change by P, which induces a natural isomorphism of generalized Dolbeault algebras. The second set of generators correspond to integral \beta-transforms, which do not change the generalized Dolbeault algebra by Lemma 5.8. The last generator corresponds to T-duality relative to a point, which leaves the generalized derived category invariant by Theorem 6.7.

Corollary

Corollary 7.9

JS3F

{\mathcal{D}}({\mathcal{J}}) is well-defined up to A_\infty-quasi-equivalence and invariant under O(2n,2n;\mathbb{Z}).

Proof of Corollary 7.9

If {\mathcal{J}} is of even cotype this is Theorem 7.7. If {\mathcal{J}} is of odd cotype we have to show that the choice of T-duality did not matter. But the composition of any two type-changing T-dualities is conjugation by an element of SO(2n,2n;\mathbb{Z}), so the first result follows from Theorem 7.10.

Invariance under O(2n, 2n; \mathbb{Z}) follows as T-duality relative to a circle is represented by a matrix with negative determinant, see Remark 3.0, generating O(2n,2n; \mathbb{Z}).

Note that we formulated Theorem 7.10 for automorphisms while T-duality involves a change of vector space. We will be lax about this point as it does not affect the proofs but would complicate notation.

Corollary

Corollary 7.10

R86Q

The generalized derived category {\mathcal{D}}({\mathcal{J}}) is invariant under integral B-transforms and integral T-duality.

Used by Proof
Proof of Corollary 7.10

The first statement is clear from the theorem. The fact that T-duality corresponds to conjugation with an element in SO(2n,2n;\mathbb{Z}) is Lemma 3.5.

Remark

Remark 7.0

2JAU

Note that this should not be confused with the action of SO(2n,2n;\mathbb{Z}) on non-commutative tori by fractional linear transforms by Morita equivalences [rieffel1999morita].

Theorem

Theorem 8.1

ICEZ

Given any translation invariant GC structure \mathcal J on the torus let \widehat {\mathcal{J}} = S{\mathcal{J}} S^{-1} be its mirror partner. Then

{\mathcal{D}}({\mathcal{J}}) \simeq {\mathcal{D}}(\widehat {\mathcal{J}}).
Used by Proof, Proof
Proof of Theorem 8.1

This is a direct corollary of Corollary 7.10 as S \in SO(2n,2n; \mathbb{Z}).

Remark

Remark 8.0

35LG

The reason to restrict attention to affine Lagrangians is that the infinite series needed to define the Fukaya category do not in general converge over \mathbb{C}. We may thus think of {\operatorname{Fuk}}_{\mathrm{aff}}(X) as a convergent Fukaya category. The true Fukaya category of X should be defined over the Novikov ring, where its objects are generated by all Lagrangians. Then the mirror is a rigid analytic variety over the Novikov ring, and the full statement of homological mirror symmetry was proven in the case of the standard symplectic 4-torus [Abouzaid10a] and is expected to hold in general – without any addition of co-isotropic branes.

Theorem

Theorem 8.3

NHON

Let X=(V/\Gamma,\omega + \mathbf{i} B) be a B-symplectic torus admitting the structure of an affine Lagrangian fibration with an affine Lagrangian section and let {\mathcal{J}}_{\omega + \mathbf{i} B} = e^{B} {\mathcal{J}}_{\omega} e^{-B} be the associated GC structure. Let \check X be the complex torus mirror to X. Then

\operatorname{Ho} {\mathcal{D}}({\mathcal{J}}_{\omega+\mathbf{i} B}) \cong {\operatorname{D^b_{coh}}}(\check X).
Proof

Proof

I9LV

The existence of the affine Lagrangian fibration with an affine Lagrangian section may be re-written as the existence of a Lagrangian splitting V = L \oplus M with L and M rational. Then \omega + \mathbf{i} B is block off-diagonal and conjugation with S turns {\mathcal{J}}_{\omega + \mathbf{i} B} into a purely complex GC structure {\mathcal{J}}_J (on L^* / \Gamma^*_L \times B / \Gamma_B). The mirror \check X of (V/\Gamma,\omega + \mathbf{i} B) is given by equipping L^* / \Gamma^*_L \times B / \Gamma_B with the complex structure J.

By Theorem 8.1 we have {\mathcal{D}}({\mathcal{J}}_{\omega + \mathbf{i} B}) \cong {\mathcal{D}}({\mathcal{J}}_J) and Example 7.2 shows that {\mathcal{D}}({\mathcal{J}}_J) \cong {\operatorname{D^b_{coh}}}(\check X).

Conjecture

Conjecture 8.1 (Homological mirror symmetry)

ORLE

Let X be the 2n-torus equipped with a translation-invariant B-symplectic structure \omega + \mathbf{i} B admitting the structure of a Lagrangian fibration together with a Lagrangian splitting. Then

{\operatorname{Fuk}}'(X) \cong {\operatorname{D^b_{coh}}}(\check X).
Corollary

Corollary 8.3

ODAH

Let X be the 2n-torus equipped with a translation-invariant B-symplectic structure \omega + \mathbf{i} B admitting the structure of a Lagrangian fibration with a Lagrangian splitting. If homological mirror symmetry holds then {\mathcal{D}}({\mathcal{J}}_{\omega + \mathbf{i} B}) \simeq {\operatorname{Fuk}}'(X).

Corollary

Corollary 8.4

D1MF

Let X be a B-symplectic torus admitting an affine Lagrangian and let {\mathcal{J}}_{\omega + \mathbf{i} B} be the associated GC structure. Then the mirror is a noncommutative torus. Moreover there is a complex torus \check X with a 2-form B' such that {\mathcal{D}}({\mathcal{J}}_{\omega+\mathbf{i} B}) is equivalent to the derived category of coherent B'-twisted sheaves on \check X.

Proof of Corollary 8.4

This is another corollary of Theorem 8.1. If there is any affine Lagrangian in X we may choose a Lagrangian lift to V and a possibly non-Lagrangian complement M with \Gamma = (\Gamma \cap L) \oplus (\Gamma \cap M). With respect to the splitting V = L \oplus M, S sends {\mathcal{J}}_{\omega + \mathbf{i} B} to the GC structure of a non-commutative torus as a mirror.

The computation is that if \omega + \mathbf{i} B admits an affine Lagrangian it may be written so that the top left quadrant of its matrix representation is 0

and the schematic computation

S {\mathcal{J}}_{\omega+\mathbf{i} B}S^{-1} = S \begin{pmatrix} 0 & 0 & 0 & * \\ 0 & 0 & * & * \\ * & * & 0 & 0 \\ *& 0 & 0 & 0 \end{pmatrix} S^{-1} = \begin{pmatrix} 0 & * & * & 0 \\ * & 0 & 0 & * \\ 0 & 0 & 0 & * \\ 0& 0 & * & 0 \end{pmatrix}

gives the \beta-transform of a purely complex torus, i.e. a noncommutative torus.

