GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES
PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Abstract. We introduce graph potentials, which are Laurent polynomials associated to (colored) trivalent graphs. These graphs encode degenerations of curves to rational curves, and graph potentials encode degenerations of the moduli space of rank 2 bundles with xed determinant. We show that the birational type of the graph potential only depends on the homotopy type of the colored graph, and thus dene a topological quantum eld theory. By analyzing toric degenerations of the moduli spaces we explain how graph potentials are related to these moduli spaces in the setting of mirror symmetry for Fano varieties. On the level of enumerative mirror symmetry this shows how invariants of graph potentials are related to Gromov–Witten invariants of the moduli space. In the con- text of homological mirror symmetry we formulate a conjecture regarding the shape of semiorthogonal decompositions for the derived category. Studying the properties of graph potentials we provide evidence for this conjecture. Finally, by studying the Grothendieck rings of varieties and categories we will give further geometric evidence.
1. Introduction 2
1.1. Graph potentials 3
1.2. Mirror symmetry 5
1.3. Contents of this paper 9
2. Graph potentials 14
2.1. Denition 15
2.2. Elementary transformations 19
2.3. Periods of Laurent polynomials 23
2.4. Critical values 25
3. Topological quantum eld theories from graph potentials 27
3.1. Restricted TQFT 27
3.2. The graph potential TQFT 28
3.3. Computing periods via the graph TQFT 33
4. Toric degenerations 41 arXiv:2009.05568v2 [math.AG] 27 Jul 2021
4.1. Manon’s degeneration 41
4.2. Analyzing the singularities 45
5. Enumerative mirror symmetry and symplectic geometry 50
5.1. Integrable systems 51
5.2. Discs of Maslov index two and Floer potentials 52
Keywords. trivalent graphs, moduli spaces of vector bundles, character varieties, topological quantum eld theories, mirror symmetry, integrable systems, conformal blocks, toric degenerations, semiorthogonal decompositions, quantum periods, cluster algebras, mutations, random walks, Bessel functions, Donaldson–Floer theory, Landau–Ginzburg models, Grothendieck rings of varieties, monotone Lagrangian tori, Witten–Dijkgraaf– Verlinde–Verlinde equation . 1 2 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
5.3. Monotone Lagrangian tori 54
5.4. Descendant invariants 57
5.5. Quantum periods 58
5.6. On mutations and cluster charts 59
6. Semiorthogonal decompositions 60
6.1. The fundamental conjecture 60
6.2. Lefschetz decompositions 61
6.3. Ringel–Samokhin-type decompositions 63
6.4. Decompositions from the point of view of mirror symmetry 64
7. Decomposition in the Grothendieck ring of varieties 67
7.1. The Thaddeus picture 68
7.2. Proof of the decomposition 71
7.3. Motivic zeta functions and a Harder-type formula 73
Appendix A. Some explicit calculations for periods 75
A.1. Second quantum period 77
A.2. Fourth period 79
A.3. General patterns 80
Appendix B. Random walks on lattices 81
References 84
1. Introduction We discuss aspects of mirror symmetry for moduli spaces of rank 2 vector bundles with xed determinant of odd degree on algebraic curves of genus 푔 ≥ 2. These are smooth projective Fano varieties of dimension 3푔 − 3 in the algebro-geometric context, and mono- tone symplectic manifolds of dimension 6푔 − 6 in the symplecto-geometric context. In this paper we propose a mirror construction for these, based on trivalent graphs, leading to graph potentials. These varieties also have manifestations as moduli spaces of at unitary SU(2)-connections on Riemann surfaces, or character varieties of conjugacy classes of representations of the fundamental group. Those alternative viewpoints allow us to import ideas, tools and results from di erent areas of mathematics, and in this introduction we will try to highlight some of these.
The main new object introduced in this paper are graph potentials. In Section 1.1 we give a general overview of them, without reference to (mirror symmetry of) moduli spaces of vector bundles. By doing so we want to invite the interested reader to nd their own variations on the construction.
In Section 1.2 we discuss the aspects of mirror symmetry relevant to understand the results in this paper, and possible future extensions. We will take some time to discuss aspects of mirror symmetry for Fano varieties, and highlight the interesting role that moduli spaces of vector bundles on curves play in this story. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 3
In Section 1.3 we will describe the contents of this paper, explaining the main results we have proved. We will also speculate on some connection between Donaldson–Floer theory, Atiyah–Floer conjecture, and our work.
1.1. Graph potentials. We will now introduce graph potentials in an abstract way, without reference to mirror symmetry or geometry. For an exposition with the applications
in mind, see Section 1.3 and Section 2 , where further details omitted in this introduction are discussed. Graph potentials. Let Y(푎,푏, 푐) be a function (dened on some space equipped with a volume form) which is symmetric in its arguments 푎,푏, 푐, such that (1) Y(푎,푏, 푠) + Y(푐,푑, 푠) = Y(푎, 푐, 푡) + Y(푏,푑, 푡) = Y(푎,푑,푢) + Y(푏, 푐,푢) for some volume-preserving transformation 휑 relating the variables. Let 푣 be a trivalent vertex, whose edges are labelled by 푎,푏, 푐, as in 푎 (2) 푣 푏 푐
We dene the vertex potential 푊푣 as Y(푎,푏, 푐). Next, let 훾 = (푉, 퐸) be a trivalent graph, possibly with half-edges (or leaves). The graph potential 푊훾 is the sum of the vertex potentials associated to the (internal) vertices, i.e. ∑︁ (3) 푊훾 := 푊푣 푣∈푉 where푊푣 = 푊푣 (푎푣,푏푣, 푐푣). This is an expression in 3(푔−푏0)+2푛 arguments, where푏0 = 푏0 (훾) is the number of connected components, 푔 = 푏1 (훾) is the genus of the graph, and 푛 is the number of half-edges. There are 3(푔 − 푏0) + 푛 internal variables, and 푛 external (or leaf) variables associated to the half-edges.
The functional equation ( 1 ) explains how the graph potentials behave under mutation. Such a mutation will produce a new trivalent graph with a new graph potential associated to it, and controlling this operation will yield important invariance properties of graph potentials. The upshot is that the di erent 푊훾 are essentially a single function 푊푔,푛 expressed in di erent coordinates. Partition functions from graph potentials. Applying a Fourier-like transform we can write ∑︁ (4) exp(Y(푎,푏, 푐)) = 퐹 (푖, 푗, 푘)푎푖푏 푗푐푘 푖,푗,푘 for a function 퐹 (푖, 푗, 푘). These functions satisfy the Frobenius equation ∑︁ ∑︁ (5) 퐹 (푖, 푗,푚)퐹 (푘, 푙, −푚) = 퐹 (푖, 푘, 푛)퐹 (푗, 푙, −푛). 푚 푛
This equation hints at an important compatibility property, culminating in Theorems A
and B .
We can moreover consider the integration over the internal variables 푤1, . . . ,푤3푔−3+푛, and dene the symmetric function ∭ (6) 푍푔 (푧1, . . . , 푧푛) := exp(푊훾 (푧1, . . . , 푧푛;푤1, . . . ,푤3푔−3+푛) 4 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY in the 푛 external variables 푧1, . . . , 푧푛. This function only depends on the genus 푔 of the graph 훾, and it satises the rules ∭ 푍푔+푔0 (푧1, . . . , 푧푛+푛0 ) = 푍푔 (푧1, . . . , 푧푛,푤)푍푔0 (푧푛+1, . . . , 푧푛+푛0,푤) (7) ∭ 푍 (푧 , . . . , 푧 ) = 푍 (푧 , . . . , 푧 ,푤,푤). 푔+1 1 푛 푔 1 푛
This type of compatibilities, ultimately governed by ( 1 ), allow us to study the partition functions associated to graph potentials. Our choice of potential. In the remainder of the paper we will focus on a specic choice of the function Y, namely the Laurent polynomial 푎 푏 푐 (푎,푏, 푐) = + + + 푝 · 푎푏푐 (8) Y 푏푐 푎푐 푎푏 for some value 푝 (which will be 1 for us). The domain of the function is the torus (C×)3.
The transformation 휑 that gives a solution to ( 1 ) is the rational change of coordinates (9) 푠 ·(푎푏 + 푐푑) = 푡 ·(푎푐 + 푏푑) = 푢 ·(푎푑 + 푏푐). This transformation preserves the logarithmic rational volume forms
(10) 휔푎 ∧ 휔푏 ∧ 휔푐 ∧ 휔∗ ∗ ∈ {푠, 푡,푢} 휔 푎 d푎 where 푎 = d log = 푎 is the Haar measure for the multiplicative group. We can give an interpretation of this choice, independent of the role it will play for us, as , 퐷 3 follows. Consider a tetrahedron inscribed in the cube [0 2 ] , whose facets are bounded by the quantum Clebsch–Gordan conditions 퐴 ≤ 퐵 + 퐶, 퐵 ≤ 퐴 + 퐶, 퐶 ≤ 퐴 + 퐵 (11) 퐴 + 퐵 + 퐶 ≤ 퐷 (12) ±퐴 ± 퐵 ± 퐶 ∈ Z This tetrahedron can be seen as the moduli space of geodesic triangles on a sphere of 푅 퐷 radius = 2휋 . The rst inequalities are the classical triangle inequalities, and the last one is a quantum bound of the perimeter. After the substitution (푎,푏, 푐,푑) = (exp 퐴, exp 퐵, exp퐶, exp(−퐷)) the Laurent polynomial
( 8 ) becomes a generating√ function of the inequalities ( 11 ), which is invariant with respect to translations by 2휋 −1(퐴, 퐵,퐶) for (퐴, 퐵,퐶) satisfying ( 12 ). If we let 퐴, 퐵,퐶 be real (resp. complex) numbers, then 푎,푏, 푐, 푝 are real positive (resp. complex non-zero), and the Haar measure 휔푎 in the coordinate 푎 becomes the Lebesgue measure in the coordinate 퐴. It would be interesting to nd other functions Y(푎,푏, 푐), study their partition functions, and relate them to geometry.
Lattice polytopes of graph potentials. As a preview to Section 1.3 we will now discuss how graph potentials (in the specic setting introduced before) are related to other objects. 푊 Í 푐 푣 푧푣 푑 For this we recall that the Newton polytope of a Laurent polynomial = 푣∈Z푑 ( ) ∈ Z[Z ] is the polytope Newt(푊 ) ⊂ R푑 dened as the convex hull of the indices 푣 ∈ Z푑 ⊂ R푑 for which 푐(푣) ≠ 0. The polar dual of a polytope 푃 ⊂ R푑 is dened as n o (13) 푃◦ := 푢 ∈ R푑 | ∀푚 ∈ 푃 : h푢,푚i ≥ −1 . This polar duality is involutive on the class of full-dimensional convex closed polytopes having the origin in their interior. It was recognised by Batyrev as a manifestation of
mirror symmetry in combinatorial geometry [ 14 , 50 ]. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 5
In the case of graph potentials, the polar dual Δ훾 of the polytope Newt(푊훾 ) is given by
the inequalities ( 11 ), whilst the congruences from ( 12 ) specify an integral lattice in the underlying space. These polytopes, and their integer points, turn out to be of central interest in mathematics, and already appeared in various contexts. Let us name a few:
• The polytopes Δ훾 and their projections to 푛 leaf variables dene the multiplicitive
or quantum Horn equalities in the works of Agnihotri–Woodward and Belkale [ 4 ,
24 ]. • They are also isomorphic to the images of the Goldmann–Je rey–Weitsman in-
tegrable systems from [ 97 , 96 ] and the Newton–Okounkov bodies arising from
conformal blocks in the works of Manon [ 64 , 128 , 129 ]. This latter point will be instrumental for this paper.
• The number of integer points in the dilations 푛Δ훾 for 푛 ≥ 0 gives the Ehrhart polynomial. This polynomial counts Bohr–Sommerfeld Lagrangian tori on the
moduli space of at SU(2)-connections in the symplectic setup [ 129 ], whilst it coincides with the Hilbert polynomial of the moduli space of rank 2 bundles in
the algebro-geometric setup [ 33 ]. They can be computed by the Verlinde formula,
and the location of their roots was studied by the authors in [ 28 ].
1.2. Mirror symmetry. Having described graph potentials, we can now introduce the second protagonist of this paper: moduli spaces of rank 2 bundles on smooth projective
curves. They have a long history, having been studied since the 1960s [ 139 , 142 ]. To dene them, x a line bundle L of odd degree on a smooth projective curve 퐶 of genus 푔 ≥ 2. Then M퐶 (2, L) is the moduli space of stable rank 2 bundles on 퐶, with xed determinant L. Occassionally we will also talk about M퐶 (2, L), but the results of this paper do not directly apply to this variety. For now it su ces to recall that M퐶 (2, L) (resp. M퐶 (2, O퐶 )) is a smooth (resp. singular) projective Fano variety of dimension 3푔 − 3, whose Picard group is rank 1, and that they are of index 2 (resp. 4).
By results of Narasimhan–Seshadri [ 140 ], Atiyah–Bott [ 6 ] and Donaldson [ 59 ] these moduli spaces have important reinterpretations:
• The moduli space of stable vector bundles M퐶 (2, L) (resp. M퐶 (2, O퐶 )) is also known as the Dolbeault moduli space, and by the Narasimhan–Seshadri corre- spondence the underlying topological space (with the Euclidean topology) is homeomorphic to the Betti moduli space, which is the twisted (resp. untwisted) character variety associated to representations of the fundamental group of the underlying oriented surface. • The Betti moduli space is isomorphic as a real analytic space to the de Rham moduli space, which is the moduli space of twisted (resp. untwisted) at unitary SU(2)- connections. The study of mirror symmetry for character varieties was started by Fock–Goncharov
[ 68 ], but it concerns di erent aspects than we study in this paper. A running example. The rst case to consider is for 푔 = 2. In this case an explicit description of M퐶 (2, L) due Newstead and Narasimhan–Ramanan exists [ 141 , 139 ]: it can be written as 5 (14) M퐶 (2, L) 푄1 ∩ 푄2 ⊂ P where 푄1, 푄2 are smooth quadric hypersurfaces determined by 퐶. This description as a complete intersection only exists for 푔 = 2, which makes them more amenable to the existing techniques, as we will discuss. 6 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
These varieties degenerate to a toric Fano threefold with six ordinary double points. For the (smooth) intersection of quadrics in P5 we have the toric degeneration given by the (singular) intersection of quadrics with equations
(15) 푍0푍1 = 푍2푍3 = 푍4푍5. As a toric variety it is described by the polytope which is the convex hull of the ver- tices (±1, ±1, ±1). This toric variety has terminal singularities, and its 6 ordinary double points admit a small resolution of singularities. The graph potential in this case, which is closely related to the polytope dening the toric degeneration, takes on the form 푥 푦 푧 1 푦푧 푥푧 푥푦 푊 = 푥푦푧 + + + + + + + . (16) e 푦푧 푥푧 푥푦 푥푦푧 푥 푦 푧 It is associated to the (colored) trivalent graph of genus 푔 = 2 푥 푦 푧 (17) 1 2 We will illustrate what is (not) known about specic aspects of mirror symmetry using this explicit example. Mirror symmetry. Under mirror symmetry the graph potentials are related to moduli spaces that parametrize isomorphism classes of holomorphic vector bundles on curves. In this paper we will consider various manifestations of mirror symmetry, which will be discussed in the next paragraphs.
The mirror dual of M퐶 (2, L) is expected to be a cluster-like variety equipped with a regular function, the so-called Landau–Ginzburg potential (closely related to the Floer potential), that will play the central rôle in this story. The graph potentials form a class of possible hierarchies of such functions adapted to the “collective” study of mirror symmetry for the moduli spaces of local systems on surfaces of all genera and all numbers of boundary components simultaneously, yet still keeping the rank xed. There are a couple of interdependent theoretical approaches to mirror symmetry for Fano varieties. These provide di erent methods, levels of rigour, constructivity, precision, predictive power, scope of interests and applications. But in what concerns experimental data, so far in any approach mirror symmetry was mostly studied either in low dimensions, or for (Schubert cycles in) homogeneous varieties,or for toric varieties, or for their products
(and some brations), or for complete intersections therein, see e.g. [ 9 , 123 , 155 , 156 , 157 ]. On the other hand, moduli spaces of rank 2 bundles are notorious for their falsications of naive conjectural extrapolations based on the data obtained from the examples above. For example, in order to apply techniques such as the abelian/non-abelian correspondence to compute quantum periods (and Givental 퐽-functions), the classication of Fano 3-folds
was reworked in terms of GIT data in [ 48 , Theorem A]. This theorem, which is proved in the rst 105 sections of the paper, says that all bers in a family of Fano threefolds (and del Pezzo surfaces) can be obtained as zero loci of regular sections of a xed vector bundle on a xed “key variety”, which is isomorphic to a product of a toric variety with a Grassmannian. As a by-product this theorem implies that moduli spaces of complex structures on such Fano manifolds are unirational.