Switching the order of the basis leads to a B'-transform of a complex torus, thus the Fukaya category agrees with the derived category of twisted sheaves, see Example 7.3.

Conjecture

Conjecture 8.2 (Noncommutative homological mirror symmetry)

FKBI

Let X be the 2n-torus equipped with a translation-invariant B-symplectic structure \omega + \mathbf{i} B Then we have

{\operatorname{Fuk}}'(X) \cong {\mathcal{D}}(S {\mathcal{J}}_{\omega + \mathbf{i} B} S^{-1}).
Example

Example 8.5

NVLJ

Let (V/\Gamma, \omega) have a space-filling brane, i.e. a line bundle equipped with a connection whose curvature F satisfies

F\omega^{-1}F + \omega = 0,

see [Kapustin03] and [kapustin2005branes]. Then one computes

e^{F} L_{{\mathcal{J}}_\omega} = e^{\frac{1}{2} (F)^{-1}} L_{{\mathcal{J}}_J}

with

J = \omega^{-1}F

and by integral invariance (Theorem 7.10) we have an equivalence

\mathcal D(\mathcal J_{\omega}) \cong \mathcal P_{\mathbf A(J, \frac{1}{2} F^{-1}, 0)}.

As this construction removed the curvature there is a free rank 1 module over {\mathbf A(J, \frac{1}{2} F^{-1}, 0)}, which we can think of as representing the space-filling brane as an object in {\operatorname{Fuk}}(X, \omega). The derived endomorphism space is {\mathbf A(J, \frac{1}{2} F^{-1}, 0)}.

This agrees with the predictions in the mathematical physics literature. There the endomorphism space represents the algebra \mathcal A of (\mathcal B_{cc}, \mathcal B_{cc})-strings, with \mathcal B_{cc} standing for the space-filling brane, see [Gukov:2008ve, Section 1.5]. With notation as above there is a holomorphic symplectic structure \Omega = F + \mathbf{i} \omega and \mathcal A is expected to be the deformation quantization of the Dolbeault algebra controlled by the inverse holomorphic bivector \Omega^{-1} (where we view \Omega as a map from holomorphic tangent to holomorphic cotangent space), see [Gukov:2008ve, Section 2.2] and [kapustin2007electric, Section 11.1]. But as in [kapustin2005branes, Section 3.2] one finds \Omega^{-1} = \frac 1 4(F^{-1} - \mathbf{i} \omega^{-1}) and this is exactly the (2,0)-part of the real bivector \frac 1 2 F^{-1}. Thus the deformation on holomorphic functions induced by \frac 1 2 F^{-1} is the one determined by \Omega^{-1}, as expected.

Definition

Definition 8.6

43GH

An affine Lagrangian object in (V/\Gamma, \omega) is an embedded affine Lagrangian torus in V equipped with a unitary local system of rank 1. It is thus determined by a triple (L,b,\alpha) where

  1. L \subset V is a linear Lagrangian subspace of V which is rational with respect to \Gamma, i.e.,

    \Gamma_L := L \cap \Gamma \cong \mathbb Z^{2n}.
  2. b \in V/L is a base vector,

  3. \alpha \in L^* is a real 1-form along L.

The image of L+b in V/\Gamma is an embedded Lagrangian torus which will be denoted as L_b.

Lemma

Lemma 8.7

CVP5

b and \alpha determine a projectively flat connection of curvature

2\pi \mathbf{i} \omega_K(L) := 2\pi \mathbf{i} \begin{pmatrix} 0 & 0 \\ 0 & \omega_K|_{IL \times IL} \end{pmatrix} \colon L \oplus IL \to L^* \oplus (IL)^*

on N^0(L) and \omega_K(L)^{(2,0)_I + (0,2)_I} = \omega_K / 2.

Proof of Lemma 8.7

Define \nabla \colon (L \oplus L_K' \oplus L_{K}) \times N^0(L) \to N^0(L) as

\nabla_{(t,l_k',l_k)} g(l,v) := \partial_{(t,l_k')}g(l,v) + 2 \pi \mathbf{i} [\omega_K(l_k+l_k',v+b)+\alpha(t)] g(l,v).

One verifies the Leibniz rule and that the curvature is indeed given by 2\pi \mathbf{i} \omega_K(L). For the final statement we note that \omega_K(L) differs from \frac{1}{2}\omega_K by \frac{1}{2}\begin{pmatrix} -\omega_K|_{L \times L}& 0 \\ 0 & \omega_K|_{IL \times IL} \end{pmatrix} which has type (1,1) with respect to I.

Definition

Definition 8.8

WIWN

We call the cohesive module, associated to an affine Lagrangian object (L,b,\alpha) obtained by endowing N^0(L) with the (0,1)-part of the connection constructed in the proof of Lemma 8.7, the Lagrangian module and denote it by

N(L,b,\alpha) = (N^0(L) \otimes \wedge^{(0,\bullet)} V_{\mathbb C}^*, \nabla^{(0,1)} ).

Note that \nabla depends on b and \alpha.

Remark

Remark 8.2

NTBY

One can also explicitly construct modules associated to Lagrangians which are complex with respect to I. We will focus on the totally real Lagrangians as this condition is open and leads to more accessible computations.

However, as the construction is of some geometric interest we now describe modules associated to complex Lagrangians.

Assume we have (X, \omega) with \omega = \omega_J as part of a Hyperkähler structure (g, I, J, K). Take an I-complex Lagrangian L on (X, \omega). It is also \omega_K-Lagrangian and thus functions on L can be used as a module over (C^\infty(T^{2n},{\omega_K^{-1}}) As L is I-complex we obtain an action of (T^*)^{0,1} and as \omega_K vanishes this gives rise to a curved module N^0(L) for \mathbf{A}(\omega). This module is not cohesive, but it admits a cohesive resolution.

The intuition here is that a non-commutative module generically needs to be coisotropic and a curved module generically needs to be isotropic, so a Lagrangian can be both.

Remark

Remark 8.2

M6D7

Recall that the choice of a Maslov index \eta(L_1,L_2) for a pair of Lagrangians depends on additional choices. Typically one chooses a lift of L to the universal cover of the Lagrangian Grassmannian. Another way of obtaining a \mathbb Z-grading is explained in [FukayaPre] which requires the choice of an auxiliary Lagrangian subspace L_{\mathrm{pt}} (see 8.19). It turns out that the Hyperkähler structure chosen here also yields a notion of Maslov index \eta_{\mathrm{HK}}(L_0,L_1), see Definition 8.12. We will show in Corollary 8.21 that there is a map L \mapsto \tau(L) \in \mathbb Z such that

\eta_{\mathrm{HK}}(L_1,L_2) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \tau(L_1) - \tau(L_2).