In contrast, the (last) Section 106 of op. cit. considers M퐶 (2, L): by the non-abelian Torelli theorem (see e.g. [ 110 , Theorem E]) the moduli space of complex structures on M퐶 (2, L) has a dominant rational map to the Deligne–Mumford moduli space of stable curves. For
푔 ≥ 23 these have Kodaira dimension at least 0 by [ 89 ]. This indicates that the mirror GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 7 map to the moduli space of complex structures on the Fano manifold M퐶 (2, L) from the complexied Kähler cone establishing mirror symmetry has to be a non-trivial map, very di erent from all the known examples of mirror maps for Fano manifolds. Enumerative mirror symmetry. The rst manifestation of mirror symmetry we will focus on is a relationship between • certain algebro-geometric invariants that are polynomial expressions in the coe - cients of the Laurent polynomial 푊 (and that can be written as the integrals in the Landau–Ginzburg B-model) • and symplectic invariants of 푀, namely disc and sphere correlators with one point insertion and gravitational descendants, for the A-model.
In particular we consider the quantum period G푀 (푡), a generating function for closed genus zero Gromov–Witten invariants with descendants and primary eld being a point, which can be also dened using the operator of quantum multiplication by the rst Chern class, whose coe cients are closed genus zero correlators without descendants but with
two arbitrary cohomology insertions. Its precise denition can be found in Section 5 . There we also dene the period of a Laurent polynomial 푊 .
The state-of-the-art on the tools to determine quantum periods can be found in [ 48 , 49 ]. In particular the case 푔 = 2 being a complete intersection makes it amenable to the existing
techniques [ 48 , §6], and its quantum period is +∞ ∑︁ (2푑)!(2푑)! (18) G (푡) = 푡 2푑 . 푀 (푑!)6 푑=0
On the other hand it is straightforward to compute the rst periods of the graph poten-
tial ( 16 ) associated to a graph 훾 of genus 2. One observes that they agree with the rst coe cients of the regularized quantum period. This numerical check hints at enumer- ative mirror symmetry between M퐶 (2, L) and the graph potential, relating two power series with integer coe cients. We will upgrade it to an equality of power series, for all 푔 ≥ 2, subject to a conjecture on the singularities of a toric degeneration which plays
an important role in the construction. After the statement of Theorem D we will discuss some alternative approaches to determining the quantum period of M퐶 (2, L), and their respective downsides. The degenerate toric variety used to prove the equality is “small” in two di erent senses: (1) It has terminal Gorenstein singularities, so it is small in the sense of Batyrev’s proposal, and appears in the classication of small toric degenerations of Fano
threefolds by [ 73 ], which was done with a view towards applications to mirror symmetry. (2) It admits a small resolution of singularities (i.e. one without exceptional divisors).
For this reason it was used by [ 144 ] to construct an integrable system and to compute Floer potential of the Lagrangian torus ber, and by Nishinou to compute
some other Gromov–Witten invariants via tropical curve counting [ 143 ]. These two properties are important in this approach to mirror symmetry, and will be
discussed at length in Section 4 .
Homological mirror symmetry and semiorthogonal decompositions. Since M퐶 (2, L) is a Fano manifold, its bounded derived category of coherent sheaves is expected to have a homological mirror dual partner in the sense of Kontsevich’s homological mirror symmetry
conjecture, cfr. [ 107 ] for Calabi–Yau varieties and [ 108 , page 30] for the Fano version. 8 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
This mirror dual partner is expected to be a pair (푌, 푓 ), where 푌 is a smooth quasiprojective variety and 푓 a regular function on it. The Laurent polynomials discussed before are then 푓 푛 푌 restrictions of the Landau–Ginzburg potential to a torus Gm ⊆ . For more information
on these pairs, and their “tamings” one is referred to [ 103 ]. The triangulated category of algebro-geometric origin associated to (푌, 푓 ) is the category
of matrix factorizations, see e.g. [ 151 ]. In an algebraic context these were studied in the
1980s by Eisenbud [ 61 ] and Buchweitz [ 43 ], whilst their role in mirror symmetry was
suggested in the physics literature by Kapustin–Li [ 101 ]. Assuming that the categories of symplectic origin, namely the so-called Fukaya category of (graded) immersed (e.g. embedded) coisotropic (e.g. Lagrangian) subvarieties in M퐶 (2, L) decorated with a at unitary connection and the analogous Fukaya–Seidel category of vanishing Lagrangian thimbles in 푌 , are well-dened, the spaces M퐶 (2, L) and (푌, 푓 ) are called homologically mirror dual if
• the derived category of coherent sheaves on M퐶 (2, L) is equivalent to the derived Fukaya–Seidel category of (푌, 푓 ),
• the derived Fukaya category of M퐶 (2, L) is equivalent to the matrix factorization category of (푌, 푓 ).
We also refer the reader to [ 102 , Conjecture 2.3] for a general discussion on Landau– Ginzburg models and HMS conjectures. An important aspect of studying properties of these triangulated categories is to under- stand their (semi)orthogonal decompositions. An overview of the state-of-the-art for derived categories of smooth projective varieties can be found in Kuznetsov’s ICM address
[ 118 ]. The running example of 푔 = 2 was in fact the rst non-trivial semiorthogonal
decomposition studied by Bondal and Orlov [ 41 ], and yields b b b (19) D (M퐶 (2, L)) = D (푄1 ∩ 푄2) = O, O(1), D (퐶) . Understanding semiorthogonal decompositions in the case of 푔 ≥ 3 is an important challenge, and is the subject of various conjectures that we will discuss. Even for 푔 = 2 the full homological mirror symmetry conjecture is not known to hold
(but see the work of [ 166 ] for partial results regarding the Fukaya category). Yet graph potentials and their enumerative properties are expected to give rise to cluster charts of the Landau–Ginzburg mirror (푌, 푓 ), and one can study (semi)orthogonal decompositions of the categories on di erent sides of the mirror already using the graph potentials without understanding the whole Landau–Ginzburg model. For the running example of 푔 = 2 this program e.g. consists of studying the singularities
of the potential ( 16 ), which has critical values 0 and ±8, with Morse points over ±8 and a 1 one-dimensional singular locus over 0. These three critical values correspond to the three
components in the Bondal–Orlov decomposition ( 19 ) of the bounded derived category of coherent sheaves on the intersection of two quadrics in P5. Quantization following Tyurin. An important source of inspiration is the work of
Tyurin on quantization and eld theories for moduli spaces of vector bundles [ 172 ]. Greatly
extending the survey of non-abelian theta functions by Beauville [ 20 ], Tyurin discusses some of the key players in both the works we build on and this paper. These include the interplay between algebraic and symplectic geometry of moduli of bundles, integrable systems, low-dimensional topology, trivalent graphs, polytopes and toric varieties, and how the geometry of degenerations can be used with great success. In this work we apply
1The so-called “banana singularity”, that consists of three rational irreducible components intersecting in two points. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 9 these ideas to study mirror symmetry aspects of the moduli spaces whose geometry forms the subject of Tyurin’s book.
1.3. Contents of this paper. Having sketched the construction of graph potentials, and discussed the aspects of mirror symmetry relevant to this paper, we can now discuss its structure and state the main theorems. In this paper we will (1) dene graph potentials and discuss their properties under elementary transforma-
tions, see Section 2 ; (2) describe an associated topological quantum eld theory with values in the Hilbert 2 space ℓ (Z), encoding the oscillating integrals of graph potentials, see Section 3 ;
(3) relate graph potentials to Manon’s construction of toric degenerations for M퐶 (2, L),
and discuss the singularities arising from them, see Section 4 ;
(4) associate a monotone Lagrangian torus 퐿훾 on M퐶 (2, L) to every trivalent graph 훾, so that (under some assumptions) the graph potential 푊e훾 is identied with the Lagrangian Floer potential of 퐿훾 , and the oscillating integrals of the potentials are
identied with the quantum periods of M퐶 (2, L), see Section 5 ;
b (5) conjecture the shape of semiorthogonal decompositions of D (M퐶 (2, L)), dis- cussing the relationship between the spectrum of the (small) quantum cohomology
algebra of M퐶 (2, L) and the critical loci of the potentials, see Section 6 ; (6) give further evidence for this conjecture in terms of an identity in the Grothendieck
rings of varieties and categories, see Section 7 . In the appendices we will
(A) describe the Fano variety of lines in M퐶 (2, L) and the lines passing through a point, and discuss some patterns in the period sequences associated to M퐶 (2, L),
see Appendix A ; (B) show how graph potentials relate to the theory of random walks, and construct random walks with distinct shapes but equal return probabilities for every number
of steps, see Appendix B . Let us now discuss the topics of this paper in more detail, indicating the three main themes of the paper. Graph potentials and topological quantum eld theories. The main new object in this paper are graph potentials, which as discussed before are a class of Laurent polynomials with remarkable symmetry properties, associated to (colored) trivalent graphs. Their
precise denition can be found in Section 2 . A trivalent graph of genus 푔 arises as a degeneration of a smooth projective curve of genus 푔 into a nodal curve with only rational components. Such a curve is also called a
graph curve in Bayer–Eisenbud [ 19 ], and its dual graph gives 훾. This link to algebraic geometry will be very important once we start considering toric degenerations of moduli spaces. Alternatively, such trivalent graphs can be seen as pair of pants decompositions of 퐶, considered as a Riemann surface. The main property that graph potentials satisfy is that they are invariant under elementary transformations à la Hatcher–Thurston, relating di erent pair of pants decompositions of the Riemann surface. The following theorem shows how graph potentials behave under these transformations. 10 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Theorem A. Let (훾, 푐) and (훾 0, 푐 0) be two trivalent colored graphs related by elementary transformations. Then the graph potentials 푊e훾,푐 and 푊e훾0,푐0 are identied via a rational change of coordinates.
Hence by tropicalization of the rational change of coordinates in Theorem A , we obtain a piecewise-linear automorphism of R3푔−3 that maps the polar dual of the Newton polytope of 푊e훾,푐 to the polar dual of the Newton polytope of 푊e훾0,푐0 . From this combinatorial point of
view Theorem A can also be realized as a geometric lift of a piecewise-linear homeomor-
phism in the sense of Berenstein–Zelevinsky [ 32 ]. We also refer the reader to Section 5.6 for a related discussion on mutations and cluster charts. Inspired by this invariance under elementary transformations, we will dene in Construc-
tion 3.2.8 a topological quantum eld theory, and call it the graph potential eld theory. In other words, we have found a new solution to the Witten–Dijkgraaf–Verlinde–Verlinde (WDVV) equation. Its partition function Zgp (푡) (for any 푡 ∈ R) assigns to the pair of pants Σ0,3 with 1 +휖 anticlockwise oriented circles, for 휖 ∈ {0, 1}, the appropriately colored vertex potential, i.e.
gp 2 ⊗3 (20) Z (푡)(Σ0,3) := exp 푡푊e푣,휖 (푥,푦, 푧) ∈ ℓ (Z) .
Here ℓ2 (Z) is the Hilbert space of square-integrable real series indexed by Z. A countable 푥푖 ℓ2 Í 푎 푥푖 orthonormal basis is given by { }푖 ∈Z and any element of (Z) can be written as 푖 ∈Z 푖 Í 푎2 such that 푖 ∈Z 푖 < ∞. Because the potential takes values in an innite-dimensional Hilbert space, we will need to restrict the bordism category suitably. The properties of the graph potential can then be used to show the following theorem.
Theorem B. Let 푡 ∈ R. The graph potential eld theory is a two-dimensional topological quantum eld theory on a suitably restricted bordism category, with values in Hilbert spaces.
Observe that our TQFT takes values in ℓ2 (Z), and that it is not the usual (1 + 1)-TQFT coming from the fusion ring of conformal blocks. The existence of this TQFT shows why it is a powerful idea to consider the moduli spaces of vector bundles on curves for all genera simultaneously when one is interested in mirror symmetry aspects of these Fano varieties. In a di erent direction, González-Prieto–Logares– Muñoz have constructed a topological quantum eld theory computing Hodge–Deligne
polynomials for representation varieties [ 84 ]. More generally, two-dimensional TQFT’s and the WDVV equation have been crucial in the development of Gromov–Witten theory and the solution of long-standing problems in classical enumerative mirror symmetry, see
e.g. Kontsevich-Manin [ 109 ]. In fact, eld theories form a common theme in this paper, and the works we build on. As
we just discussed, in Section 3 we have introduced a new topological quantum eld theory
to compute periods of graph potentials. The results we build upon in Section 4 use the
conformal eld theory given by conformal blocks [ 23 , 173 , 171 ]. Section 5 features both
cohomological eld theories and symplectic eld theories (see [ 109 ] and [ 1 , 42 , 62 ]). The cohomological and symplectic eld theories in turn can be compared using a tropical eld theory, which can be seen as an intermediary between them. Using holomorphic-to- tropical and Lagrangian-to-tropical correspondences one can establish mirror symmetry
[ 131 , 132 , 130 , 126 , 127 ]. Finally, at the end of this introduction we speculate on the relationship between our work and the construction of the (still conjectural) 4-dimensional Donaldson–Floer quantum eld theory. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 11
Enumerative mirror symmetry. As discussed before, an important tool in mirror sym-
metry is the notion of a toric degeneration, see [ 93 , 17 , 81 , 15 , 36 ]. For moduli spaces of (quasi-parabolic) bundles of rank 2 on a curve, toric degenerations have been introduced
by Manon using the theory of conformal blocks [ 129 , 128 ]. Behind this construction is the
work of Faltings, Kumar–Narasimhan–Ramanathan, Beauville–Laszlo and Pauly [ 63 , 21 ,
114 , 154 ], who show that the bundles of conformal blocks restricted to the smooth locus of the moduli space of curves M푔,푛 are isomorphic to the bundles of non-abelian 휃-functions for SL2, i.e. holomorphic sections of line bundles on moduli spaces of (parabolic) bundles of rank 푟. These toric degenerations are described by Manon using trivalent graphs, associated to the most degenerate elements in the stratication of the boundary divisor in the Deligne– Mumford compactication (albeit with a di erent type of decoration than a coloring). Their construction and the relationship to the toric varieties associated to graph potentials is
discussed in Section 4.1 . To the best of our knowledge, the bers over the boundary divisor have not yet found any alternative interpretation as moduli spaces of objects associated to the degenerate curves. Certain geometric properties of the toric degeneration depend on the choice of trivalent
graph, unlike the discussion in [ 128 ] which is concerned mostly about properties invariant under this choice. In particular, the singularities of the toric variety can di er signicantly.
We show in Theorem 4.2.3 that the singularities are terminal if and only if the trivalent graph does not have any separating edges, i.e. one needs to remove at least 2 edges to
make it disconnected. We apply Theorem 4.2.3 and recover an earlier result of Kiem–Li
[ 105 ] about terminality of M퐶 (2, O퐶 ) (for a generic curve 퐶). To apply results of Nishinou–Nohara–Ueda we need the toric variety to have a small toric resolution of singularities, i.e. a subdivision of the fan into basic cones based on original rays. We conjecture the following criterion
Conjecture C. Let 훾 be a connected trivalent graph and 푐 its coloring of odd degree. The toric variety with moment polytope Δ훾,푐 has a small resolution of singularities if and only if the graph 훾 is maximally connected, i.e. one needs to remove at least 3 edges to make it disconnected.
As discussed in Remark 4.2.10 , we have been able to check this conjecture for 푔 = 2, 3, 4. We now come to the third main theorem of this paper
Theorem D. Let 퐶 be a smooth projective curve of genus 푔 ≥ 2. Let (훾, 푐) be a colored trivalent graph of genus 푔, with an odd number of colored vertices such that the cor- responding toric variety has a small resolution of singularities. Then mirror symmetry holds between the A-model of the Fano manifold M퐶 (2, L) and the B-model of the graph potential 푊e훾,푐 in the following senses:
(1) There is a monotone Lagrangian torus 푇훾 on M퐶 (2, L) such that its Floer potential is equal to 푊e훾,푐 ,
(2) The quantum period of the Fano manifold M퐶 (2, L) is equal to the integral of the exponent of a graph potential over the unit torus ∫ 푡 ∑︁ 휓푑−2 푝푡 푡푑 휔 푡 푊 푧 (21) GM퐶 (2,L) ( ) := 1 + h [ ]i = 푧 · exp( · 훾,푐 ( )) | | 푑 푧 =1 In particular, (the main irreducible component of) the Laplace transform of the quantum connection has geometric origin, i.e. it is isomorphic to a submodule of the Picard–Fuchs D-module for a pencil of algebraic varieties. 12 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
By Conjecture C , it is expected that we can always nd many such trivalent graphs.
Theorem D is unconditional for 푔 = 2, 3, 4.
Using the computational tools provided by Theorem B it is possible to compute the period sequences e ciently, up to a certain cuto , for all 푔 at the same time. Moreover using these, in principle one can e ciently compute the Picard–Fuchs operators for 푊e훾,푐 (푡) or the quantum di erential equation of the Fano variety M퐶 (2, L) up to a su ciently high genus.
In Tables 1 and 2 we have collected a small portion of these computations, and one can
notice certain patterns. These are discussed in Appendix A . In particular we will give in
Appendix A.1 a geometric proof, unconditional on the existence of a small resolution, for the vanishing of the second quantum period for 푔 ≥ 3.
It is also possible to interpret results on periods in terms of probability theory and random walks, highlighting another interesting property that graph potentials exhibit: in contrast to all other known examples, all graph potentials with 푛 = 0 are balanced, that is 1 is their critical point and the value 푊 (1) equals 8푔(훾) − 8 for any trivalent graph 훾. We do this in
Appendix B .