This relationship means that the choices \eta_{\mathrm{Fuk}} and \eta_{\mathrm{HK}} are compatible: If we shift each Lagrangian L in the derived category by \tau(L) (see Definition 8.15), then the gradings induced by \eta_{\mathrm{HK}} and \eta_{\mathrm{Fuk}} coincide.

Proposition

Proposition 8.9

QPXO

There is an isomorphism

N(L,b,\alpha)^\vee \cong N(L,-b,-\alpha)

of \mathbf A(I,\omega_K^{-1},-\omega_K/2)-modules.

Proof (Proof sketch).

Consider the subalgebra \mathcal S(\Lambda_L^*, -\omega_K^{-1}) \subset \mathcal S(\Gamma^*, -\omega_K^{-1}) = C^\infty( V/\Gamma, -\omega_K^{-1}). We have that

S(\Gamma^*, -\omega_K^{-1}) \widehat \otimes_{\mathcal S(\Lambda_L^*, -\omega_K^{-1})} H(-\omega_K^{-1}) = N(L) \quad( = (N(L,b,\alpha))^0 )

where H(-\omega_K^{-1}) is the Heisenberg-module constructed in ?. The statement of the lemma then essentially follows from the following claim: For any non-commutative torus V/\Gamma on an invertible bivector \sigma, there is an isomorphism

\operatorname{Hom}_{C^\infty(V/\Gamma,\sigma)}(H(\sigma), C^\infty(V/\Gamma,\sigma)) = H(\sigma)^\vee \cong H(-\sigma)

of C^\infty(V/\Gamma,-\sigma)-modules. To prove the claim, choose a Lagrangian splitting V = L \oplus L'. Then

\begin{aligned} &H(\sigma) = \mathcal S(L), \quad (u_\gamma.f)(l) = u_\gamma(l + \frac{1}{2} \sigma(\gamma|_{L'})) f(l+\sigma(\gamma|_{L'})) \\&H(-\sigma) = \mathcal S(L), \quad (u_\gamma.g)(l) = u_\gamma(l - \frac{1}{2} \sigma(\gamma|_{L'})) g(l-\sigma(\gamma|_{L'})). \end{aligned}

We define the isomorphism \Phi \colon H(-\sigma) \to H(\sigma)^\vee as:

\Phi \colon f \mapsto \left(g \mapsto \sum_{\gamma \in \Gamma^*} \int_L f(l) (u_{-\gamma}.g)(l) \cdot u_\gamma\right)

(which depends on the choice of a volume form on L). For every \gamma we have the following identity:

\int_L (u_{\gamma}.f)(l) g(l) = \int_L f(l) (u_{\gamma}.g)(l).

It follows from this formula that \Phi is C^\infty(V/\Gamma,-\sigma)-linear in f and C^\infty(V/\Gamma,\sigma)-linear in g.

The proof of the lemma is completed by comparing connections.

Proof

Proof (Proof sketch)

K77M

Consider the subalgebra \mathcal S(\Lambda_L^*, -\omega_K^{-1}) \subset \mathcal S(\Gamma^*, -\omega_K^{-1}) = C^\infty( V/\Gamma, -\omega_K^{-1}). We have that

S(\Gamma^*, -\omega_K^{-1}) \widehat \otimes_{\mathcal S(\Lambda_L^*, -\omega_K^{-1})} H(-\omega_K^{-1}) = N(L) \quad( = (N(L,b,\alpha))^0 )

where H(-\omega_K^{-1}) is the Heisenberg-module constructed in ?. The statement of the lemma then essentially follows from the following claim: For any non-commutative torus V/\Gamma on an invertible bivector \sigma, there is an isomorphism

\operatorname{Hom}_{C^\infty(V/\Gamma,\sigma)}(H(\sigma), C^\infty(V/\Gamma,\sigma)) = H(\sigma)^\vee \cong H(-\sigma)

of C^\infty(V/\Gamma,-\sigma)-modules. To prove the claim, choose a Lagrangian splitting V = L \oplus L'. Then

\begin{aligned} &H(\sigma) = \mathcal S(L), \quad (u_\gamma.f)(l) = u_\gamma(l + \frac{1}{2} \sigma(\gamma|_{L'})) f(l+\sigma(\gamma|_{L'})) \\&H(-\sigma) = \mathcal S(L), \quad (u_\gamma.g)(l) = u_\gamma(l - \frac{1}{2} \sigma(\gamma|_{L'})) g(l-\sigma(\gamma|_{L'})). \end{aligned}

We define the isomorphism \Phi \colon H(-\sigma) \to H(\sigma)^\vee as:

\Phi \colon f \mapsto \left(g \mapsto \sum_{\gamma \in \Gamma^*} \int_L f(l) (u_{-\gamma}.g)(l) \cdot u_\gamma\right)

(which depends on the choice of a volume form on L). For every \gamma we have the following identity:

\int_L (u_{\gamma}.f)(l) g(l) = \int_L f(l) (u_{\gamma}.g)(l).

It follows from this formula that \Phi is C^\infty(V/\Gamma,-\sigma)-linear in f and C^\infty(V/\Gamma,\sigma)-linear in g.

The proof of the lemma is completed by comparing connections.

Lemma

Lemma 8.10

HHH2

We have an isomorphism

N^0(L \times \tilde L) / \Gamma^* \cong \mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^* \times C_1).
Proof.

Consider the intermediate quotient

\mathrm{id} \otimes p \otimes \mathrm{id} \colon N^0(L \times \tilde L) \to N^0(L \times \tilde L) / (\Lambda_L \cap \Lambda_{\tilde L}) = \mathcal S(\Gamma^*_L \times \Gamma^*_{\tilde L}) \widehat \otimes \mathcal S((IL \cap I\tilde L)^+) / (\Lambda_L \cap \Lambda_{\tilde L}) \widehat \otimes \mathcal S(C_1),

where

p\colon\mathcal S((IL \cap I\tilde L)^+) \to \mathcal S((IL \cap I\tilde L)^+) / (\Lambda_L \cap \Lambda_{\tilde L})

is the quotient map and \Lambda_L \cap \Lambda_{\tilde L} acts as

\gamma.g(Il,Il) := u_{\gamma}.g(Il,Il) = g(Il-\omega_K^{-1}(\gamma),Il-\omega_K^{-1}(\gamma)).

We identify

\mathcal S((IL \cap I\tilde L)^+) / (\Lambda_L \cap \Lambda_{\tilde L}) \cong C^\infty(IL \cap I\tilde L / \omega_K^{-1}(\Lambda_L \cap \Lambda_{\tilde L})) \cong \mathcal S((\Lambda_L \cap \Lambda_{\tilde L})^*)

by g \mapsto (Il \mapsto \sum_{\lambda \in \omega_K^{-1}(\Lambda_L \cap \Lambda_{\tilde L})} g((Il+\lambda),(Il+\lambda))). We have

N^0(L \times \tilde L) / \Gamma^* = [N^0(L \times \tilde L) / (\Lambda_L \cap \Lambda_{\tilde L})] / (\Gamma^* / \Lambda_L \cap \Lambda_{\tilde L}).

Consider the subspace

\mathcal S(S) \otimes \mathrm{id} \colon \mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/\Gamma^*) \widehat \otimes \mathcal S((\Lambda_L \cap \Lambda_{\tilde L})^*) \widehat \otimes \mathcal S(C_1) \hookrightarrow N^0(L \times \tilde L) / (\Lambda_L \cap \Lambda_{\tilde L}).

The composition of \mathcal S(S) \otimes \mathrm{id} with the projection onto N^0(L \times \tilde L) / \Gamma^* is the desired isomorphism.

Proof

Proof

W4SE

Consider the intermediate quotient

\mathrm{id} \otimes p \otimes \mathrm{id} \colon N^0(L \times \tilde L) \to N^0(L \times \tilde L) / (\Lambda_L \cap \Lambda_{\tilde L}) = \mathcal S(\Gamma^*_L \times \Gamma^*_{\tilde L}) \widehat \otimes \mathcal S((IL \cap I\tilde L)^+) / (\Lambda_L \cap \Lambda_{\tilde L}) \widehat \otimes \mathcal S(C_1),

where

p\colon\mathcal S((IL \cap I\tilde L)^+) \to \mathcal S((IL \cap I\tilde L)^+) / (\Lambda_L \cap \Lambda_{\tilde L})

is the quotient map and \Lambda_L \cap \Lambda_{\tilde L} acts as

\gamma.g(Il,Il) := u_{\gamma}.g(Il,Il) = g(Il-\omega_K^{-1}(\gamma),Il-\omega_K^{-1}(\gamma)).

We identify

\mathcal S((IL \cap I\tilde L)^+) / (\Lambda_L \cap \Lambda_{\tilde L}) \cong C^\infty(IL \cap I\tilde L / \omega_K^{-1}(\Lambda_L \cap \Lambda_{\tilde L})) \cong \mathcal S((\Lambda_L \cap \Lambda_{\tilde L})^*)

by g \mapsto (Il \mapsto \sum_{\lambda \in \omega_K^{-1}(\Lambda_L \cap \Lambda_{\tilde L})} g((Il+\lambda),(Il+\lambda))). We have

N^0(L \times \tilde L) / \Gamma^* = [N^0(L \times \tilde L) / (\Lambda_L \cap \Lambda_{\tilde L})] / (\Gamma^* / \Lambda_L \cap \Lambda_{\tilde L}).

Consider the subspace

\mathcal S(S) \otimes \mathrm{id} \colon \mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/\Gamma^*) \widehat \otimes \mathcal S((\Lambda_L \cap \Lambda_{\tilde L})^*) \widehat \otimes \mathcal S(C_1) \hookrightarrow N^0(L \times \tilde L) / (\Lambda_L \cap \Lambda_{\tilde L}).

The composition of \mathcal S(S) \otimes \mathrm{id} with the projection onto N^0(L \times \tilde L) / \Gamma^* is the desired isomorphism.

Lemma

Lemma 8.11

VDJE

There is a canonical isomorphism

\Psi \colon V/(L \cap \tilde L \oplus IL \cap I \tilde L) \xrightarrow{\cong} Z = ((IL \cap I\tilde L)^+)^\perp / (IL \cap I \tilde L)^+,\quad v \mapsto (\pi_{IL}(v),\pi_{I \tilde L}(v)).

Hence, Z naturally carries a complex structure induced by I. Moreover, the symplectic form \omega' on Z is of type (1,1) for I.

Proof.

The map V \ni v \mapsto (\pi_{IL}(v), \pi_{I\tilde L}(v)) \in IL \oplus I\tilde L factors through the subspace ((IL \cap I\tilde L)^+)^\perp because \omega_K|_{IL \times L} = \omega_K|_{I\tilde L \times \tilde L} = 0. Its kernel is L \cap \tilde L, and the preimage of (IL \cap I\tilde L)^+ is IL \cap I \tilde L. This shows that \Psi is well-defined and injective. For dimensional reasons it is an isomorphism. The form \Psi^* \omega' is given by

(\Psi^* \omega')(v,v') = \omega_K((\pi_{I \tilde L}-\pi_{IL})(v),v').

Thus,

\begin{aligned} (\Psi^* \omega')(Iv,Iv') &= \omega_K((\pi_{I \tilde L}-\pi_{IL})(Iv),Iv')\\ &= \omega_K(I(\pi_{\tilde L}-\pi_{L})(v),Iv') \\ &= -\omega_K(I(\pi_{I\tilde L}-\pi_{IL})(v),Iv') \\ &= \omega_K((\pi_{I\tilde L}-\pi_{IL})(v),v') = (\Psi^* \omega')(v,v') \end{aligned}

as \pi_{I\tilde L}-\pi_{IL} = \pi_{L} - \pi_{\tilde L} and since \omega_K is of type (2,0)+(0,2) with respect to I.

Uses Definition 2.4 · Used by Proof
Proof

Proof

5SOE

The map V \ni v \mapsto (\pi_{IL}(v), \pi_{I\tilde L}(v)) \in IL \oplus I\tilde L factors through the subspace ((IL \cap I\tilde L)^+)^\perp because \omega_K|_{IL \times L} = \omega_K|_{I\tilde L \times \tilde L} = 0. Its kernel is L \cap \tilde L, and the preimage of (IL \cap I\tilde L)^+ is IL \cap I \tilde L. This shows that \Psi is well-defined and injective. For dimensional reasons it is an isomorphism. The form \Psi^* \omega' is given by

(\Psi^* \omega')(v,v') = \omega_K((\pi_{I \tilde L}-\pi_{IL})(v),v').

Thus,

\begin{aligned} (\Psi^* \omega')(Iv,Iv') &= \omega_K((\pi_{I \tilde L}-\pi_{IL})(Iv),Iv')\\ &= \omega_K(I(\pi_{\tilde L}-\pi_{L})(v),Iv') \\ &= -\omega_K(I(\pi_{I\tilde L}-\pi_{IL})(v),Iv') \\ &= \omega_K((\pi_{I\tilde L}-\pi_{IL})(v),v') = (\Psi^* \omega')(v,v') \end{aligned}

as \pi_{I\tilde L}-\pi_{IL} = \pi_{L} - \pi_{\tilde L} and since \omega_K is of type (2,0)+(0,2) with respect to I.

Definition

Definition 8.12

89ZN

Let L_1,L_2 be \omega_J Lagrangian subspaces which are totally real with respect to I. Hence V = L_1 \oplus IL_1 = L_2 \oplus IL_2. We define on V the quadratic form

t(v) = \omega_J( (\pi_{L_2} - \pi_{L_1})(v),v).