Nohara–Ueda [ 148 , 149 ] construct potential functions for moduli spaces of Euclidean polygons using equivariant toric degenerations of Grassmannians of planes. We discuss
(in the last part of Section 3 and Remark 3.3.19 ) how the Nohara–Ueda potential can be recovered as the limit of our graph potentials for moduli spaces of spherical polygons. We
refer the reader to [ 149 , Page 164], and to Denition 4.1.5 and ( 145 ) in Section 4 for the comparisons between the equations dening the relevant polytopes.
There are alternative methods to compute quantum periods of M퐶 (2, L). A rst one is using the Desale–Ramanan description of M퐶 (2, L) when 퐶 is hyperelliptic [ 57 ], which implies that the symplectic variety M퐶 (2, L) is symplectomorphic to the zero locus of a section of the homogeneous vector bundle (Sym2 U∨) ⊕2 on Gr(푔 − 1, 2푔 + 2). Hence
the abelian/non-abelian correspondence from [ 46 ] can be applied to it. But this method
is only feasible for low values of 푔. A second method is the construction from [ 94 , 106 ], again using the Desale–Ramanan description. One rst produces a mirror dual gauged linear sigma model of higher dimension and then abelianizes and reduces the number of variables. But we have only been able to make this approach work for 푔 = 2, 3.
The benet of the approach in this paper is that the construction is very explicit and direct, and leads to the graph potential TQFT. This allows us to compute quantum periods simultaneously for all 푔 using a xed amount of memory. Semiorthogonal decompositions and homological mirror symmetry. As discussed in Section 1.2 , the derived categories of the varieties M퐶 (2, L) are expected to have
interesting semiorthogonal decompositions, generalizing the result of Bondal-Orlov [ 41 ]. The main conjecture in this setting is the following.
Conjecture E. Let퐶 be a smooth projective curve of genus푔. Then there exists a semiorthog- onal decomposition
b D (M퐶 (2, L)) = Db (pt), Db (퐶), Db (Sym2 퐶),... (22) ..., Db (Sym푔−2 퐶), Db (Sym푔−1 퐶), Db (Sym푔−2 퐶),...... , Db (Sym2 퐶), Db (퐶), Db (pt) . GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 13
This was conjectured independently by Narasimhan (as communicated in [ 122 ]) and the
authors (see [ 25 , Conjecture 7]). We also give a further renements of this conjecture in
Section 6.2 and Section 6.3 . The derived category of a Fano variety is one of the four main actors in the homological mirror symmetry between Fano varieties and Landau–Ginzburg models, and one can try
to leverage information on the other side of the mirror to understand Conjecture E . In
particular we will discuss in Section 6.4 the relationship between
• eigenvalues of quantum multiplication by c1 (푋) on quantum cohomology (index- ing components in the spectral decomposition of the Fukaya category of 푋, and under the generalized Dubrovin conjecture also of the derived category of 푋), as studied by Muñoz; • critical values of the Landau–Ginzburg potential to give evidence for this conjecture.
Another piece of evidence for Conjecture E can be found by giving identities for invariants 푖 involving M퐶 (2, L) and Sym 퐶. This can take on several forms (identities in the category
of Chow motives, on the level of Poincaré or Hodge–Deligne polynomials, ... [ 11 ]). We will prove the following identity in the Grothendieck ring of varieties, where L = [A1] is the Lefschetz class.
Theorem F. We have the equality
푔−2 푔−1 푔−1 ∑︁ 푖 3푔−3−2푖 푖 (23) [M퐶 (2, L)] = L [Sym 퐶] + (L + L )[Sym 퐶] + 푇 푖=0 in K0 (Var/푘), for some class 푇 such that (1 + L)· 푇 = 0.
We expect that 푇 = 0, but our method of proof is not strong enough to remove this error term.
The identity ( 23 ) gives evidence for Conjecture E : the semiorthogonal decomposition ( 22 ) would give an identity in the Grothendieck ring of categories, which agrees with the image (up to 2-torsion) of the identity from the theorem after taking the Bondal–Larsen–Lunts motivic measure.
The term 푇 in Theorem F vanishes in any realization of K0 (Var/푘) where (1 + L) is
invertible. In particular Theorem F implies [ 83 , Theorem 6.1], where a similar identity is shown to hold in the dimensional completion of the Grothendieck ring, using classes
of algebraic stacks coming from the Behrend–Dhillon identity [ 22 ]. What remains to be found is a proof of the expected identity, without error term and in K0 (Var/푘) itself. Outlook: mirror approach to the Atiyah–Floer conjecture. To end the introduction, we speculate on a variation on the theme of the Atiyah–Floer conjecture, suggested by mirror symmetry.
The Atiyah–Floer conjecture from [ 7 ] states that two homology theories, both introduced by Floer, are isomorphic. One is the instanton Floer homology HFinst (퐵) of a 3-manifold 퐵
[ 65 ]. The other is the symplectic Floer homology of two Lagrangian subvarieties 푅(퐵−) and 푅(퐵+) inside a symplectic variety 푅(Σ) associated to a Heegaard splitting 퐵 = 퐵− ∪Σ 퐵+
into two handlebodies along a common boundary surface Σ [ 66 ]. The notation 푅(∗) stands for the moduli space of at SU(2)-connections on ∗ = Σ, 퐵±, 퐵. The (smooth locus of the) variety 푅(Σ) is equipped with the Narasimhan–Atiyah–Bott–Goldman symplectic structure, for which 푅(퐵±) → 푅(Σ) are Lagrangian embeddings. 14 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Donaldson proposed to extend and categorify the Atiyah–Floer conjecture. For an overview
one is referred to [ 53 ]. Associated to the surface Σ one then has a category C(Σ), whilst the handlebodies 퐵± with boundary 휕퐵± = Σ dene objects in C(Σ). This assignment is subject to the condition that the morphism spaces are identied with the symplectic Floer homology. This (provisional) extended 4-dimensional TQFT is known as Donaldson–Floer theory. The category C(Σ) should be a Fukaya-like category for the symplectic variety 푅(Σ), whereas the Lagrangian subvarieties 푅(퐵±) dene objects in it. On the other side of the mirror we have graph potentials (and more complicated Landau– Ginzburg models constructed using graph potentials), and their categories of matrix factorizations. Similar to how categories of matrix factorizations of the potential 푥푛 appear as basic building blocks of Khovanov–Rozansky eld theories, the categories of matrix factorizations of graph potentials are expected to be mirror dual to the Fukaya categories of the symplectic varieties 푅(Σ). The upshot of this approach is that it is purely algebraic, whereas on the symplectic side one needs to do complicated analysis. Likewise, 푅(Σ) is singular, complicating the study
of the Fukaya category even further. In Remark 5.2.5 we discuss some observations on how graph potentials can be related under mirror symmetry to a suitable modication of the singular variety M퐶 (2, O퐶 ). In particular, one could try to construct Donaldson–Floer theory on the other side of the mirror. Acknowledgments. We want to thank Najmuddin Fakhruddin, Yuri Lima, M.S. Narasimhan, Maxim Smirnov and Catharina Stroppel for interesting discussions. This collaboration started in Bonn in January–March 2018 during the second author’s visit to the “Periods in Number Theory, Algebraic Geometry and Physics” Trimester Program of the Hausdor Center for Mathematics (HIM) and the rst and third author’s stay in the Max Planck Institute for Mathematics (MPIM), and the remaining work was done in the Tata Institute for Fundamental Research (TIFR) during the second author’s visit in December 2019–March 2020 and the rst author’s visit in February 2020. We would like to thank HIM, MPIM and TIFR for the very pleasant working conditions. The initial progress on di erent parts of this work was reported at “Patchworking in Geometry and Topology” conference in Belalp (dedicated to the 70th anniversary of Oleg Viro) and “Geometry and Topology motivated by Physics” at the Monte Verità Conference center in Ascona (Ticino). The feedback from the participants greatly widened our understanding.
Several computations and experiments were performed using Pari/GP [ 152 ] and Sage [ 163 ]
(in particular we used PALP [ 111 ] routines to analyze lattice polytopes). The rst author was partially supported by the FWO (Research Foundation—Flanders). The third author was partially supported by the Department of Atomic Energy, India, under project no. 12-R&D-TFR-5.01-0500 and also by the Science and Engineering Research Board, India (SRG/2019/000513).
2. Graph potentials We will consider trivalent graphs of genus 푔. Such graphs are associated to pair of pants
decompositions of compact orientable surfaces in the sense of Hatcher–Thurston [ 90 ]. These are collections of disjoint circles on a surface Σ such that their complement is the disjoint union of spheres with 3 holes, also known as “pairs of pants” or trinions. In this section we associate a Laurent polynomial to each trivalent graph, and study how di erent Laurent polynomials associated to di erent graphs are related to each other. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 15
The dual of the pair of pants can be encoded as follows:
(24) = and to a decomposition of Σ into such pairs of pants we associate the graph whose vertices are the pairs of pants, and whose edges indicate how the pairs of pants are glued together. Loops arise from cutting a genus 1 surface with 1 puncture by a circle. For every pants decomposition one thus obtains a trivalent graph. For now we will consider trivalent graphs without making this link to geometry and topology explicit, but this will be important later.
2.1. Denition. Let 훾 = (푉, 퐸) be an undirected trivalent graph (possiby containing loops), which we will assume to be connected, and whose rst Betti number is 푔. Hence #푉 = 2푔 − 2, (25) #퐸 = 3푔 − 3. Recall that the homology (resp. cohomology) of a graph with coe cients in a ring 푅 (which 푉 퐸 for us will be either Z or F2) takes C0 (훾, 푅) = 푅 and C1 (훾, 푅) = 푅 with di erential given by the incidence matrix after choosing an orientation of the graph (resp. the 푅-linear dual of the incidence matrix). Hence rk H0 (훾, 푅) is the number of connected compo- nents, and rk H1 (훾, 푅) is the genus of the graph. Because we only consider the di erential when 푅 = F2 the choice of orientation is irrelevant for us. We will denote 1 (26) 푁e훾 := C (훾, Z) the free abelian group of 1-cochains on 훾, and
(27) 푀e훾 := C1 (훾, Z) the free abelian group of 1-chains. 푣 푉 푒 , 푒 , 푒 퐸 Let ∈ be a vertex. It is adjacent to the edges 푣푘 푣푙 푣푚 ∈ (which might coincide if there is a loop). We will denote the sublattice of 푁e훾 generated by the cochains 푥푖, 푥 푗, 푥푘 for 푥 푒 훿 푁 which 푖 ( 푣푎 ) = 푖,푣푎 by e푣.
Inside 푁e푣 we consider the sublattice 푁푣 generated by the eight cochains {±푥푖 ± 푥 푗 ± 푥푘 }. Using these we dene the sublattice
(28) 푁훾 ⊆ 푁e훾 É 푁 푁 as the image of the natural morphism 푣∈푉 푣 → e훾 . 푇 ∨ 푁 푇 ∨ 푁 푁 푁 We have the associated tori 훾 := Spec C[ 훾 ] and e훾 := Spec C[ e훾 ], so that 훾 and e훾 푇 ∨ 푇 ∨ are realized as the character lattice of 훾 and e훾 respectively, and hence as the cocharacter lattices of 푇훾 and 푇e훾 . ∨ ∨
푇 푇 The inclusion (28) induces an isogeny of the tori e훾 → 훾 , whose kernel × ∨ (29) 퐴훾 := Hom(푁e훾 /푁훾, C ) = (푁e훾 /푁훾 ) , ⊕푔 (which is isomorphic to F2 ) we wish to describe explicitly. 푀 푁 ∨ 푀 푁 ∨ 푀 푀 Consider the dual lattices 훾 := 훾 and e훾 := e훾 , for which we have e훾 ⊆ 훾 , and 퐴훾 푀훾 /푀e훾 . The following lemma summarizes the situation. 16 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Lemma 2.1.1. We have that ( ) ∑︁ 푀훾 = 푚 = 푤푒푒 | 푤푒 ∈ R, ∀푛 ∈ 푁훾 : h푛,푚i ∈ Z 푒 ∈퐸 (30) ( ) ∑︁ 1 = 푚 = 푤 푒 | 푤 ∈ Z, ∀푣 ∈ 푉 : 푤 + 푤 + 푤 ∈ Z 푒 푒 2 푒푖 푒푗 푒푘 푒 ∈퐸 and
(31) 퐴훾 H1 (훾, F2).
Here 푒푖, 푒푗, 푒푘 are edges adjacent to the vertex 푣.
In particular we also see that rational functions on the torus 푇훾 are 퐴훾 -invariant functions on the torus 푇e훾 . Colorings. Next we introduce colorings. We will use these to conveniently deal with a
generalization of the class of trivalent graphs, where half-edges are allowed, see Section 3.3
and Section 4.1 . Using colorings we can reduce such situations (which will correspond to punctured surfaces) in the cases that we are interested in to the already considered case of trivalent graphs without half-edges.
Denition 2.1.2. Let 훾 = (푉, 퐸) be a trivalent graph. A coloring is a function 푐 : 푉 → F2, interpreted as an F2-valued 0-chain on 훾.
If 푐(푣) = 0 we say that 푣 is uncolored, and if 푐(푣) = 1 we say that 푣 is colored. When drawing a graph, uncolored corresponds to a circle whilst colored corresponds to a disc . If it can be either we will indicate this by drawing .
We can now introduce one of the main objects of this paper. In Section 3.3 we will generalize this denition to also include half-edges, but for now this denition su ces.
Denition 2.1.3. Let 훾 = (푉, 퐸) be a trivalent graph, and let 푐 : 푉 → F2 be a coloring. Let 푒1, . . . , 푒3푔−3 be an enumeration of the edges. We will denote 푥푖 the coordinate variable in Z[푁e훾 ] associated to 푒푖 .
Let 푣 ∈ 푉 be a vertex, and denote 푒푖, 푒푗, 푒푘 the three edges incident to it. Then the vertex potential is the Laurent polynomial
∑︁ 푠푖 푠푗 푠푘 푊 := 푥 (−1) 푥 (−1) 푥 (−1) (32) e푣,푐 푖 푗 푘 ⊕3 (푠푖,푠푗 ,푠푘 ) ∈F2 푠푖 +푠푗 +푠푘 =푐 (푣) in Z[푁e푣]. Observe that 푊e푣,푐 is in the image of Z[푁푣] in Z[푁e푣]. Then we dene the graph potential as the Laurent polynomial ∑︁ (33) 푊e훾,푐 := 푊e푣,푐 푣∈푉 in Z[푁e훾 ]. Similarly 푊e훾,푐 is in the image of Z[푁훾 ] in Z[푁e훾 ]. If 푐 is the zero 0-chain, we will write 푊e푣,0 and 푊e훾,0.
푊 푇 ∨ 푊 We can also think of e훾,푐 as a regular function on the torus e훾 . Since e훾,푐 lies in the image 푁 푁 푊 푇 ∨ of Z[ 훾 ] in Z[ e훾 ], it follows that e훾,푐 descends to a regular function on the torus 훾 . This function will be denoted by 푊훾,푐 . This is summarized in the following commutative GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 17
푎 푏 1 푐 2
(a) Theta graph 푏 푎 푐 1 2 (b) Dumbbell graph
Figure 1. Trivalent genus 2 graphs diagram: 푇 ∨ e훾 (34) 푊f훾,푐
푇 ∨ 1 훾 A 푊훾,푐
Remark 2.1.4. In Section 4 we will study toric varieties with cocharacter lattice 푁훾 whose moment polytope is the polar dual of the Newton polytope of 푊훾,푐 . However the lattice 푁훾 has no natural basis, hence we will not express 푊훾,푐 in terms of a basis of 푁훾 . Instead we will work with 푊e훾,푐 and reinterpret everything we need for applications in terms of 푊e훾,푐 . For the vertex potential we have precisely the following two cases.
Example 2.1.5. We consider the following local picture of a trivalent vertex.
푎
푏 푐
Then the vertex potentials are precisely 푎 푏 푐 푊 = 푎푏푐 + + + , e훾,0 푏푐 푎푐 푎푏 (35) 1 푎푏 푎푐 푏푐 푊 = + + + . e훾,1 푎푏푐 푐 푏 푎
There are two trivalent graphs with 2 vertices, which we will call the Theta graph and
dumbbell graph respectively, and they are given in Figures 1(a) and 1(b) .
Example 2.1.6 (Theta graph). We consider the Theta graph, labeled as in Figure 1(a) . It is a trivalent graph of genus 2. We will consider two colorings for this graph, which will
su ce for us by Corollary 2.1.9 . Let 푐 denote the non-trivial coloring given by 푐(1) = 0, 푐(2) = 1. Then we have 푎 푏 푐 푊 = 푎푏푐 + + + , (36) e훾,0 2 푏푐 푎푐 푎푏 푎 푏 푐 1 푏푐 푎푐 푎푏 푊 = 푎푏푐 + + + + + + + . (37) e훾,푐 푏푐 푎푐 푎푏 푎푏푐 푎 푏 푐
Example 2.1.7 (Dumbbell graph). Alternatively for 푔 = 2 we can consider the dumbbell
graph, labeled as in Figure 1(b) . We will consider two colorings for this graph, which will 18 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
su ce for us by Corollary 2.1.9 . Let 푐 denote the non-trivial coloring given by 푐(1) = 0, 푐(2) = 1. Then we have 푎 4 푎 (38) 푊 = 푎푏2 + + + 푎푐2 + , e훾,0 푏2 푎 푐2 푎 2 1 푐2 (39) 푊 = 푎푏2 + + + + 2푎 + . e훾,푐 푏2 푎 푎푐2 푎
Independence of coloring for graph potentials. The graph potential depends a priori on the graph 훾 and the coloring 푐 ∈ C0 (훾, F2). But in fact we can identify many graph potentials as follows. If 푒 ∈ 퐸 is an edge connecting the vertices 푣 and 푣 0 of the graph, then we have an 0 associated 1-chain [푒] ∈ C1 (훾, F2). Its boundary is 휕[푒] = [푣] + [푣 ] ∈ C0 (훾, F2). This allows us to modify colorings 푐 by ipping the color of 푣 and 푣 0, preserving the parity of the coloring. Its e ect on the graph potential is explained by the following lemma.