Define the Hyperkähler Maslov index

\eta_{\mathrm{HK}}(L_1,L_2) = \frac{1}{2} n^-(t)

where n^-(t) denotes the negative index of t, i.e., the number of negative eigenvalues.

Lemma

Lemma 8.13

U1OW

Let \omega' and I be as in Lemma 8.11. \omega'(I(-),(-)) defines a symmetric form on Z and we have the following:

\frac{1}{2} n^+(\omega'(I(-),(-))) = \eta_{\mathrm{HK}}(L,\tilde L)

where n^+ denotes the positive index.

Proof.

This follows from the proof of Lemma 8.11.

Used by Proof
Definition

Definition 8.14

31M7

Let C be a real vector space. Let \alpha be a polynomial closed 1-form on C. The Schwartz-de-Rham-Witten complex of \alpha is the real cochain complex

\mathcal S_{DRW}^\bullet(C, \alpha) := (\mathcal S_{\mathbb R}(C) \otimes \scalebox{1.2}{$\wedge$}^\bullet C^*, d_{DR} + \alpha \wedge (-) ). \qedhere
Definition

Definition 8.15

0CAG

Fix an I-totally real, \omega_J-Lagrangian subspace L_{\mathrm{pt}} \subset V. Given an affine Lagrangian (L,b,\alpha) we define the corrected Lagrangian module

\tilde N(L,b,\alpha) := N(L,b,\alpha)[-\tau(L)]

where \tau(L) = \eta_{\mathrm{HK}}(L,L_{\mathrm{pt}}).

Used by Proof
Proposition

Proposition 8.16

M2HG

There is a spectral sequence E^{pq}_* converging to the cohomology

E^{pq}_* \Rightarrow H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}(N(L,b,\alpha),N(\tilde L, \tilde b, \tilde \alpha))).

Its first page is

E_1^{pq} = H^q(L_b \cap \tilde L_{\tilde b}, \mathcal L_{\tilde \alpha - \alpha}) \otimes_{\mathbb R} \mathcal S^p_{DRW}(C_1, dh)

where h is a quadratic function on C_1 whose negative index is equal to the Hyperkähler Maslov index \eta_{\mathrm{HK}}(L, \tilde L). It collapses on the second page to

E_2^{pq} = \begin{cases} &H^q(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha}) \text{, if } p = \eta_{HK}(L,\tilde L) \\ &0 \text{, else. } \end{cases}
Proof.

On the exterior algebra \scalebox{1.2}{}^{(0,\bullet)} V^* we introduce the decreasing filtration

F^i [\scalebox{1.2}{$\wedge$}^{(0,\bullet)} V^*] = \scalebox{1.2}{$\wedge$}^{(0,\geq i)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \wedge \scalebox{1.2}{$\wedge$}^{(0,\bullet)} V^*.

Its filtered pieces are

F^i / F^{i+1} = \scalebox{1.2}{$\wedge$}^{(0,i)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*.

F induces a (bounded) filtration also on Q^\bullet:

F^i Q^\bullet = \mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^* \times C_1) \otimes F^i [\scalebox{1.2}{$\wedge$}^{(0,\bullet)} V^*]

The differential d = \nabla^{(0,1)} on Q^\bullet is compatible with it. The associated spectral sequence has terms

\begin{aligned} E_0^{pq} &= [\mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^*) \otimes \scalebox{1.2}{$\wedge$}^{(0,q)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*] \\ & \quad \widehat \otimes [\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,p)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^*] \end{aligned}

with differential d_0 \colon E_0^{pq} \to E_0^{p(q+1)} given by ?, in particular acting only on the first factor. Now

\mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^*) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*

is isomorphic to

\oplus_{t \in T} (\mathcal S((\Gamma_L \cap \Gamma_{\tilde L})^* \times (\Lambda_L \cap \Lambda_{\tilde L})^*) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*, \nabla^{(0,1)}|_{L \cap \tilde L \oplus IL \cap I \tilde L})

by equation ?. The summands are isomorphic to Dolbeault complexes with twisted differentials:

\Omega^{(0,\bullet)}( (L \cap \tilde L) / (\Gamma_L \cap \Gamma_{\tilde L}) \times (IL \cap I \tilde L) / \omega_K^{-1}(\Lambda_L \cap \Lambda_{\tilde L})), \bar \partial + 2\pi \mathbf{i} (\tilde \alpha - \alpha + \omega_K(b-\tilde b))^{(0,1)} ).

Its cohomology is hence

= \begin{cases} &\scalebox{1.2}{$\wedge$}^\bullet (L \cap \tilde L)_{\mathbb C} \text{, if } (\tilde \alpha - \alpha)|_{L \cap \tilde L} \in (\Gamma_L \cap \Gamma_{\tilde L})^* \text{ and } L_b \cap \tilde L_{\tilde b} \neq \emptyset \\ & 0 \text{, else. } \end{cases}

Moreover, as T parametrises the connected components of L_b \cap \tilde L_{\tilde b}, we get

H^\bullet(L_b \cap \tilde L_{\tilde b}, \mathcal L_{\tilde \alpha - \alpha})

after summing over T (here \mathcal L_{\tilde \alpha - \alpha} denotes the local system associated to the 1-form \tilde \alpha - \alpha). Hence, the first page of the spectral sequence reads

E_1^{pq} = H^q(L_b \cap \tilde L_{\tilde b}, \mathcal L_{\tilde \alpha - \alpha}) \otimes [\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,p)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^*]

and the differential E_1^{pq} \to E_1^{(p+1)q} is induced by \tilde \nabla, see ?, and acts on the latter factor only. Recall that

C_1 \oplus C_2 = Z \cong V/(L \cap \tilde L \oplus IL \cap I \tilde L)

is a Lagrangian splitting for \omega' and \omega' is of type (1,1) for I. We can choose C_1 to be totally real for I and C_2 := IC_1. Then

\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \cong (\mathcal S_{\mathbb R}(C_1) \otimes_{\mathbb R} \scalebox{1.2}{$\wedge$}^\bullet C_1) \otimes_{\mathbb R} \mathbb C

where the differential on \mathcal S_{\mathbb R}(C_1) \otimes_{\mathbb R} \scalebox{1.2}{}^\bullet C_1 reads

d \colon C_1 \times \mathcal S_{\mathbb R}(C_1) \to \mathcal S_{\mathbb R}(C_1), \quad d_c(g)(v) = \partial_c g(v) - 2\pi \omega'(Ic,v) g(v).

Note that h := \omega'(I(-),(-)) is a symmetric non-degenerate bilinear form on C_1. Diagonalise h as

A h A^{*} = \mathrm{diag}(-1,...,-1,1,...,1).

Let r be the number of negative eigenvalues and s the number of positive eigenvalues. By Lemma 8.13, s coincides with the Hyperkähler Maslov index \eta_{\mathrm{HK}}(L_1,L_2). Define the complexes

S^- := \mathcal S_{\mathbb R}(\mathbb R) \xrightarrow{\partial_x + 2\pi x } \mathcal S_{\mathbb R}(\mathbb R),\quad S^+ := \mathcal S_{\mathbb R}(\mathbb R) \xrightarrow{\partial_x - 2\pi x } \mathcal S_{\mathbb R}(\mathbb R).

Using A, we obtain an isomorphism of Fréchet complexes \mathcal S_{\mathbb R}(C_1) \otimes_{\mathbb R} \scalebox{1.2}{}^\bullet C_1 \cong (S^-)^{\widehat \otimes r} \widehat \otimes (S^+)^{\widehat \otimes s}. There are continuous homotopy equivalences

S^- \simeq (\mathbb R \to 0) = \mathbb R,\quad S^+ \simeq (0 \to \mathbb R) =\mathbb R[1].

We conclude that

\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \simeq \mathbb C[1]^{\otimes s} = \mathbb C[s]

so

E_2^{pq} = \begin{cases} &H^q(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha}), \text{ if } p = s \\ &0, \text{ else. } \end{cases} \qedhere
Uses Definition 8.12 · Used by Proof
Proof

Proof

ZILN

On the exterior algebra \scalebox{1.2}{}^{(0,\bullet)} V^* we introduce the decreasing filtration

F^i [\scalebox{1.2}{$\wedge$}^{(0,\bullet)} V^*] = \scalebox{1.2}{$\wedge$}^{(0,\geq i)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \wedge \scalebox{1.2}{$\wedge$}^{(0,\bullet)} V^*.

Its filtered pieces are

F^i / F^{i+1} = \scalebox{1.2}{$\wedge$}^{(0,i)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*.

F induces a (bounded) filtration also on Q^\bullet:

F^i Q^\bullet = \mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^* \times C_1) \otimes F^i [\scalebox{1.2}{$\wedge$}^{(0,\bullet)} V^*]

The differential d = \nabla^{(0,1)} on Q^\bullet is compatible with it. The associated spectral sequence has terms

\begin{aligned} E_0^{pq} &= [\mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^*) \otimes \scalebox{1.2}{$\wedge$}^{(0,q)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*] \\ & \quad \widehat \otimes [\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,p)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^*] \end{aligned}

with differential d_0 \colon E_0^{pq} \to E_0^{p(q+1)} given by ?, in particular acting only on the first factor. Now

\mathcal S((\Gamma^*_L \times \Gamma^*_{\tilde L})/ \Gamma^* \times (\Lambda_L \cap \Lambda_{\tilde L})^*) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*

is isomorphic to

\oplus_{t \in T} (\mathcal S((\Gamma_L \cap \Gamma_{\tilde L})^* \times (\Lambda_L \cap \Lambda_{\tilde L})^*) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (L \cap \tilde L \oplus IL \cap I \tilde L)^*, \nabla^{(0,1)}|_{L \cap \tilde L \oplus IL \cap I \tilde L})

by equation ?. The summands are isomorphic to Dolbeault complexes with twisted differentials:

\Omega^{(0,\bullet)}( (L \cap \tilde L) / (\Gamma_L \cap \Gamma_{\tilde L}) \times (IL \cap I \tilde L) / \omega_K^{-1}(\Lambda_L \cap \Lambda_{\tilde L})), \bar \partial + 2\pi \mathbf{i} (\tilde \alpha - \alpha + \omega_K(b-\tilde b))^{(0,1)} ).

Its cohomology is hence

= \begin{cases} &\scalebox{1.2}{$\wedge$}^\bullet (L \cap \tilde L)_{\mathbb C} \text{, if } (\tilde \alpha - \alpha)|_{L \cap \tilde L} \in (\Gamma_L \cap \Gamma_{\tilde L})^* \text{ and } L_b \cap \tilde L_{\tilde b} \neq \emptyset \\ & 0 \text{, else. } \end{cases}

Moreover, as T parametrises the connected components of L_b \cap \tilde L_{\tilde b}, we get

H^\bullet(L_b \cap \tilde L_{\tilde b}, \mathcal L_{\tilde \alpha - \alpha})

after summing over T (here \mathcal L_{\tilde \alpha - \alpha} denotes the local system associated to the 1-form \tilde \alpha - \alpha). Hence, the first page of the spectral sequence reads

E_1^{pq} = H^q(L_b \cap \tilde L_{\tilde b}, \mathcal L_{\tilde \alpha - \alpha}) \otimes [\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,p)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^*]

and the differential E_1^{pq} \to E_1^{(p+1)q} is induced by \tilde \nabla, see ?, and acts on the latter factor only. Recall that

C_1 \oplus C_2 = Z \cong V/(L \cap \tilde L \oplus IL \cap I \tilde L)

is a Lagrangian splitting for \omega' and \omega' is of type (1,1) for I. We can choose C_1 to be totally real for I and C_2 := IC_1. Then

\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \cong (\mathcal S_{\mathbb R}(C_1) \otimes_{\mathbb R} \scalebox{1.2}{$\wedge$}^\bullet C_1) \otimes_{\mathbb R} \mathbb C

where the differential on \mathcal S_{\mathbb R}(C_1) \otimes_{\mathbb R} \scalebox{1.2}{}^\bullet C_1 reads

d \colon C_1 \times \mathcal S_{\mathbb R}(C_1) \to \mathcal S_{\mathbb R}(C_1), \quad d_c(g)(v) = \partial_c g(v) - 2\pi \omega'(Ic,v) g(v).

Note that h := \omega'(I(-),(-)) is a symmetric non-degenerate bilinear form on C_1. Diagonalise h as

A h A^{*} = \mathrm{diag}(-1,...,-1,1,...,1).

Let r be the number of negative eigenvalues and s the number of positive eigenvalues. By Lemma 8.13, s coincides with the Hyperkähler Maslov index \eta_{\mathrm{HK}}(L_1,L_2). Define the complexes

S^- := \mathcal S_{\mathbb R}(\mathbb R) \xrightarrow{\partial_x + 2\pi x } \mathcal S_{\mathbb R}(\mathbb R),\quad S^+ := \mathcal S_{\mathbb R}(\mathbb R) \xrightarrow{\partial_x - 2\pi x } \mathcal S_{\mathbb R}(\mathbb R).

Using A, we obtain an isomorphism of Fréchet complexes \mathcal S_{\mathbb R}(C_1) \otimes_{\mathbb R} \scalebox{1.2}{}^\bullet C_1 \cong (S^-)^{\widehat \otimes r} \widehat \otimes (S^+)^{\widehat \otimes s}. There are continuous homotopy equivalences

S^- \simeq (\mathbb R \to 0) = \mathbb R,\quad S^+ \simeq (0 \to \mathbb R) =\mathbb R[1].