Lemma 2.1.8. Let (훾, 푐) be a graph together with a coloring 푐. Let {푥푖 } be a system of 0 coordinates associated to 훾. Let 푒푘 be an edge, and set 푐 := 푐 + 휕[푒푘 ]. Then we have the equality of graph potentials
(40) 푊e훾,푐 = 푊e훾,푐0 after the biregular automorphism ( 푥 푖 ≠ 푘 (41) 푥 ↦→ 푖 푖 푥−1 푖 푘 푖 = 푥±, . . . , 푥± of the algebraic torus Spec Z[ 1 3푔−3]. Moreover we get a biregular automorphism 푇 ∨ 푊 푊 of 훾 that identies 훾,푐 and 훾,푐0 .
Proof. We only need to consider 푣 and 푣 0, as the vertex potentials for the other vertices are not modied. It also su ces to consider 푣. The vertex potentials are
∑︁ 푠푖 푠푗 푠푘 푊 = 푥 (−1) 푥 (−1) 푥 (−1) e푣,푐 푖 푗 푘 ⊕3 (푠푖,푠푗 ,푠푘 ) ∈F2 푠푖 +푠푗 +푠푘 =푐 (푣) (42) ∑︁ (−1)푠푖 (−1)푠푗 (−1)푠푘 푊 0 = 푥 푥 푥 e푣,푐 푖 푗 푘 ⊕3 (푠푖,푠푗 ,푠푘 ) ∈F2 푠푖 +푠푗 +푠푘 =푐 (푣)+1 and these sums agree after the given biregular automorphisms. More precisely, we par- ⊕3 , , , , , , , , , , , , , , , , , , , , , , tition F2 as {(0 0 0) (0 1 1) (1 0 1) (1 1 0)} t {(0 0 1) (0 1 0) (1 0 0) (1 1 1)} and the given biregular automorphism exchanges these subsets, but then the vertex potentials are precisely identied using the given biregular automorphism.
The second part of the proposition follows directly from the fact that each element of 퐴훾 is of order two and it acts on the variables 푥푖 ’s by a character and the biregular automor-
phism ( 42 ) either xes the variables or inverts them.
Because C1 (훾, F2) is generated by the 1-chains [푒푘 ] where 푒푘 runs over the edges, we obtain the following corollary.
Corollary 2.1.9. Let 훾 be a trivalent graph, and 푐 a coloring. Then the graph potential푊e훾,푐 only depends on the homology class [푐] ∈ H0 (훾, F2), up to biregular automorphism of the torus. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 19
←→
(a) Elementary transformation on a surface
푖 푙 푖 푙
←→
푗 푘 푗 푘
(b) Elementary transformation on the dual graph
Figure 2. Elementary transformations of trivalent graphs
푎 푐 푎 푐 푣 푥 1 ↦→ 푥 푣 푣 1 2 푣 푏 푑 2 푏 푑
(a) Elementary transformation with colors
푎 푐 푎 푐 푣 푥 1 ↦→ 푥 푣 푣 1 2 푣 푏 푑 2 푏 푑
(b) Elementary transformation without colors
Figure 3. Elementary transformations of trivalent graphs
2.2. Elementary transformations. For a given surface of genus 푔 there exist many (isotopy classes of) pair of pants decompositions. But they can be related via certain
operations, which in [ 90 , Appendix] are called (I), ..., (IV). It is remarked that (I) and (IV) in fact su ce to relate di erent decompositions for a surface, and that (IV) does not change the associated graph. Hence for our purposes we are only interested in the operation (I),
which we will call an elementary move. Topologically this can be described as in Figure 2(a)
on the surface and its dual graph in Figure 2(b) .
For later reference we summarize the discussion from [ 90 ] as follows.
Proposition 2.2.1 (Hatcher–Thurston). Elementary transformations act transitively on the set of colored trivalent graphs of genus 푔 with 푛 colored vertices, for 푛 = 0,..., 2푔 − 2. 20 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Elementary transformations will give us a family of mutations of Laurent polynomials, as
discussed in [ 5 ]. We are mostly interested in the behavior of their periods under mutations induced by operations on the trivalent graph. When considering the behavior of an elementary transformation of a trivalent graph on its associated graph potential, we can always split the graph potential as 푊 푊 mut 푊 frozen (43) e훾,푐 = e훾,푐 + e훾,푐 where the mutated part involves the variables associated to the vertices 푣1 and 푣2 adjacent to the edge corresponding to the variable 푥. The frozen part of the graph potential is not changed, and can be ignored. Elementary transformation with colors. Let us now describe how the graph potential changes when two edges attached to vertices of di erent colors come together, i.e. we con- 0 sider Figure 3(a) . We will denote by 푐 the coloring of both 훾 and 훾 , under the identication of the vertices. Before the transformation we have 푥 푑 푑 1 푎푏 푎푥 푏푥 푊 mut = 푥푐푑 + + + + + + + e훾,푐 푐푑 푐푥 푐푥 푎푏푥 푥 푏 푎 (44) 1 1 푐 푑 1 푎 푏 = 푎푏 + + + + 푥 푐푑 + + + . 푥 푎푏 푑 푐 푐푑 푏 푎 Denoting 1 휇 := (푎 + 푏푐푑)(푏 + 푎푐푑) 푎푏푐푑 (45) 1 휈 := (푐 + 푎푏푑)(푑 + 푎푏푐) 푎푏푐푑 we can write it as 휇 푊 mut = + 휈푥. (46) e훾,푐 푥
After the transformation we have 푥 0 푏 푑 푎푐 푐 푥 0 mut 0 1 푊 0 = 푥 푏푑 + + + + + + + e훾 ,푐 푏푑 푑푥 0 푏푥 0 푎푐푥 0 푥 0 푎푥 0 푎푐 (47) 1 1 푏 푑 1 푎 푐 = 푎푐 + + + + 푥 0 푏푑 + + + . 푥 0 푎푐 푑 푏 푏푑 푐 푎 Denoting 1 휇0 := (푏 + 푎푐푑)(푑 + 푎푐푑) 푎푏푐푑 (48) 1 휈 0 := (푎 + 푏푐푑)(푐 + 푎푏푑) 푎푏푐푑 we can write it as 휇0 mut 0 0 (49) 푊 0 = + 휈 푥 . e훾 ,푐 푥 0
Lemma 2.2.2. Let 휇, 휈, 휇0, 휈 0 be Laurent polynomials such that 휇휈 = 휇0휈 0. Then the Laurent 0 휇 + 휈푥 휇 + 휈 0푥 0 polynomials 푥 and 푥0 are identied after a rational change of coordinates.
Proof. Setting 푧 = 휈푥 we have 휇 휇휈 + 휈푥 = + 푧. (50) 푥 푧 푧 휇0 On the other hand we can do the rational change of coordinates = 푥0 to get 휇 휇0 휇휈푥 0 (51) + 휈푥 = + . 푥 푥 0 휇0 GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 21
0 휇휈 휈 0 휇 + 휈푥 휇 + 휈 0푥 0 But by assumption we have 휇0 = , hence have identied 푥 and 푥0 .
Theorem 2.2.3. Let 훾 and 훾 0 be trivalent graphs related via an elementary transformation at an edge with di erent colors. Then
(1) the graph potentials 푊e훾,푐 and 푊e훾0,푐 are identied after a rational change of coordi- nates;
(2) the rational transformation is invariant under the action of 퐴훾 and 퐴훾0 and hence identies 푊훾,푐 and 푊훾0,푐0 .
Proof. The rst point follows from Lemma 2.2.2 , as we have 1 (52) 휇휈 = 휇0휈 0 (푎 + 푏푐푑)(푏 + 푎푐푑)(푐 + 푎푏푑)(푑 + 푎푏푐). (푎푏푐푑)2 푥 0 휇0 Hence the rational transformation we are using is given by = 휈푥 .
To prove the second point, observe that the group 퐴훾 acts on the variables 푎,푏, 푐,푑, 푥 via characters 휎푎, 휎푏, 휎푐, 휎푑, 휎푥 , satisfying the relations
(53) 휎푎휎푏 = 휎푥 = 휎푐휎푑 .
퐴 0 푎,푏, 푐,푑, 푥 0 휎 0, 휎 0, 휎 0, 휎 0, 휎 0 Likewise 훾 acts on the variables via characters 푎 푏 푐 푑 푥0 , satisfying the relations 휎 0휎 0 휎 0 휎 0휎 0 . (54) 푎 푐 = 푥0 = 푏 푑
To compute the action of 퐴훾 and 퐴훾0 we write the rational change of coordinates more explicitly as (푏 + 푎푐푑)(푑 + 푎푏푐) (55) 푥 0 = 푥 (푎 + 푏푐푑)(푏 + 푎푐푑) 0 where we have removed the factors 푎푏푐푑 (one coming from훾, the other from훾 ) as 휎푎휎푏휎푐휎푑 휎 0휎 0휎 0휎 0 퐴 0 and 푎 푏 푐 푑 are trivial so they do in fact cancel. The group 훾 acts on the numerator, the group 퐴훾 on the denominator, and we need to check that for every 푢 ∈ 퐴훾 퐴훾0 the action on the left- and right-hand side agrees. On the left-hand side we have that 푢 푥 0 휎 0 푢 푥 0 (56) · = 푥0 ( ) whilst on the right-hand side we have that 휎 0 (푢)푏 + 휎 0휎 0휎 0 (푢)푎푐푑 휎 0 (푢)푑 + 휎 0휎 0휎 0(푢)푎푏푐 (푏 + 푎푐푑)(푑 + 푎푏푐) 푏 푎 푐 푑 푑 푎 푏 푐 푢 · = 푥 (푎 + 푏푐푑)(푏 + 푎푐푑) 휎푥 (푢)푥 (휎푎 (푎) + 휎푏휎푐휎푑 (푢)푏푐푑)(휎푏 (푏) + 휎푎휎푐휎푑 (푢)푎푐푑) 휎 0 (푢)푏 + 휎 0 (푢)푎푐푑 휎 0 (푢)푑 + 휎 0 (푢)푎푏푐 (57) 푏 푏 푑 푑 = 휎푥 (푢)푥 (휎푎 (푎) + 휎푎 (푢)푏푐푑)(휎푏 (푏) + 휎푏 (푢)푎푐푑) 휎 0휎 0 휇0 = 푏 푑 (푢) . 휎푥휎푎휎푏 휈푥
But by the relations in ( 53 ) and ( 54 ) we have the necessary equality of characters. 22 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Elementary transformations without colors. Let us now describe how the graph po- tential changes when two edges attached to vertices of the same colors (and hence it is enough to assume that they have no coloring) come together, i.e. we consider Fig-
ure 3(b) . The proof is similar to the previous case, so not all details are given. Before the transformation we have 푥 푐 푑 푎 푥 푏 푊 mut = 푥푐푑 + + + + 푎푏푥 + + + e훾 푐푑 푑푥 푐푥 푏푥 푎푏 푎푥 (58) 1 푎 푏 푐 푑 1 1 = + + + + 푥 푐푑 + + + 푎푏 . 푥 푏 푎 푑 푐 푐푑 푎푏 Denoting 1 휇 := (푎푑 + 푏푐)(푎푐 + 푏푑) 푎푏푐푑 (59) 1 휈 := ( + 푎푏푐푑)(푐푑 + 푎푏) 푎푏푐푑 1 we can write it as 휇 푊 mut = + 휈푥. (60) e훾 푥
After the transformation we have 푥 0 푏 푑 푎 푐 푥 0 푊 mut = 푥 0푏푑 + + + + 푎푐푥 0 + + + e훾 푏푑 푑푥 0 푏푥 0 푐푥 0 푎푥 0 푎푐 (61) 1 푏 푑 푎 푐 1 1 = + + + + 푥 푏푑 + + 푎푐 + . 푥 0 푑 푏 푐 푎 푏푑 푎푐 Denoting 1 휇0 := (푎푏 + 푐푑)(푎푑 + 푏푐) 푎푏푐푑 (62) 1 휈 0 := ( + 푎푏푐푑)(푎푐 + 푏푑) 푎푏푐푑 1 we can write it as 휇0 (63) 푊 mut = + 휈 0푥 0. e훾 푥 0
We obtain the following analogue of Theorem 2.2.3 , where we use that 1 (64) 휇휈 = 휇0휈 0 = (1 + 푎푏푐푑)(푎푐 + 푏푑)(푎푏 + 푐푑)(푎푑 + 푏푐). (푎푏푐푑)2
Theorem 2.2.4. Let 훾 and 훾 0 be trivalent graphs related via an elementary transformation at an edge with the same colors. Then
(1) the graph potentials 푊e훾,푐 and 푊e훾0,푐 are identied after a rational change of coordi- nates;
(2) the rational transformation is invariant under the action of 퐴훾 and 퐴훾0 and hence identies 푊훾,푐 and 푊훾0,푐 .
As a direct application of Theorem 2.2.3 and Theorem 2.2.4 , we obtain the following corollary by tropicalizing the rational change of coordinates.
Corollary 2.2.5. Let (훾, 푐) and (훾 0, 푐 0) be two trivalent colored graphs related by elemen- 3푔−3 tary transformations, then there are piecewise linear automorphisms 푇훾,푐 of R that maps
(65) 푇훾,푐 : 푃훾,푐 → 푃훾0,푐0, GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 23
푎 푐 푎 푐 푥 ↦→ 푣1 푣2 푣 푏 푑 푏 푑
Figure 4. Edge contraction for an elementary transformation where 푃훾,푐 (resp. 푃훾0,푐0 ) are the polar duals of the Newton polytope of 푊e훾,푐 (resp. 푊e훾0,푐0 ). Elementary transformations in terms of edge contractions. More conceptually we can describe these elementary transformations in terms of edge contractions and potentials for quadrivalent vertices.
We will consider Figure 3 , and contract the edge 푥 between the vertices 푣1 and 푣2 into a ± vertex 푣 as in Figure 4 . Associated to this vertex we add the variable 푧 to the coordinate ring of the torus. Let 푣 ∈ 푉 be the quadrivalent vertex, and denote 푎,푏, 푐,푑 the four edges incident to it. Then the quadrivalent vertex potential is the Laurent polynomial (푎푏 + 푐푑)(푎푑 + 푏푐)(푎푐 + 푏푑)(1 + 푎푏푐푑) 1 + 푧 푐(푣) = 0 푎푏푐푑 2 푧 푊 := ( ) (66) e푣,푐 (푎 + 푏푐푑)(푏 + 푎푐푑)(푐 + 푎푏푑)(푑 + 푎푏푐) 1 + 푧 푐(푣) = 1. 푎푏푐푑 2 푧 ( ) Here the induced coloring 푐(푣) is dened as 푐(푣1) + 푐(푣2). Then the mutation of the graph potential can be described as the transformation (푎푏 + 푐푑)(푎푑 + 푏푐)(푎푐 + 푏푑)(1 + 푎푏푐푑) 1 (67) 푊 + 푊 ↦→ + 푧 e푣1,푐 e푣2,푐 (푎푏푐푑)2 푧 if 푐(푣1) + 푐(푣2) = 푐(푣) = 0, and (푎 + 푏푐푑)(푏 + 푎푐푑)(푐 + 푎푏푑)(푑 + 푎푏푐) 1 (68) 푊 + 푊 ↦→ + 푧 e푣1,푐 e푣2,푐 (푎푏푐푑)2 푧 if 푐(푣1) + 푐(푣2) = 푐(푣) = 1.
2.3. Periods of Laurent polynomials. Consider an 푛-cycle {|푥1| = ... = |푥푛 | = 1} in 푛 the torus (C×)푛, along with the normalized volume form √1 d푥1 ... d푥푛 . 2휋 −1 푥1 푥푛 푊 푥±, . . . , 푥± 푊 Denition 2.3.1. Let ∈ C[ 1 푛 ] be a Laurent polynomial and denote by [ ]0 its constant term. The period휋푊 (푡) of 푊 is dened as 1 푛 ∫ ∫ d푥 d푥 휋 (푡) := ··· 푊 1 ··· 푛 , (69) 푊 √ 푥 푥 2휋 −1 |푥1 |=...=|푥푛 |=1 1 푛 which can be identied with ∑︁ 푘 푘 (70) 휋푊 (푡) = [푊 ]0푡 푘 ≥0
for 푡 < 1/푊 (1). b 푘 We will often denote the constant term [푊 ]0 of the 푘th power of 푊 as 휋푘 . The inverse Laplace transform 휋푊 (푡) of 휋푊 (푡) is dened as
b ∑︁ [푊 푘 ] (71) 휋 (푡) := 0 푡푘 푊 푘! 푛≥0
24 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY b
The inverse Laplace transform 휋푊 (푡) converges for all 푡 and can also be written as 1 푛 ∫ ∫ d푥 d푥 휋 (푡) = ··· (푡푊 ) 1 ··· 푛 . (72) 푊 √ exp 푥 푥 2휋 −1 |푥1 |=...=|푥푛 |=1 1 푛
Finally we remark that in Section 3 , we will express 휋푊 (푡) as the trace of a trace-class operator in the Hilbert space ℓ2 (Z). In this way we have associated an integer sequence to a graph potential. Let us discuss the easiest example of this, where 푔 = 2.
Example 2.3.2 (Genus two). The periods for the two colorings of the Theta graph as
discussed in Example 2.1.6 are 휋 푡 푡 4 푡 8 푡 12 ... 푊 ( ) = 1 + 384 + 645120 + 1513881600 + (73) f훾,0 휋 (푡) = 1 + 8푡 2 + 216푡 4 + 8000푡 6 + 343000푡 8 + 16003008푡 10 + 788889024푡 12 + ... 푊f훾,푐
These are in fact the same as for the dumbbell graph, which will follow from Corollary 2.3.4 .
The goal of this section is to prove that the periods of the graph potential only depend on the genus of the graph and the parity of the coloring. This will follow from the following
lemma, which is a mild generalization of [ 5 , Lemma 1], whose proof we include for completeness’ sake.
Lemma 2.3.3. Let 휑 be an automorphism of the eld C(푥1, . . . , 푥푛) corresponding to a × 푛 푥±, . . . , 푥± rational transformation of the torus (C ) = Spec C[ 1 푛 ], given by a collection of × 푛 rational functions (푝1/푞1, . . . , 푝푛/푞푛). Assume that for some 훼 ∈ C (74) 휑∗휔 = 훼휔, where 휔 = dlog 푥1 ∧ ... ∧ dlog 푥푛 is the volume form.
푊 ,푊 0 푥±, . . . , 푥± 휑∗푊 푊 0 휋 푡 휋 0 푡 Let be Laurent polynomials in C[ 1 푛 ]. If = , then 푊 ( ) = 푊 ( ).
푍 푓 Î푛 푝 푞 Proof. Let 푓 be the vanishing locus of = 푖=1 푖 푖 . Consider the morphism × 푛 푛 (75) Log: (C ) → R : (푧1, . . . , 푧푛) ↦→ (log |푧1|,..., log |푧푛 |).
By [ 80 , Corollary 6.1.8] the image 퐴 := Log(푍푓 ), called the amoeba, is a proper subset 푛 푛 퐴 푟 푛 of R , such that R \ is a disjoint union of convex sets. There exists an element ∈ R>0 such that 휑 is regular in all points of the torus × 푛 (76) 푇푟 := {(푧1, . . . , 푧푛) ∈ (C ) | ∀푖 = 1, . . . , 푛 : |푧푖 | = 푟푖 } ⊆ 푈 .
푈 × 푛 푍 푇 푇 0 × 푛 푟, 푟 0 푛 Here := (C ) \ 푓 . The cycles 푟 , 푟 are homologous in (C ) for any ∈ R>0, hence × 푛 they dene the same homology class in H푛 ((C ) , Z) Z훾, where 훾 = [푇푟 ].
This implies that [휑 (푇푟 )] = 푘훾 ∈ H푛 (푇, Z) for some 푘 ∈ Z. But we also see that ∫ ∫ ∫ ∫ ∫ 훼 훼 휔 훼 푖∗ 휔 휑∗ 휔 휔 휔 푘 (77) = = 푈 ( 푈 ) = 푈 ( 푈 ) = = = 푇푟 푇푟 푇푟 휑 (푇푟 ) 푘푇푟0 × 푛 where 푖푈 : 푈 → (C ) and 휑푈 is the composition of the rational map 휑 with 푖푈 , which is everywhere dened. Hence 훼 = 푘 is an non-zero integer. 푟 푛 푇 푈 Now for any ∈ R>0 such that 푟 ⊆ the equality ∫ 휔 ∫ 휔 ∫ 휔 ∫ 휔 (78) 훼 = 휑∗ = = 푘 − 푡푊 0 − 푡푊 − 푡푊 − 푡푊 푇푟 1 푇푟 1 휑 (푇푟 ) 1 푇푟 1 and non-vanishing of 훼 = 푘 implies the equality of periods.
Hence we obtain the following corollary. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 25
푎 푐 푥
푣1 푣2 푏 푑
(a) Local picture of trivalent graph with 푐 trivial
푎 푐 푥
푣1 푣2 푏 푑
(b) Local picture of trivalent graph with 푐 non-trivial
Figure 5. Local pictures of colored trivalent graphs
Corollary 2.3.4. Let (훾, 푐) and (훾 0, 푐 0) be related via elementary transformations or change of colors by the boundary of a 1-chain C1 (훾, F2). Then their periods agree, i.e. (79) 휋 (푡) = 휋 (푡). 푊f훾,푐 푊f훾0,푐0 and moreover are independent of the choice of ambient lattice, i.e. (80) 휋 (푡) = 휋 (푡), 푊f훾,푐 푊훾,푐
This explains why in Example 2.3.2 we could claim that the periods for the two distinct
genus two graphs agree. And combining the corollary with Proposition 2.2.1 we have that the periods only depend on the genus and the parity of the coloring.
2.4. Critical values. In Section 6.4 we will consider graph potentials as the potential for a Landau–Ginzburg model. For this reason it is important to analyze its critical values (and also its critical locus). For the purposes of this paper it su ces to describe the critical values, a more careful analysis of the critical locus will be useful for understanding other aspects of homological mirror symmetry. We only describe what we expect are all the
critical values (see also Section 6.4 ), we do not prove that these are indeed all.
Let 훾 be a trivalent graph, and 푐 be a coloring. By Corollary 2.1.9 we can assume that 푐 has at most one colored vertex. To determine the critical values of the graph potential 푊e훾,푐 we need to determine solutions to the equations 휕 (81) 푥푖 푊e훾,푐 = 0, ∀푖 = 1,..., 3푔 − 3. 휕푥푖 As for the description of the behavior of the graph potential under elementary transforma- tions in the previous section, we will study the local picture for (colored) trivalent graphs.
These are given in Figure 5 , where Figure 5(a) (resp. Figure 5(b) ) corresponds to the case where the coloring is trivial (resp. the vertex 푣2 is the unique colored vertex). We can write the potential in the uncolored (resp. colored) case as 1 1 푎 푏 푐 푑 푊 = 푥 푎푏 + + 푐푑 + + 푥−1 + + + + 푊 frozen (82) e훾,0 푎푏 푐푑 푏 푎 푑 푐 e훾,0 resp. 1 1 푎 푏 1 푊 = 푥 푎푏 + + 푐푑 + + 푥−1 + + 푐푑 + + 푊 frozen (83) e훾,푐 푎푏 푐푑 푏 푎 푐푑 e훾,푐 26 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY where we have split o the frozen part which does not involve the variable 푥. We call these expressions (without the frozen part) edge potentials. 푥 휕 푊 푥 휕 푊 Hence we are interested in solutions to the equations 휕푥 e훾,0 = 0, resp. 휕푥 e훾,푐 = 0, which can be rewritten by clearing the denominators as (84) 푥 (1 + 푎푏푐푑)(푎푏 + 푐푑) = 푥−1 (푎푐 + 푏푑)(푎푑 + 푏푐) resp. (85) 푥 (푎 + 푏푐푑)(푏 + 푎푐푑) = 푥−1 (푐 + 푎푏푑)(푑 + 푎푏푐).
Perfect matchings. The graph potential is dened as a sum over the vertex potentials associated to the vertices. But for certain graphs we can alternatively write the graph potential as a sum of edge potentials. For this we need a perfect matching, i.e. a subset 푃 ⊆ 퐸 of edges such that every vertex is contained in precisely one 푒 ∈ 푃. If the genus of 훾 is 푔 then #푃 = 푔 − 1. Given a perfect matching 푃 for 훾 we can rewrite the graph potential by summing over the
edges in the matching, and using the expressions ( 82 ), ( 83 ) for the edge potentials. Let 푒 denote the edge between 푣1 and 푣2, and label the variables in Figure 5 as 푎푒,푏푒, 푐푒,푑푒, 푥푒 . Then in the uncolored case we have ∑︁ 1 1 푎 푏 푐 푑 푊 = 푥 푎 푏 + + 푐 푑 + + 푥−1 푒 + 푒 + 푒 + 푒 (86) e훾,0 푒 푒 푒 푎 푏 푒 푒 푐 푑 푒 푏 푎 푑 푐 푒 ∈푃 푒 푒 푒 푒 푒 푒 푒 푒 whilst in the colored case we let 푒푐 ∈ 퐸 denote the colored vertex and omit the subscript 푒푐 from the variables, so that 1 1 푎 푏 1 푊 = 푥 푎푏 + + 푐푑 + + 푥−1 + + 푐푑 + e훾,푐 푎푏 푐푑 푏 푎 푐푑 (87) 푎 푏 푐 푑 ∑︁ 푥 푎 푏 1 푐 푑 1 푥−1 푒 푒 푒 푒 . 푒 푒 푒 + + 푒 푒 + + 푒 + + + 푎푒푏푒 푐푒푑푒 푏푒 푎푒 푑푒 푐푒 푒 ∈푃\{푒푐 }
By Petersen’s theorem, a trivalent graph which is bridgeless (i.e. we cannot remove an edge to make it disconnected) has at least one perfect matching. Such a graph exists for every
genus 푔. Using Theorems 2.2.3 and 2.2.4 we have that the graph potentials for di erent graphs of the same genus are related via rational changes of coordinates, hence its critical values do not depend on the choice of graph.
Proposition 2.4.1. Let 훾 be a trivalent graph. Let 푐 be a coloring with at most one colored vertex. Then the following are critical values of the graph potential 푊e훾,푐 : • purely real critical values: (88) 8푔 − 8 − 16푘 for 푘 = 0, . . . ,푔 − 1;
• purely imaginary critical values: √ (89) (8푔 − 16 − 16푘) −1 for 푘 = 0, . . . ,푔 − 2.
The critical value 0 is listed twice, depending on the parity of 푔 as a purely real or a purely imaginary critical value.
Proof. If we assign ±1 to the variables associated to the edges in the perfect matching, then it is immediate that assigning 1 to all variables associated to edges outside the perfect
matching gives a solution to the equations ( 84 ), ( 85 ). If we choose 푘 edges from 푃 ⊆ 퐸 to
have value −1, evaluating ( 86 ), ( 87 ) gives the value 8푔 − 8 − 16푘. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 27 √ If we assign ± −1 to the variables√ associated to the edges in the perfect matching, then it is immediate that assigning −1 to all variables associated to edges outside the perfect
푘 푃 퐸 matching gives√ a solution to the equations (84), (85). If we choose edges√ from ⊆ to have value − −1, evaluating ( 86 ), ( 87 ) gives the value (8푔 − 16 − 16푘) −1.
One can also show that the set of critical points used in the analysis is preserved under elementary transformations. This follows from the precise formulae in the rational change of coordinates, but we leave the details to the interested reader. Conifold points. For future consideration, we want to look at a certain critical value. If 푊 is a Laurent polynomial with positive coe cients such that the origin is contained in the Newton polytope (which is the convex hull of the exponents of monomials with
nonzero coe cients), then it is shown in [ 74 ] that there is a unique critical point with strictly positive (real) coordinates. This critical point is called the conifold point 푥con, and plays a role in the context of (homological) mirror symmetry. We can perform this analysis for the graph potentials we have just introduced.
Proposition 2.4.2. Let 훾 be a trivalent graph, and let 푐 be a coloring. The conifold point × 3푔−3 of the graph potential 푊e훾,푐 is given by (1,..., 1) ∈ (C ) , and the value of the graph potential at the conifold point is 8푔 − 8.
Proof. This follows from direct computation: for every vertex 푣 we have d푊푣,푐 (1, 1, 1) = 0 푊 , , 푊 Í 푊 푊 ,..., Í 푊 , , 푉 and 푣,푐 (1 1 1) = 4, hence d 훾,푐 = 푣∈푉 d 푣 = 0 and 훾,푐 (1 1) = 푣∈푉 푣,푐 (1 1 1) = 4# .
3. Topological antum field theories from graph potentials
The invariance under elementary transformations from Section 2.2 can be used to dene a two-dimensional topological quantum eld theory (or 2d TQFT for short), which will give an e cient computational tool to compute period sequences. Let us quickly recall what
2d TQFTs are, using the functorial description from Atiyah [ 8 ]. A two-dimensional topological quantum eld theory is a symmetric monoidal functor
(90) 푍 : (Bord2, t) → (C, ⊗) from the symmetric monoidal category Bord2 of 2-bordisms to a symmetric monoidal category (C, ⊗). An object in Bord2 is an oriented closed curve (a closed topological manifold of real dimension one), i.e. a (possibly empty) disjoint union of copies of 푆1, and a morphism (or bordism) from 퐸 to 퐹 is an equivalence class of oriented compact surfaces 푀 together with identications of the boundary 휕푀 to 퐸 and 퐹. The equivalence relation identies bordisms via orientation-preserving di eomorphisms, keeping only the 푆1 푆1 topological information. We will write 0 (respectively 1) for the circle with anticlockwise (respectively clockwise) orientation. 3.1. Restricted TQFT. For our purposes, we will take as our target category the category of real Hilbert spaces (not necessarily nite-dimensional), with the space of bounded operators between Hilbert spaces as morphisms. This category is a monoidal category under the Hilbertian tensor product but it is not rigid. As rigidity implies the existence of the trace function for all endomorphisms which forces the Hilbert spaces to be nite- dimensional. This also means that we cannot consider the traces of the identity operator
as in [ 8 , page 180], and hence we must not consider cups and caps. However to every Hilbert space H, we can consider its continuous dual H∗ which is linearly isomorphic to H. Moreover, this gives a natural linear isomorphism between the Hilbertian ∗, tensor product H1 ⊗ H2 and the space of Hilbert–Schmidt operators HS(H1 H2). 28 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
The category of Hilbert spaces has additional structure, by the Riesz representation the- orem. Namely, to every bounded operator 푓 : H1 → H2 we can assign the adjoint ∗ operator 푓 : H2 → H1 of a morphism, which satises the following properties: (1) Id∗ = Id; (2) (푓 ◦ 푔)∗ = 푔∗ ◦ 푓 ∗; (3) 푓 ∗∗ = 푓 . Categories with morphisms satisfying the above conditions are known as ∗-categories. The category Bord2 is naturally a ∗-category as to every bordism we can assign the opposite bordism.
If 푀 is an object in Bord2 and we let 푀 be the surface 푀 with opposite orientation, and we take our target category as the category of Hilbert spaces as above, then 푍 (푀) is nite- 푀 푀, dimensional. This follows from the existence of both the evaluation map HomBord2 ( t ∅) and the coevaluation map which in turn implies that the vector space 푍 (푀) has a rigid
dual, see for example [ 124 , Proposition 1.1.8]. Hence to allow innite-dimensional spaces we also need to slightly modify our source category, by discarding some bordisms in Bord2 while the objects remain the same. The new category will be denoted by RBord2. Namely we will only consider bordisms given by surfaces Σ푔,푛 (where 푔 is the genus and 푛 the number of boundary components) for which the Euler characteristic 2 − 2푔 − 푛 is strictly negative, together with cylinders (to ensure we have identity morphisms), braidings and handles considered as elements 푆1 푆1, of HomRBord2 ( t ∅). This will then su ce to dene invariants of closed surfaces of genus 푔 ≥ 2.
We have thus dened RBord2 as a non-full subcategory of Bord2. Observe that by the
symmetries inherent in the denition of these TQFT’s (see Lemmas 3.2.4 and 3.2.6 ), the pair 1 1 1 1 1 1 of pants Σ0,3 appears as a morphism 푆 t 푆 → 푆 and 푆 → 푆 t 푆 with the appropriate orientations, but also as a morphism 푆1 t 푆1 t 푆1 → ∅ and ∅ → 푆1 t 푆1 t 푆1.
Remark 3.1.1. The category RBord2 is not a ∗-category. Observe that we are not allowing , 푆1 푆1 opposite handles (nor cups or caps) in the morphisms HomRBord2 (∅ t ) in the restricted bordisms category RBord2. A handle combined with an opposite handle gives a torus whose corresponding assignment is the trace of the identity operator. Since our target category is the category of Hilbert spaces, the trace of identity may not be dened. This is one of the main reasons for considering the restricted bordism category RBord2.
In fact, we will describe a family of 2d TQFT’s, parametrized by 푡 ∈ R. It is only by considering the entire family of TQFT’s that we can e ciently compute period sequences.
Remark 3.1.2. Two-dimensional TQFT’s (with values in nite-dimensional vector spaces)
can equivalently be described using Frobenius algebras [ 58 , 3 ]. Because we consider a more general target category, and restricted the possible bordisms, we do not get the usual notion of a unital Frobenius algebra, but rather we get a Hilbertian algebra. The 2 algebra structure on ℓ (Z) with an associated Spin-structure, in the sense of [ 150 ] should be analyzed. We leave this for future work.