We conclude that

\mathcal S(C_1) \otimes \scalebox{1.2}{$\wedge$}^{(0,\bullet)} (V/(L \cap \tilde L \oplus IL \cap I \tilde L)^* \simeq \mathbb C[1]^{\otimes s} = \mathbb C[s]

so

E_2^{pq} = \begin{cases} &H^q(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha}), \text{ if } p = s \\ &0, \text{ else. } \end{cases} \qedhere
Theorem

Theorem 8.22

M0HW

Let (L,b,\alpha) and (\tilde L, \tilde b, \tilde \alpha) be affine Lagrangian objects. Then the derived Hom between \tilde N(L,b,\alpha) and \tilde N(\tilde L,\tilde b, \tilde \alpha) coincides with Floer cohomology:

H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}(\tilde N(L,b,\alpha),\tilde N(\tilde L, \tilde b, \tilde \alpha))) \cong HF^{\bullet}( (L,b,\alpha), (\tilde L, \tilde b, \tilde \alpha) ).
Proof.

From the spectral sequence in Proposition 8.16 we find that

H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}( N(L,b,\alpha),N(\tilde L, \tilde b, \tilde \alpha))) \cong H^{\bullet-\eta_{\mathrm{HK}}(L,\tilde L)}(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha})

whereas

HF^\bullet((L,b,\alpha),(\tilde L, \tilde b, \tilde \alpha)) = H^{\bullet-\eta_{\mathrm{Fuk}}(L,\tilde L)}(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha}),

according to equation ?. Hence by Definition 8.15 and Corollary 8.21 below it follows that

H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}(\tilde N(L,b,\alpha),\tilde N(\tilde L, \tilde b, \tilde \alpha))) \cong HF^{\bullet}( (L,b,\alpha), (\tilde L, \tilde b, \tilde \alpha) ). \qedhere
Proof

Proof

9E4X

From the spectral sequence in Proposition 8.16 we find that

H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}( N(L,b,\alpha),N(\tilde L, \tilde b, \tilde \alpha))) \cong H^{\bullet-\eta_{\mathrm{HK}}(L,\tilde L)}(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha})

whereas

HF^\bullet((L,b,\alpha),(\tilde L, \tilde b, \tilde \alpha)) = H^{\bullet-\eta_{\mathrm{Fuk}}(L,\tilde L)}(L_b \cap \tilde L_{\tilde b},\mathcal L_{\tilde \alpha - \alpha}),

according to equation ?. Hence by Definition 8.15 and Corollary 8.21 below it follows that

H^\bullet(\operatorname{Hom}^\bullet_{\mathbf A(I,-\omega_K^{-1},\omega_K/2)}(\tilde N(L,b,\alpha),\tilde N(\tilde L, \tilde b, \tilde \alpha))) \cong HF^{\bullet}( (L,b,\alpha), (\tilde L, \tilde b, \tilde \alpha) ). \qedhere
Definition

Definition 8.18

EXUG

Let L_0,L_1,L_2 \subset V be Lagrangian subspaces. Define the quadratic form

q \colon L_1 \oplus L_2 \oplus L_3 \to \mathbb R, \quad q(x_1,x_2,x_3) = \omega(x_1,x_2) + \omega(x_2,x_3) + \omega(x_3,x_1).

The negative Kashiwara index is defined as

\tau^-(L_1,L_2,L_3) := n^-(q)

where n^- is the negative index of q (i.e., the number of negative eigenvalues).

Definition

Definition 8.19

6QHH

Let L_1,L_2 \subset V be Lagrangian subspaces. Define Fukaya’s Maslov index (relative to L_{\mathrm{pt}}) as

\eta_{\mathrm{Fuk}}(L_1,L_2) := 2n - \tau^-(L_1,L_2,L_{\mathrm{pt}}) - \dim(L_2 \cap L_{\mathrm{pt}}) - \dim(L_1 \cap L_{2}).

where 2n = \dim_{\mathbb R}(V).

Proposition

Proposition 8.20

S8R1

Let \eta \in \{\eta_{\mathrm{HK}},\eta_{\mathrm{Fuk}} \}. Then

  1. \tau^-(L_1,L_2,L_3) = \eta(L_2,L_1) + \eta(L_3,L_2) + \eta(L_1,L_3),

  2. \eta(L_1,L_2) + \eta(L_2,L_1) = n - \dim(L_1 \cap L_2).

Proof.

The statements for \eta = \eta_{\mathrm{Fuk}} follow from [FukayaPre, Lemma-Definition 2.12]. Let us now show the identities for \eta = \eta_{\mathrm{HK}}. The equality \eta_{\mathrm{HK}}(L_1,L_2) + \eta_{\mathrm{HK}}(L_2,L_1) = n - \dim(L_1 \cap L_2) follows from the definition of \eta_{\mathrm{HK}} (as the kernel of the quadratic form is L_1 \cap L_2 \oplus IL_1 \cap IL_2). It remains to show that

\tau^-(L_1,L_2,L_3) = \eta_{\mathrm{HK}}(L_2,L_1) + \eta_{\mathrm{HK}}(L_3,L_2) + \eta_{\mathrm{HK}}(L_1,L_3).

We define on V \oplus V \oplus V the quadratic form

t(v_1,v_2,v_3) = \omega((\pi_{L_1}-\pi_{L_2})(v_1),v_2) + \omega((\pi_{L_2}-\pi_{L_3})(v_2),v_3) + \omega((\pi_{L_3}-\pi_{L_1})(v_3),v_1).

We decompose

V \oplus V \oplus V = (L_1 \oplus IL_1) \oplus (L_2 \oplus IL_2) \oplus (L_3 \oplus IL_3).

We have

\begin{aligned} t((l_1+il_1'),(l_2+il_2'),(l_1+il_2') = &\omega(l_1,l_2) + \omega(l_2,l_3) + \omega(l_3,l_1) \\ &+ \omega(l_1',l_2') + \omega(l_2',l_3') + \omega(l_3',l_1'). \end{aligned}

Hence the negative index of t is

n^-(t) = 2 \tau^-(L_1,L_2,L_3).