3.2. The graph potential TQFT. We x 푡 ∈ R. Using graph potentials we will construct a TQFT with values in Hilbert spaces for every value of 푡. Later we will allow 푡 to vary, after we have made the identication ensuring that at least the partition function for surfaces without punctures is related to the period of a graph potential, and therefore is well-behaved when we let 푡 vary. This su ces for our purposes. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 29
Witten–Dijkgraaf–Verlinde–Verlinde (WDVV) equations. We can describe a 2d TQFT in terms of the WDVV equations. For this, let (H, h−, −i) be a Hilbert space. This will be 푆1 ℓ2 the value of our 2d TQFT for 휖 , and later on we will take it to be (Z). We have a multilinear form h−, −i ⊗푛 : H⊗푛 → R for all 푛 ≥ 1, and a natural pair- ⊗푛 ⊗푛−2 ing h−, −i푖,푗 : H → H for all 푖 < 푗 by pairing the 푖th and 푗th factor. ⊗3 Consider an assignment M3 (푡) = M3 (푡;푖, 푗, 푘) ∈ H , where we will use 푖, 푗, 푘 to refer to the three tensor factors. This labeling allows us to refer to specic factors in repeated tensor products of the element M3 (푡).
Assume that M3 (푡) is symmetric in its factors. Then we can dene the assignment ⊗4 (91) M4 (푡) = M4 (푡;푖, 푗, 푘, 푙) := h−, −i푚,푛 (M3 (푡;푖, 푗,푚) ⊗ M3 (푡;푘, 푙, 푛)) ∈ H .
Here h−, −i푚,푛 refers to the natural pairing of the third and sixth factor.
⊗3 ⊗4 Denition 3.2.1. We say that M3 (푡) ∈ H is a solution to the WDVV equation if M4 (푡) ∈ H is symmetric in 푖, 푗, 푘, 푙.
Let M3 (푡) be such an assignment which satises the WDVV equation. Then we can construct a 2d TQFT as follows.
Construction 3.2.2. Let Σ푔,푛 be an oriented surface of genus 푔 with 푛 punctures, such that 2 − 2푔 − 푛 < 0. For every pair of pants, we will consider its dual graph. This is an oriented trivalent graph with one vertex and three half-edges. If the boundary circle is 푆1 oriented anticlockwise i.e. 0, then the half-edge is oriented outwards and vice versa as
shown in Figure 6 .
휖 = 0
=
휖 = 1 휖 = 1
Figure 6. Dual graph to oriented pair of pants
By choosing a pair of pants decomposition of Σ푔,푛 we assign the dual graph 훾. Observe that 훾 has oriented half-edges and internal edges are unoriented. We can assign ! 푡 Ì , Ì 푡 푖, 푗, 푘 ⊗푛 (92) MΣ푔,푛 ( ) := h− −i푎,푏 M3 ( ; ) ∈ H 푒 ∈퐸int 푣∈푉
where we use the labeling for internal edges and trivalent vertices as in Figure 7 , and the tensor product over the internal edges means we apply all possible pairings h−, −i푎,푏 where 푎 푏 Ë 푡 푖, 푗, 푘 the vertices and refer to specic factors in the tensor product 푣∈푉 M3 ( ; ). Using the relationship between 2d TQFT’s and solutions to the WDVV equation (see
e.g. [ 70 , §9], where the WDVV equation is referred to as the associativity equation) we obtain the following result.
푖 푒 푎 푏 푗 푣 푘
Figure 7. Labeling for internal edges and trivalent vertices 30 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
⊗3 Proposition 3.2.3. Let M3 (푡) ∈ H be a solution to the WDVV equations. Then Con-
struction 3.2.2 is independent of the choice of the pair of pants decomposition.
Graph potentials as solutions to the WDVV equations. Now we will revisit the set-
ting introduced in the introduction, in particular the functional equation in ( 1 ) and the ensuing discussion. We set H to be ℓ2 (Z). By Fourier expansion, any 푓 ∈ ℓ2 (Z) can be written as
∑︁ 푖 (93) 푎푖푧 , 푖 ∈Z
푖 2 where the collection {푧 }푖 ∈Z is a complete orthonormal basis of ℓ (Z) with respect to the standard pairing on ℓ2 (Z) given by * + 푓 푧 ,푔 푧 ∑︁ 푎 푧푖, ∑︁푏 푧 푗 h ( ) ( )i푧 = 푖 푗 푖 ∈Z 푗 푧 ∑︁ (94) = 푎푖푏푖 푖 ∈Z 1 ∫ d푧 = 푓 (푧)푔(푧−1) . √ 푧 2휋 −1 푆1
The WDVV equation can be rephrased as follows in this particular setting. Consider 2 ⊗3 a 푓 (푥1, 푥2, 푥3) ∈ ℓ (Z) . Assume now that there exists a function 휑 such that
(95) 푓 (푥1, 푥2,푦)푓 (푥3, 푥4,푦) = 푓 (푥1, 푥3, 푧)푓 (푥2, 푥4, 푧)
2 ⊗6 in ℓ (Z) , where 푦 = 휑 (푧, 푥1, 푥2, 푥3, 푥4) is a function such that the Jacobian of 휑 is the 휑 푑푧 푑푦 identity. In that case we get that ∗ 푧 = 푦 , and hence ∫ 푦 −1 d h푓 (푥1, 푥2,푦) ⊗ 푓 (푥3, 푥4,푦 )i푦 = 푓 (푥1, 푥2,푦)푓 (푥3, 푥4,푦) 푆1 푦 (96) ∫ d푧 = 푓 (푥1, 푥3, 푧)푓 (푥2, 푥4, 푧) 푆1 푧 −1 = h푓 (푥1, 푥3, 푧) ⊗ 푓 (푥2, 푥4, 푧 )i푧 in ℓ2 (Z) ⊗4 giving a solution to the WDVV equation. Taking the logarithm of the function 푓 we can interpret the multiplicative condition as an additive condition, and this brings us
in the setting of Section 2 .
In ( 32 ) we have dened the vertex potential at a vertex 푣 ∈ 푉 of a colored graph (훾, 푐). Using the variables 푥,푦, 푧, and writing 휖 = 푐(푣) we have that
(−1)휖 −1 −1 (−1)휖 −1 −1 (−1)휖 −1 −1 (−1)휖 (97) 푊e푣,휖 = (푥푦푧) + (푥푦 푧 ) + (푥 푦푧 ) + (푥 푦 푧) .
We observe that 푊e푣,휖 is symmetric in the variables 푥,푦, 푧. We also have the following symmetries, aside from the symmetry in the variables.
Lemma 3.2.4. The vertex potential satises
−1 −1 푊e푣,0 (푥,푦, 푧) = 푊e푣,0 (푥 ,푦 , 푧) −1 = 푊e푣,1 (푥 ,푦, 푧) (98) −1 푊e푣,1 (푥,푦, 푧) = 푊e푣,0 (푥 ,푦, 푧) −1 −1 −1 = 푊e푣,0 (푥 ,푦 , 푧 ). GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 31
Denition 3.2.5. Let 휖 be 0 or 1, and 푡 ∈ R. We dene (99) 푓휖 (푥,푦, 푧; 푡) := exp 푡푊e푣,휖 (푥,푦, 푧) 푎+푏+푐+푑 ∑︁ 푡 휖 휖 휖 = 푥 (−1) ((푎+푏)−(푐+푑))푦 (−1) ((푎+푐)−(푏+푑))푧 (−1) ((푎+푑)−(푏+푐)) . 푎!푏!푐!푑! 푎,푏,푐,푑 ≥0
The following lemma follows directly.
2 ⊗3 Lemma 3.2.6. For all 푡 ∈ R we have that 푓휖 (푥,푦, 푧; 푡) ∈ ℓ (Z) . It is moreover symmetric in 푥,푦, 푧.
Finally, as observed above, we have translated between an additive and a multiplicative
form of the WDVV equation, and hence by Section 2.2 we obtain the following
Proposition 3.2.7. For all 푡 ∈ R we have that 푓휖 (푥,푦, 푧; 푡) satises the WDVV equation.
We now come to the essential construction.
Construction 3.2.8. Let 푡 ∈ R. The graph potential eld theory Zgp (푡) for 푡 is dened as follows. 휖 , 휖 , 휖 Let Σ0,3;휖1,휖2,휖3 be a pair of pants, and 1 2 3 denote the orientation of the three boundary circles. Let 훾 be the trivalent graph on one vertex and three oriented half-edges which is dual to the pair of pants. Then we dene the partition function as
휖 휖 휖 gp 푡 푡푊 푥 (−1) 1 ,푦 (−1) 2 , 푧 (−1) 3 (100) Z ( )(Σ0,3;휖1,휖2,휖3 ) := exp e푣,0 ( ) ℓ2 ⊗3 푣 훾 푥,푦, 푧 in (Z) . Here is the unique vertex of the dual graph of Σ0,3;휖1,휖2,휖3 . We assign (resp. 푥−1,푦−1, 푧−1) as coordinate variables corresponding to the half-edges that are oriented
outwards (resp. inwards) as shown in Figure 8 .
푥 푥 푥−1
푧−1 푦−1 푧 푦 푧 푦−1
Figure 8. Variable attachment to oriented graphs
The assignment to the oriented pair of pants is well-dened by virtue of Lemma 3.2.4 .
Moreover Lemma 3.2.4 also implies that gp 푡 푡푊 푥,푦, 푧 , (101) Z ( )(Σ0,3;휖1,휖2,휖3 ) = exp e푣,휖 ( ) where 휖 = 휖1 + 휖2 + 휖3.
Hence all the graphs shown in Figure 8 have the same partition function exp 푡푊e푣,0 (푥,푦, 푧) .
Similarly all the graphs in Figure 9 have the same partition function exp 푡푊e푣,1 (푥,푦, 푧) .
푥−1 푥−1 푥
푧 푦 푧−1 푦−1 푧−1 푦
Figure 9. Variable attachment to oppositely oriented graphs 32 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
This is an additional feature of our partition function and we color the vertex of 훾 with color 휖 = 휖1 + 휖2 + 휖3 to mark this feature. (푆1, 푆1 ) (ℓ2 ( )) To the cylinder in HomRBord2 휖 휖+1 we assign the identity morphism in End Z , and similarly for any braiding we just assign the identity morphism on all the factors. The (푆1 t 푆1 , ∅) handle in HomRBord2 휖 휖+1 is assigned the evaluation map which is also natural pairing on the Hilbert space ℓ2 (Z).
Let Σ푔,푛 be a connected oriented surface of genus 푔 with 푛 boundary components satis- fying 2 − 2푔 − 푛 < 0. Consider a pair of pants decomposition for Σ푔,푛, and let 훾 be the trivalent dual graph with 푛 half-edges determined by the pair of pants decomposition. We incorporate the orientation of Σ푔,푛 as follows. Cut 훾 along the internal edges 퐸int to form a forest consisting of 2푔 − 2 trivalent graphs with one vertex and three half-edges with appropriate orientations. The coloring of the vertices is determined by the parity of the orientations of the number of clockwise circles of each pair of pants. We set ! gp Ì Ì (102) Z (푡)(Σ푔,푛) := h−, −i푎,푏 exp 푡푊e푣,휖 (푥,푦, 푧) 푒 ∈퐸int 푣∈푉 2 ⊗푛 in ℓ (Z) . As in ( 92 ), we use the labeling for internal edges and trivalent vertices as
in Figure 7 , and the tensor product over the internal edges means we apply all possible pairings h−, −i푎,푏 where the vertices 푎 and 푏 refer to specic factors in the tensor product indexed by the set of vertices of 푉 .
By ( 100 ), ( 101 ) and Lemma 3.2.4 , we get
(−1)휖1 (−1)휖2 (−1)휖3 (−1)휖1 (−1)휖2+1 (−1)휖3+1 (103) 푊푣,0 (푥 ,푦 , 푧 ) = 푊푣,0 (푥 ,푦 , 푧 ). Equating the variables 푦 and 푧 and taking inner product in the variable 푦, we get
(−1)휖1 (−1)휖 (−1)휖+1 (−1)휖1 (−1)휖+1 (−1)휖 (104) h푊푣,0 (푥 ,푦 ,푦 )i푦 = h푊푣,0 (푥 ,푦 ,푦 )i푦 This equality encodes that we obtain the same partition function for one-holed torus with decompositions obtained from cutting and gluing with two di erent circles as shown in gp Figure 10 . Thus the partition function Z (푡)(Σ1,1) for a one-holed torus is well-dened.
= 휖 휖
Figure 10. Fundamental relation for one-holed torus
Hence by Proposition 3.2.3 , the fact that cylinders and braidings are the identity, and that the handle is assigned the natural pairing, we obtain the following
Corollary 3.2.9. For all 푡 ∈ R, the assignment Zgp (푡) denes a two-dimensional TQFT on the restricted bordism category RBord2.
This proves Theorem B . It would be interesting to formalize this notion of a family of innite-dimensional TQFT’s, so that one can consider all of them together. We will not develop such a formalism. Rather we will only consider the family of partition functions for closed surfaces, dened by this
TQFT. This will allow us in Section 3.3 to obtain an e cient method to compute periods of graph potentials. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 33
푒 푒 0 푒 00 푎 푏 푎 푏
Figure 11. Cutting an internal edge
3.3. Computing periods via the graph TQFT. We can now turn our discussion to a practical method to compute the periods of graph potentials using topological quantum
eld theories. We also refer to [ 55 , 56 ] for various interesting interpretations of periods, and the use of innite-dimensional Hilbert spaces (albeit using di erent methods) to compute them.
The following denition extends Denition 2.1.3 , where we now allow leaves (or half- edges) to be present in the graph, as the denition of the vertex potentials does not depend on whether we have half-edges or not.
Denition 3.3.1. Let (훾, 푐) be a colored trivalent graph of genus 푔 with 푛 leaves. We dene the graph potential 푊e훾,푐 as the sum of vertex potentials.
For a graph with half-edges we let 퐸int denote the set of internal edges, i.e. we remove any half-edges from 퐸. Let us enumerate the variables associated to internal edges 퐸int as 푥1, . . . , 푥3푔−3+푛. Then we will use the following notation that interpolates between the value of the partition function on the surface Σ푔,푛 and the periods of the Laurent
polynomials in Section 5 .
Notation 3.3.2. Let (훾, 푐) be a colored trivalent graph of genus 푔 with 푛 leaves such that 2 − 2푔 − 푛 < 0. Orient the half-edges of 훾 such that half-edges attached to vertices 푣 of color 푐(푣) = 0 are pointing outwards while those attached to vertices of 푐(푣) = 1 are pointing inwards. This orientation is consistent with the orientation coming from pair of pants decomposition. Denote by (105) #퐸int 1 ∫ ∫ d푥 d푥3푔−3+푛 K (푡) := ··· 푡푊 (푥 , . . . , 푥 ) 1 ··· 훾,푐 √ exp e훾,푐 1 3푔−3+2푛 푥 푥 2휋 −1 (푆1)퐸int 1 3푔−3+푛 the corresponding element of ℓ2 (Z) ⊗푛.
We record the following important observation as a lemma. Lemma 3.3.3. If 푛 = 0, then ( 105 ) reduces to b
(106) K (푡) = 휋 (푡) b 훾,푐 푊f훾,푐 where 휋 (푡) is the inverse Fourier–Laplace transform of the period 휋 (푡) of the graph 푊f훾,푐 푊f훾,푐 potential.
In other words, if 휋 (푡) = Í 휋 푡푛, then K (푡) = Í 푝 푡푛, where 푝 = 휋 /푛!. 푊f훾,푐 푛≥0 푛 훾,푐 푛≥0 푛 푛 푛 We have the following proposition, which follows from the change of variables formula for integrals. It is an important computational tool in what follows.
Proposition 3.3.4. Let훾 0 be a trivalent graph with 푛+2 leaves. Let 푐 be a coloring. Consider half-edges 푒 0 and 푒 00 adjacent to vertices 푎 and 푏. We dene a new colored trivalent graph 훾 (with 푛 leaves) by replacing two leaves at the vertices 푎 and 푏 by the internal edge 푒 34 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
connecting 푎 and 푏, as in Figure 11 . Then 1 ∫ d푧 (107) K (푡) = √ K 0 (푡)| 0 00 . 훾,푐 훾 ,푐 푥푒 =푥푒 =푧 푧 2휋 −1 푆1 0 00 where 푥푒0 and 푥푒00 are variables associated to the leaves 푒 and 푒 attached to the vertices 푎 and 푏.
Observe that not all trivalent colored graphs that we considered in Section 2 arise as the dual graph of a pair of pants decomposition of a orientable surface whose boundary has induced orientations. Since the category RBord2 only consists of objects of this form, we need the following results to use the TQFT partition function Zgp (푡) e ectively to compute periods of arbitrary (훾, 푐).
2 The following proposition relates K훾,푐 (푡) to the partition function of the ℓ (Z)-valued TQFT that we constructed.
Proposition 3.3.5. Let Σ푔,푛 be a oriented surface of genus 푔 with 푛 boundary components and let (훾, 푐) be the dual graph obtained from a pair of pants decomposition of Σ푔,푛. If 3푔 − 3 + 푛 is even, then gp (108) K훾,푐 (푡) = Z (푡)(Σ푔,푛).