Now define an isomorphism \Psi \colon V \oplus V \oplus V \to V \oplus V \oplus V

\begin{aligned} \Psi = \frac{1}{2} \begin{pmatrix} 1 & 1 & -1 \\ 1 & -1 & -1 \\ 1 & -1 & 1 \end{pmatrix} = \begin{pmatrix} 1 & 0 & 1 \\ 1 & -1 & 0 \\ 0 & -1 & 1\end{pmatrix}^{-1}. \end{aligned}

Writing \Psi(v_1,v_2,v_3) = (u_1,u_2,u_3) we find that

\begin{aligned} t(v_1,v_2,v_3) = \omega((\pi_{L_1}-\pi_{L_2})(u_1),u_1) + \omega((\pi_{L_2}-\pi_{L_3})(u_2),u_2) + \omega((\pi_{L_3}-\pi_{L_1})(u_3),u_3) \end{aligned}

and so

2 \tau^-(L_1,L_2,L_3) = n^-(t) = 2 (\eta_{\mathrm{HK}}(L_2,L_1) + \eta_{\mathrm{HK}}(L_3,L_2) + \eta_{\mathrm{HK}}(L_1,L_3)). \qedhere
Proof

Proof

3XNY

The statements for \eta = \eta_{\mathrm{Fuk}} follow from [FukayaPre, Lemma-Definition 2.12]. Let us now show the identities for \eta = \eta_{\mathrm{HK}}. The equality \eta_{\mathrm{HK}}(L_1,L_2) + \eta_{\mathrm{HK}}(L_2,L_1) = n - \dim(L_1 \cap L_2) follows from the definition of \eta_{\mathrm{HK}} (as the kernel of the quadratic form is L_1 \cap L_2 \oplus IL_1 \cap IL_2). It remains to show that

\tau^-(L_1,L_2,L_3) = \eta_{\mathrm{HK}}(L_2,L_1) + \eta_{\mathrm{HK}}(L_3,L_2) + \eta_{\mathrm{HK}}(L_1,L_3).

We define on V \oplus V \oplus V the quadratic form

t(v_1,v_2,v_3) = \omega((\pi_{L_1}-\pi_{L_2})(v_1),v_2) + \omega((\pi_{L_2}-\pi_{L_3})(v_2),v_3) + \omega((\pi_{L_3}-\pi_{L_1})(v_3),v_1).

We decompose

V \oplus V \oplus V = (L_1 \oplus IL_1) \oplus (L_2 \oplus IL_2) \oplus (L_3 \oplus IL_3).

We have

\begin{aligned} t((l_1+il_1'),(l_2+il_2'),(l_1+il_2') = &\omega(l_1,l_2) + \omega(l_2,l_3) + \omega(l_3,l_1) \\ &+ \omega(l_1',l_2') + \omega(l_2',l_3') + \omega(l_3',l_1'). \end{aligned}

Hence the negative index of t is

n^-(t) = 2 \tau^-(L_1,L_2,L_3).

Now define an isomorphism \Psi \colon V \oplus V \oplus V \to V \oplus V \oplus V

\begin{aligned} \Psi = \frac{1}{2} \begin{pmatrix} 1 & 1 & -1 \\ 1 & -1 & -1 \\ 1 & -1 & 1 \end{pmatrix} = \begin{pmatrix} 1 & 0 & 1 \\ 1 & -1 & 0 \\ 0 & -1 & 1\end{pmatrix}^{-1}. \end{aligned}

Writing \Psi(v_1,v_2,v_3) = (u_1,u_2,u_3) we find that

\begin{aligned} t(v_1,v_2,v_3) = \omega((\pi_{L_1}-\pi_{L_2})(u_1),u_1) + \omega((\pi_{L_2}-\pi_{L_3})(u_2),u_2) + \omega((\pi_{L_3}-\pi_{L_1})(u_3),u_3) \end{aligned}

and so

2 \tau^-(L_1,L_2,L_3) = n^-(t) = 2 (\eta_{\mathrm{HK}}(L_2,L_1) + \eta_{\mathrm{HK}}(L_3,L_2) + \eta_{\mathrm{HK}}(L_1,L_3)). \qedhere
Corollary

Corollary 8.21

3XMI

The indices \eta_{\mathrm{HK}} and \eta_{\mathrm{Fuk}} are cohomologous, i.e., there exists for every L a number \tau(L) \in \mathbb Z such that \eta_{\mathrm{HK}}(L_1,L_2) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \tau(L_1) - \tau(L_2).

Proof.

By the first part of the proposition we have

\eta_{\mathrm{HK}}(L_1,L_2) + \eta_{\mathrm{HK}}(L_2,L_3) + \eta_{\mathrm{HK}}(L_3,L_1) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \eta_{\mathrm{Fuk}}(L_2,L_3) + \eta_{\mathrm{Fuk}}(L_3,L_1).

Apply this to L_3 = L_{\mathrm{pt}} to get

\eta_{\mathrm{HK}}(L_1,L_2) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \eta_{\mathrm{Fuk}}(L_2,L_{\mathrm{pt}}) + \eta_{\mathrm{Fuk}}(L_{\mathrm{pt}},L_1) - \eta_{\mathrm{HK}}(L_2,L_{\mathrm{pt}}) - \eta_{\mathrm{HK}}(L_{\mathrm{pt}},L_1).

By the second part of the proposition we obtain

\eta_{\mathrm{HK}}(L_1,L_2) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \eta_{\mathrm{Fuk}}(L_2,L_{\mathrm{pt}})-\eta_{\mathrm{HK}}(L_2,L_{\mathrm{pt}}) -\eta_{\mathrm{Fuk}}(L_1,L_{\mathrm{pt}}) + \eta_{\mathrm{HK}}(L_1,L_{\mathrm{pt}})

which has the desired form. In fact, \eta_{\mathrm{Fuk}}(L,L_{\mathrm{pt}}) = 0 so it suffices to let \tau(L) = \eta_{\mathrm{HK}}(L,L_{\mathrm{pt}}).

Proof

Proof

YYWY

By the first part of the proposition we have

\eta_{\mathrm{HK}}(L_1,L_2) + \eta_{\mathrm{HK}}(L_2,L_3) + \eta_{\mathrm{HK}}(L_3,L_1) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \eta_{\mathrm{Fuk}}(L_2,L_3) + \eta_{\mathrm{Fuk}}(L_3,L_1).

Apply this to L_3 = L_{\mathrm{pt}} to get

\eta_{\mathrm{HK}}(L_1,L_2) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \eta_{\mathrm{Fuk}}(L_2,L_{\mathrm{pt}}) + \eta_{\mathrm{Fuk}}(L_{\mathrm{pt}},L_1) - \eta_{\mathrm{HK}}(L_2,L_{\mathrm{pt}}) - \eta_{\mathrm{HK}}(L_{\mathrm{pt}},L_1).

By the second part of the proposition we obtain

\eta_{\mathrm{HK}}(L_1,L_2) = \eta_{\mathrm{Fuk}}(L_1,L_2) + \eta_{\mathrm{Fuk}}(L_2,L_{\mathrm{pt}})-\eta_{\mathrm{HK}}(L_2,L_{\mathrm{pt}}) -\eta_{\mathrm{Fuk}}(L_1,L_{\mathrm{pt}}) + \eta_{\mathrm{HK}}(L_1,L_{\mathrm{pt}})

which has the desired form. In fact, \eta_{\mathrm{Fuk}}(L,L_{\mathrm{pt}}) = 0 so it suffices to let \tau(L) = \eta_{\mathrm{HK}}(L,L_{\mathrm{pt}}).

Source and licence

The text on this page is from arXiv:2610.21960v1 by Patrick Antweiler, Vicente Cortés, Julian Holstein, Markus Röser and Aldo Witte, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2610/2610.21960.jsonl.