Proof. Proposition 3.3.4 allows us to write K훾,푐 (푡) as the iterated integral over the pair of pants decomposition. Since we need 3푔 − 3 + 푛 cuts to get to the pair of pants, the parity
being even guarantees that we can use Lemma 3.2.4 to match up the integral with the 2 gp norms in ℓ (Z) that appears in the denition of Z (푡)(Σ푔,푛).
Remark 3.3.6. If 3푔 − 3 + 푛 is odd, then we can compute K훾,푐 (푡) by rst cutting (훾, 푐) 0 0 0 0 along one edge to produce a new graph (훾 , 푐 ), and then use Proposition 3.3.5 for (훾 , 푐 )
and apply Proposition 3.3.4 .
Bessel functions. To adequately work with the partition functions of this topological quantum eld theory we recall that the modied Bessel function of the second kind is dened as ∑︁ 1 푧 2푚+훼 (109) I (푧) := . 훼 푚!Γ(푚 + 훼 + 1) 2 푚≥0 For our purposes we are only interested in the case 훼 = 0, with a rescaling of the argument. We will use the following notation.
Notation 3.3.7. We denote
∑︁ 1 2푚 (110) B(푧) := I = (2푧) = 푧 . 훼 0 (푚!)2 푚≥0
The following lemma explains why this function is relevant to us. It allows us to give an explicit expression for the partition function for the open necklace graph 훾1,2 from
Figure 12(a) . The necklace graph is the dual graph of the two-holed torus as shown in
Figure 13
Lemma 3.3.8. Let훾 = 훾1,2 be the open necklace graph as in Figure 12(a) , with one half-edge oriented outwards and the other one oriented inwards. Then −1 −1 (111) K훾,1 (푡) = B(푡 (푥 + 푦 )) B(푡 (푥 + 푦)). GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 35
푥 푦−1
1 2 (a) Open necklace graph ...
(b) Open necklace graph with many beads
Figure 12. Necklace graphs
푥 푦−1 = 휖 = 0 휖 = 1 1 2
Figure 13. Dual graph of the two-holed torus
Proof. The colored graph (훾, 1) in the statement of the lemma is the dual graph obtained from a pair of pants decomposition of a two-holed torus with one hole oriented anticlock- wise and the other one oriented clockwise. This is obtained by gluing two pairs of pants
as shown in Figure 14 .
휖 = 1 휖 = 0
휖 = 0 휖 = 1 휖 = 0 휖 = 1 휖 = 1 휖 = 0
Figure 14. Pants decomposition for the two-holed torus
Now the graph 훾1,2 is obtained two half-edges of a trivalent graph with one vertex as
shown in Figure 15
Figure 15. Genus one graph from gluing
By denition, Proposition 3.3.5 , and ( 92 ), we have (112)
D −1 −1 −1 E K훾,1 (푡) = exp(푡 (푊e푣,0 (푥,푢 , 푣 ) + 푊e푣,0 (푦 ,푢, 푣))) 푢,푣 1 ∬ d푢 d푣 = (푡 (푊 (푥,푢−1, 푣−1) + 푊 (푦−1,푢−1, 푣−1))) (113) √ exp e푣,0 e푣,0 푢 푣 (2휋 −1)2 푆1×푆1 1 ∬ d푢 d푣 = (푡 (푊 (푥,푢, 푣) + 푊 (푦−1,푢, 푣))) (114) √ exp e푣,0 e푣,0 푢 푣 (2휋 −1)2 푆1×푆1 which we can interpret as 푡 푊 푥,푢, 푣 푊 푦−1,푢, 푣 . (115) = [exp( ( e푣,0 ( ) + e푣,0 ( )))]푢0푣0 36 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Via the following sequence of rewrites
(푎+푏+푐+푑)+(푎0+푏0+푐0+푑0) ∑︁ 푡 0 0 0 0 (116) = 푥 (푎+푏)−(푐+푑)푦−(푎 +푏 )+(푐 +푑 ) 푎!푎0!푏!푏0!푐!푐 0!푑!푑 0! 푎,푏,푐,푑,푎0,푏0,푐0,푑0 푏+푑+푏0+푑0=푎0+푐0+푎+푐 푎0+푑0+푎+푑=푏+푐+푏0+푐0 (푎+푏+푐+푑)+(푎0+푏0+푐0+푑0) ∑︁ 푡 0 0 0 0 (117) = 푥 (푎+푏)−(푐+푑)푦−(푎 +푏 )+(푐 +푑 ) 푎!푎0!푏!푏0!푐!푐 0!푑!푑 0! 푎,푏,푐,푑,푎0,푏0,푐0,푑0 푐+푐0=푑+푑0 푎+푎0=푏+푏0 ∑︁ 푡푎1+푏1+푐1+푑1 (118) = 푥푎1−푐1푦−푏1+푑1 푎 푎 푎 푎0 푏 푎0 푐 푐 푐 푐 0 푑 푐 0 0 0 !( 1 − )! !( 1 − )! !( 1 − )! !( 1 − )! 푎1,푏1,푐1,푑1,푎,푎 ,푐,푐 0 푐1+푑1=2(푐+푐 ) 0 푎1+푏1=2(푎+푎 ) ∑︁ ∑︁ 푡 2(푚+푛) 푥푎1−푐1푦−푏1+푑1 (119) = 0 0 0 0 푎!(푎1 − 푎)!푎 !(푏1 − 푎 )!푐!(푐1 − 푐)!푐 !(푑1 − 푐 )! 푚,푛 푎1,푏1,푐1,푑1,푎,푐 푎≤푚,푐 ≤푛 푐1+푑1=2푛;푎1+푏1=2푚 ∑︁ ∑︁ 푡 2(푚+푛) 푎 푏 푐 푑 (120) = 푥푎1−푐1푦−푏1+푑1 1 1 1 1 푎1!푏1!푐1!푑1! 푎 푚 − 푎 푐 푛 − 푐 푚≥0,푛≥0 푎1,푏1,푐1,푑1,푎,푐 푎≤푚,푐 ≤푛 푐1+푑1=2푛;푎1+푏1=2푚 푏 −푑 ∑︁ ∑︁ 2푚 2푛 푥푎1−푐1 푦−1 1 1 (121) = 푡 2(푚+푛) 푚 푛 푎1!푏1!푐1!푑1! 푚≥0,푛≥0 푎1,푏1,푐1,푑1 푎1+푏1=2푚;푐1+푑1=2푛 we nally obtain
(122) = B(푡 (푥 + 푦−1)) B(푡 (푥−1 + 푦)) as desired.
As a direct corollary of Lemma 3.3.3 and Remark 3.3.6 we obtain
Corollary 3.3.9. Let 훾 = 훾1,2 be the graph 훾1,2 with no colored vertices, then
−1 −1 (123) K훾,0 (푡) = B(푡 (푥 + 푦)) B(푡 (푥 + 푦 )).
Before we discuss the general case, we will give a formula in the genus two case, which
will be relevant to the discussion in Remark A.1.2 .
Corollary 3.3.10. Let 훾 be a genus two graph, without half-edges. We consider the case 휖 = 1. Then the inverse Fourier–Laplace transform of the period 휋 (푡) is given by 푊f훾,휖
∑︁ (2푛!)2 (124) 푡 2푛. 푛!6 푛≥0
Proof. If we cut the colored graph (훾, 1) along any edge then we get back the graph
considered in Lemma 3.3.3 . Hence, by Lemma 3.3.3 , ( 92 ), and Lemma 3.3.8 , we have putting GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 37
푥 = 푦 b(125) ∑︁ ∑︁ 2푚 2푛 1 1 ∫ d푥 휋 (푡) = 푡 2(푚+푛) √ 푥푎1−푐1 ·(푥−1)푏1−푑1 푊f훾,1 푚 푛 푎 푏 푐 푑 푥 1! 1! 1! 1! 2휋 −1 푆1 푚≥0,푛≥0 푎1,푏1,푐1,푑1 푎1+푏1=2푚;푐1+푑1=2푛 ∑︁ ∑︁ 2푚 2푛 1 1 ∫ d푥 = 푡 2(푚+푛) 푥푎1−푐1−푏1+푑1 푚 푛 푎 푏 푐 푑 √ 푥 1! 1! 1! 1! 2휋 −1 푆1 푚≥0,푛≥0 푎1,푏1,푐1,푑1 푎1+푏1=2푚;푐1+푑1=2푛 −1 −1 = [B(푡 (푥 + 푥 ))B(푡 (푥 + 푥))]푥0 . Using the denition of the twisted Bessel function, the Vandermonde identity and some elementary manipulations, we obtain ! −1 2 ∑︁ ∑︁ 1 −1 2푛 2푛 [B(푡 (푥 + 푥 )) ] 0 = [(푥 + 푥 ) ] 0푡 푥 푎!2푏!2 푥 푛≥0 푎+푏=푛 푛 ! ∑︁ ∑︁ 푛!2 2푛 푡 2푛 = 푎!2 (푛 − 푎)!2 푛 푛!2 (126) 푛≥0 푎=0 ∑︁ 2푛2 푡 2푛 = 푛 푛!2 푛≥0 ∑︁ (2푛!)2 = 푡 2푛. 푛!6 푛≥0
Applying the machinery. For 푔 ≥ 3 we can describe an inductive procedure. Denote −1 −1 the product of Bessel functions B(푡 (푥 + 푦) B(푡 (푥 + 푦 ))) by T1 (푥,푦). Observe that −1 −1 −1 −1 T1 푥,푦 = T1 푥 ,푦 = B(푡 (푥 + 푦 ) B(푡 (푥 + 푦))) (127) −1 −1 −1 −1 T1 푥 ,푦 = T1 (푥,푦) = B(푡 (푥 + 푦) B(푡 (푥 + 푦 ))). This calculates the e ect of changing the orientation of the boundary of the two-holed torus.
Denition 3.3.11. Dene inductively using convolution the function
(128) T푘+1 (푥,푦) := [T푘 (푥, 푧) T1 (푧,푦)]푧0
The following proposition is an application of the usual machinery of determining the partition function by cutting a closed surface to easier pieces, and it explains the denition of T푘+1 (푥,푦).
Proposition 3.3.12. Let Σ푔,2 be a genus 푔 surface with two holes, one oriented anticlock- wise and the other oriented clockwise. Then the partition function of Σ푔,2 is given by gp −1 Z (푡)(Σ푔,2) = T푔 푥,푦 .
Proof. Consider the necklace graph 훾1,2 as in Lemma 3.3.8 show in Figure 12(a) . The genus two surface with two holes Σ2,2 with one hole oriented anti-clockwise and the other hole oriented clockwise can be obtained gluing two Σ1,2 as shown in Figure 16(a) .
In terms of the dual graph this is obtained by gluing two open necklace graphs Figure 13
as shown in Figure 16(b) . 38 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
+ = 휖 = 0 휖 = 1 휖 = 0 휖 = 1 휖 = 0 휖 = 1
(a) Gluing surfaces 푥 푦−1 푦 푧−1 푥 푧−1 + =
(b) Dual picture
Figure 16. Two-holed surfaces of genus two by gluing
Then the partition function is given by gp −1 −1 Z (푡)(Σ2,2) = hT1 (푥, 푧 ), T1 (푧,푦 )i푧 1 ∫ d푧 = (푥, 푧−1) (푧−1,푦−1) (129) √ T1 T1 푧 2휋 −1 푆1 −1 = T2 (푥,푦 ).
Hence by repeating this process, gluing copies of 훾1,2 to increase the genus, and applying
the TQFT formalism given by Proposition 3.2.3 we obtained the required result.
We now describe this result in terms of the period of graph potentials that we are interested in computing.
Proposition 3.3.13. Let (훾푔,2, 휖) be the (colored) dual graph of a surface with two holes obtained by gluing 푔 copies of Σ1,2 in a row. Then
푔 휖 푡 푥,푦 (−1) 푥,푦 (−1) . (130) K훾푔,2,휖 ( ) = T푔 = T푔
Proof. First of all observe that 휖 has the same parity as the genus 푔 of the surface by construction. Furthermore, if 푔 is odd, then we need to make even number of cuts to get a disjoint union gp
푡 푡 of Σ1,2’s. Hence by Proposition 3.3.5, we get that K훾푔,2,휖 ( ) = Z ( )(Σ푔,2). Then we are
done by Proposition 3.3.12 when 푔 is odd.
Now if 푔 is even, then consider the graph 훾푔−1,2 and an open necklace graph. Now by
Proposition 3.3.4 , we get −1 푡 푡 , 푧,푦 0 K훾푔,2,0 ( ) = [K훾푔−1,2,1 ( ) T1 ]푧 −1 −1 (131) = [T푔−1 푥, 푧 , T1 푧 ,푦 ]푧0
= T푔 (푥,푦) .
A direct very useful corollary of the above proposition that removes the restriction on the
matching of the parity of coloring and genus in Proposition 3.3.5 is the following
Corollary 3.3.14. Let 훾푔,2 be the open necklace graph of genus 푔 with two half-edges as shown in Figure 12(b) . Let 푐 be a coloring of 훾푔,2 and 휖 be the parity of the coloring, then
휖 푡 푥,푦 (−1) . (132) K훾푔,2,푐 ( ) = T푔 GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 39
Proof. If the parity of 푐 matches up with the parity of the genus as in Proposition 3.3.5 , then we are done. Otherwise choose one of the two half-edges and reverse the orientation 푡 of that half-edge. In the denition of K훾푔,2,푐 ( ), we never integrate over a half-edge variable of the original graph 훾, hence we are now reduced to the situation in Proposition 3.3.5 .
Now we can use Proposition 3.3.13 to get a formula for the periods of the closed (i.e. without half-edges) genus 푔 trivalent colored graphs (훾, 푐).
Proposition 3.3.15. Let 훾 be a trivalent graph of genus 푔 ≥ 2 without leaves and 푐 a coloring of the vertices. Let 휖 ∈ F2 denote the parity of the coloring. Then h (−1)휖 i (133) K훾,푐 (푡) = T푔−1 푥, 푥 . 푥0
Proof. By Corollary 2.1.9 we know that K훾,푐 (푡) only depends on the parity of the coloring. Hence we can assume that the graph has either one or zero colored vertices. Moreover since the periods are invariants under mutations, we can assume that 훾 is the closed
necklace graph of genus 푔 as in Figure 17 with one or zero colored vertex.
...
Figure 17. Closed necklace graph with many beads
First consider the case when 푔 and 푐 have the same parity. Then the result directly follows
from Proposition 3.3.13 by applying Proposition 3.3.4 . Now in the case 푔 and 푐 have opposite 푡 푒 훾 훾 , 휖 parity, then to compute K(훾푔,0,휖) ( ), cut an edge of the graph to get the graph ( 푔−1,2 )
considered in Proposition 3.3.13 . Then by we are done by rst applying Corollary 3.3.14
and then Proposition 3.3.4 .
The main result. This proposition gives an e ective method to compute the period of a graph potential, giving the main result of this section. It allows us to express (the inverse
Fourier–Laplace transform of) the periods of the Laurent polynomials from Section 2 using trace-class operators in ℓ2 (Z). Dene
• 퐴 to be the Hilbert–Schmidt operator given by T1 (푥,푦); 푆 푆 Í 푎 푥푛 Í 푎 푥−푛 • to be the bounded linear operator ( 푛∈Z 푛 ) = 푛∈Z 푛 . We consider 퐴 as matrix with respect to the orthonormal basis and reinterpret Corol-
lary 3.3.10 as follows.
Lemma 3.3.16. Let (훾2, 1) be a genus two graph with one colored vertex, then ∑︁ (2푛!)2 (134) K (푡) = tr(퐴푆) = 푡 2푛. 훾2,1 푛!6 푛≥0 In particular the composition 퐴푆 is trace class.
Since trace-class operators form an ideal, it follows that 퐴푎+1푆푏 is also trace-class for non-negative integers 푎,푏. Moreover 푆 commutes with 퐴. The following is the important computational tool for computing periods of graph poten- tials. 40 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Theorem 3.3.17. Let (훾, 푐) be a colored trivalent graph of genus 푔 ≥ 2 (without half-edges). Then, b
(135) 휋 (푡) = tr(퐴푔−1푆휖+푔) 푊f훾,푐 where 휖 denotes the parity of the number of colored vertices in 훾.
Proof. By Corollary 2.1.9 we can assume that the number of colored vertices is either zero
or one. To show the equality in ( 135 ) we use Lemma 3.3.3 and Proposition 3.3.15 , so that we need to show that
(−1)휖 푔−1 휖+푔 (136) [T푔−1 (푥, 푥 )]푥0 = tr(퐴 푆 ).
We do this by induction. If 푔 = 2, then this follows from Corollary 3.3.10 for 휖 = 1, whilst the case of 휖 = 0 is analogous. Next, observe that by the symmetry properties of 퐴, we have that 푆 commutes with 퐴. 푔 푥,푦 ℓ2 ⊗2 Í 푏 푥푖푦 푗 Now if is arbitrary, we consider T푔−1 ( ) ∈ (Z) , writing it as 푖,푗 ∈Z 푖,푗 for 푖 푔−1 푔−1 some matrix 퐵 = (푏푖,푗 ) in the orthonormal basis {푥 }푖 ∈Z. We claim that 퐵 = 퐴 푆 . This follows by induction, using that 푆 commutes with 퐴. Hence we get
(−1)휖 ∑︁ [T푔−1 (푥, 푥 )]푥0 = 푏푖,푗 푖,푗 ∈Z 푖+(−1)휖 푗=0 ∑︁ = 푏푖,푗 (137) 푖,푗 ∈Z 푖=(−1)휖+1 푗 ∑︁ 휖+1 푔−1 푔−1 = (푆 퐴 푆 )푖,푖 푖 ∈Z = tr(퐴푔−1푆휖+푔), again using that 푆 commutes with 퐴.
Remark 3.3.18. This gives an explicit and e cient method to compute the period se- quences, by truncating the power series in 푡. This method is independent of 푔, and the number of periods that one can compute depends only on the degree of the truncation, and not on 푔. Alternatively one can compute quantum periods of these moduli spaces via the abelian/non- abelian correspondence in Gromov–Witten theory. But in this approach one xes 푔 and does a special analysis for each value of 푔 (which is only feasible for low 푔). Developing the details of the abelian/non-abelian correspondence in this case and comparing the two methods is left for future work.
Some connections to other works in the literature. There exists a degeneration of our setup, which links it to mirror symmetry for Grassmannians of planes. For this we let 훾 be a genus zero graph with 푛 ≥ 3 half-edges and all vertices uncolored. The cardinality of the set of edges 퐸 is 2푛 − 3 and we assign variables 푥1, . . . , 푥2푛−3 as before to all the half-edges. Consider the substitution 휏 푌 푍 (푥,푦, 푧) → , , . (138) 푋 휏 휏 GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 41 where 푥,푦, 푧 are variables assigned to adjacent edges on a vertex 푣; the 푋, 푌, 푍 are new variables and 휏 is a formal parameter. In this set-up 푥 푦 푧 푊 (푥,푦, 푧) = 푥푦푧 + + + e푣,0 푦푧 푥푧 푥푦 (139) 푌푍 1 푋푌 푍푋 = 휏−1 + 휏4 + + 푋 푋푌푍 푍 푌 푊 Í 푊 푥,푦, 푧 휏푊 Now consider the Laurent polynomial e훾,0 = 푣∈푉 e푣,0 ( ), and take lim휏→0 e훾 . This is a Laurent polynomial G훾 in the variables 푋1, . . . , 푋2푛−3. This is exactly the Laurent
polynomial considered by Nohara–Ueda in [ 149 , Theorem 1.6], and brings us to the following remark.
Remark 3.3.19. The above observation and the TQFT results from Theorem B and Propo-
sition 3.2.3 tell us that the potential functions from Nohara and Ueda [ 149 ] associated to the Grassmannian of planes in an 푛-dimensional vector space also give a TQFT which one can call the Grassmannian TQFT. To the best of our knowledge this was not observed before.
In particular, by the above discussion and ( 139 ) the graph potential TQFT coming from moduli of bundles that we develop recovers the Grassmannian TQFT as a limit. This should be compared with a corresponding fact that genus zero conformal blocks recover invariants of tensor product representations in the limit.
4. Toric degenerations
The graph potentials introduced in Section 2 are part of the mirror symmetry picture for M퐶 (2, L).
Abstracting from his earlier work with Ciocan-Fontanine, Kim and van Straten [ 17 , 16 ] on mirror symmetry for Grassmannians and partial ag varieties, Batyrev proposed in
[ 15 , 18 ] a passage from a “small” toric degeneration 푋 of a Fano variety 퐹 to a Laurent polynomial 푊 , which shall be mirror dual to 퐹.
In this section we will recall a class of toric degenerations of M퐶 (2, L) (and M퐶 (2, O퐶 ))
introduced by Manon [ 128 ], which turn out to be closely related to the graph potentials we
have introduced in Section 2 : the Newton polytope of these Laurent polynomials agrees with the moment polytope of Manon’s degenerations.
We will recall the construction of the toric degenerations in Section 4.1 . For later appli- cations we need to analyze the singularities on the toric degeneration. In particular, we need to show that they have terminal singularities and admit a small resolution. These
properties will depend on the choice of trivalent graph, and in Section 4.2 we will formulate a criterion to check the former and provide a conjectural criterion for the latter.
4.1. Manon’s degeneration. In [ 128 ] Manon studied the homogeneous coordinate rings of the moduli stack M퐶;푝1,...,푝푛 (SL2) of quasi-parabolic bundles of rank 2. The Picard group 푛 of this stack is 푋 (퐵) × Z where 푋 (퐵) = Z휔1 is the character group of the Borel subgroup of SL2. Manon studies the algebraic properties of the various graded algebras one obtains by choosing a line bundle on the stack.
For us it will su ce to take 푛 ≤ 1, and hence we are interested in the moduli stacks M퐶 (SL2) and M퐶,푝 (SL2). Then we denote Ê 0 ⊗ℓ 푅퐶 (푏) := H (M퐶 (SL2), L(푏) ) ℓ ≥0 (140) Ê 0 ⊗ℓ 푅퐶,푝 (푎,푏) := H (M퐶,푝 (SL2), L(푎휔1,푏) ) ℓ ≥0 42 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY the graded algebras associated to the corresponding line bundles. Using the theory of conformal blocks Manon denes a at degeneration of these graded algebras to the boundary M푔,푛 \ M푔,푛 of the Deligne–Mumford compactication of the stack of pointed curves. This stack has a stratication by closed substacks, indexed by the weighted dual graphs of the stable curves. The most degenerate curves one gets (with only rational components) correspond to trivalent graphs of genus 푔 and 푛 half-edges with all weights being zero, corresponding to pair of pants decompositions we discussed earlier. The algebras obtained in this way are semigroup algebras associated to polytopes, hence we get toric varieties. To understand the relationship beteween these algebras and the varieties M퐶 (2, O퐶 ) and M퐶 (2, L), recall that by Drézet-Narasimhan [ 60 , Théorème B] the Picard groups of these varieties are generated by the class of the Theta divisor. By
Beauville and Laszlo [ 21 , Theorem 9.4] (and more generally the works [ 154 , 114 , 133 ]) we have identications Ê 0 ⊗ℓ 푅퐶 (2) H (M퐶 (2, O퐶 ), Θ ) ℓ ≥0 (141) Ê 0 ⊗ℓ 푅퐶,푝 (2, 2) H (M퐶 (2, L), Θ ). ℓ ≥0 Hence we get a toric degeneration of the moduli spaces we are interested in, for every choice of trivalent graph 훾 with at most one half-edge.
Graphs from stable curves. Let (퐶; 푝1, . . . , 푝푛) be a stable nodal curve with 푛 marked points. The dual graph associated to (퐶; 푝1, . . . , 푝푛) is the weighted graph 훾 = (푉, 퐸,푔) with 푛 half-edges dened as
• each vertex 푣푖 ∈ 푉 corresponds to an irreducible component 퐶푖 of 퐶, with weight 푔(푣푖 ) the genus of 퐶푖 ;
• the number of edges between 푣푖 and 푣 푗 is #퐶푖 ∩ 퐶푗 ;
• for each marked point 푝 in the component 퐶푖 we add a half-edge to the vertex 푣푖 . 퐶 퐸 푉 Í 푔 푣 The arithmetic genus of the curve is then # − # + 1 + 푣∈푉 ( ).
We will only consider nodal curves with marked points (퐶; 푝1, . . . , 푝푛), for which the dual graph 훾 (containing half-edges) is trivalent (and all weights are zero). These correspond to the zero-dimensional strata in the stratication of M푔,푛 \ M푔,푛 in terms of the type of the dual graph. Equivalently, all components 퐶푖 are required to be rational. We can rephrase
the denition of Manon’s polytope from [ 128 , Denition 1.1] as follows.
Denition 4.1.1. Let 훾 be a trivalent weighted graph of genus 푔 with at most one half- #퐸 edge. The polytope 푃훾 ⊆ R is the set of non-negative real weightings of the edges of 훾, such that • the weights of half-edges are precisely 2;
• for every vertex 푣 ∈ 푉 , if 푤1 (푣),푤2 (푣),푤3 (푣) are the weights of the edges incident to it, then
(142) 2 max{푤1 (푣),푤2 (푣),푤3 (푣)} ≤ 푤1 (푣) + 푤2 (푣) + 푤3 (푣) ≤ 4.
#퐸 We dene the lattice M훾 ⊂ 푃훾 using the lattice of integer points in R , using the condition that 푤1 (푣) + 푤2 (푣) + 푤3 (푣) ∈ 2Z for every 푣 ∈ 푉 .
The following theorem then describes the sought-after toric degeneration, which is a
special case of [ 128 , Theorem 1.2] and [ 129 , Theorem 1.3]. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 43
Theorem 4.1.2 (Manon). Let 퐶 be a smooth projective curve of genus 푔 ≥ 2. Let 훾 be a trivalent graph of genus 푔, with at most one half-edge. Then
• if there are no half-edges, then the homogeneous coordinate ring 푅퐶 (2) for M퐶 (2, O퐶 ) can be atly degenerated to the semigroup algebra associated to the polytope 푃훾 in the lattice M훾 ;
• if there is precisely one half-edge, then the homogeneous coordinate ring 푅퐶 (2, 2) for M퐶 (2, L) can be atly degenerated to the semigroup algebra associated to the polytope 푃훾 in the lattice M훾 .
푌 Denition 4.1.3. The toric variety from Theorem 4.1.2 will be denoted 푃훾 ,M훾 . Observe that 푃훾 is the moment polytope and M훾 is the character lattice.
We will rewrite the inequality in Denition 4.1.1 so that the origin becomes an internal point of the polytope and the polytope becomes full-dimensional in the character lattice of the toric variety. Setting 푢푖 (푣) := 푤푖 (푣) − 1 for the weight of the 푖th edge at the vertex 푣 we obtain the following lemma.
Lemma 4.1.4. Let 훾 be a trivalent graph of genus 푔 ≥ 2 with at most one half-edge. The polytope 푃훾 is dened by the inequalities • if 푣 is a vertex without half-edges, then
−푢1 (푣) − 푢2 (푣) − 푢3 (푣) ≥ −1
−푢1 (푣) + 푢2 (푣) + 푢3 (푣) ≥ −1 (143) 푢1 (푣) − 푢2 (푣) + 푢3 (푣) ≥ −1
푢1 (푣) + 푢2 (푣) − 푢3 (푣) ≥ −1;
• if 푣 is a vertex with a (necessarily unique) half-edge, then 푢1 = −푢2 for the weights of the other two edges.
Removing half-edges. To make the link to the setup of colored trivalent graphs from
Section 2.1 we will explain how to reduce trivalent graphs with one half-edge to colored trivalent graphs. A more general procedure can be constructed, but in this paper we are only interested in the case with one half-edge. Let 훾 = (푉, 퐸) be a trivalent graph of genus 푔 ≥ 2 with precisely one half-edge. Let 푣 ∈ 푉 be the vertex to which the half-edge 푒 is attached. Let 푒1 and 푒2 be the other edges in 푣, which 0 are incident to the vertices 푣1 and 푣2. Then we construct a colored trivalent graph (훾 , 푐) as follows: • the vertices 푉 0 are 푉 \{푣}; 0 • the edges 퐸 are (퐸 ∪ {푒1,2})\{푒, 푒1, 푒2}, where 푒1,2 is an edge between 푣1 and 푣2
• we color precisely one vertex from {푣1, 푣2}.
Graphically we have the situation from Figure 18 . 푣 푣 푣 푣 1 푒 푣 푒 2 1 푒1,2 2 1 2 ↦→ 푒
Figure 18. Removing a half-edge
With this procedure we dene a new polytope as follows. 44 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Denition 4.1.5. Let 훾 be a trivalent graph of genus 푔 ≥ 2 with one half-edge. Let (훾 0, 푐) be the colored trivalent graph obtained by removing the half-edge, where we choose #퐸0 to color the vertex 푣1. The polytope 푃훾0,푐 ⊆ R is dened by the same inequalities for 0 all 푣 ∈ 푉 \{푣1}, and in 푣1 they take on the inequalities
푢1 (푣1) + 푢2 (푣1) + 푢3 (푣1) ≥ −1
푢1 (푣1) − 푢2 (푣1) − 푢3 (푣1) ≥ −1 (144) −푢1 (푣1) + 푢2 (푣1) − 푢3 (푣1) ≥ −1
−푢1 (푣1) − 푢2 (푣1) + 푢3 (푣1) ≥ −1. Hence this polytope is dened in a subspace of R#퐸 given by the equality of the weights 푢1 (푣1) = 푢2 (푣2) where 푢1 and 푢2 refer to the edges 푒1 and 푒2 (recall that half-edges have a constant value assigned to them).
Vertices of the polytope 푃훾,푐 . Let (훾, 푐) be a colored trivalent graph of genus 푔 (with 푐 (푣)
푀 푁 퐻 no half-edges) We use e훾 and e훾 as in Section 2.1. Consider the half-spaces 푖 (for 푖 = 1, 2, 3, 4) dened by the inequalities 푐 (푣) (−1) (−푢1 − 푢2 − 푢3) ≥ −1, 푐 (푣) (−1) (−푢1 + 푢2 + 푢3) ≥ −1, (145) 푐 (푣) (−1) (푢1 − 푢2 + 푢3) ≥ −1, 푐 (푣) (−1) (푢1 + 푢2 − 푢3) ≥ −1, where 푢1,푢2,푢2 correspond to the three edges incident at the vertex 푣. Observe that the
inequalities ( 145 ) specialize to the inequalities ( 143 ) (resp. ( 144 )) if 푐(푣) = 0 (respectively if 푐(푣) = 1).
Denition 4.1.6. Let 훾, 푐 be a colored trivalent graph with no half-edges. Let 푉 be the set of vertices and 퐸 be the set of edges of 훾. Consider the polytope dened by the intersection of half-edges in 푀e훾 푃 Ù 퐻푐 (푣) 퐻푐 (푣) 퐻푐 (푣) 퐻푐 (푣) . (146) 훾,푐 = 1 ∩ 2 ∩ 3 ∩ 4 푣∈푉
푃◦ 푁 푣 푉 We can also describe the polar dual 훾,푐 ∈ e훾 as follows. For ∈ a vertex with color- ing 푐(푣), and 푒푖, 푒푗, 푒푘 the edges adjacent to 푣 we can consider the subgraph 훾푣, and dene the vector space
(147) 퐶푣,푐 := {푠 ∈ C1 (훾푣, F2) | d(푠) = 푐(푣)}.
For 푣 ∈ 푉 and 푠 ∈ 퐶푣,푐 we can then dene the point
푠푖 푠푗 푠푘 #퐸 (148) 푝(푣, 푠) := (0,..., 0, (−1) , 0,..., 0, (−1) , 0,..., 0, (−1) , 0,..., 0) ∈ 푁훾 ⊂ R with ±1 in positions 푖, 푗, 푘. The following standard lemma then makes the description as a convex hull of the polytope we are interested in explicit.
Lemma 4.1.7. Let (훾, 푐) be a trivalent graph of genus 푔 ≥ 2 with no half-edges. The 푃◦ 푃 푝 푣, 푠 푣 푉 polar dual 훾,푐 of the polytope 훾,푐 is the convex hull of the points ( ), where ∈ and 푠 ∈ 퐶푣,푐 .
Denition 4.1.8. Let 훾, 푐 be a colored trivalent graph of genus 푔 with no half-edges. The graph potential toric varieties 푋 (respectively 푋 ) are dened as the toric varieties 푃훾,푐 ,푀훾 푃훾,푐 ,푀e훾 with moment polytope 푃훾,푐 in the character lattice 푀훾 (respectively 푀e훾 ). GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 45
We can now compare Manon’s toric degeneration to the toric varieties obtained from the graph potentials, and show that they are the same.
Proposition 4.1.9. Let 훾 be a trivalent graph of genus 푔, with at most one half-edge. Let (훾 0, 푐) be the associated colored trivalent graph obtained from removing half-edges. There exists an isomorphism 푌 푋 . (149) 푃훾 ,M훾 푃훾0,푐 ,푀훾0
Proof. By [ 51 , Proposition 2.3.9], for any 휔 in the character lattice of a toric variety, the moment polytope 푃 and the polytope 휔 + 푃 has the same normal fan, and in turn gives isomorphic toric varieties.
푌 휔 ,..., Faust–Manon [64] translates the moment polytope of 푃훾 ,M훾 by = (2 2) ∈ M훾 to get the origin as the unique interior point. Observe that we have the identication of the lattices M훾 = 2푀훾0 . Hence, we have to scale down by a factor of 2. Starting from Manon’s equations for the polytope 푃훾 , we translate it by the vector (1, 1,..., 1) in 푀훾0 to get the equations of the re exive polytope 푃훾0,푐 . Thus the result follows.
In particular have we found another realization of Manon’s toric degeneration using an explicit full-dimensional re exive polytope.
Remark 4.1.10. The graph potential 푊e훾,푐 discussed in Section 2 descends to a regular 푊 푇 ∨ 푇 ∨ function 훾,푐 on the torus 훾 (respectively the toric variety e훾 ) of the toric variety of 푋 (respectively 푋 ). Moreover, by construction the Newton polytope of 푃훾,푐 ,푀훾 푃훾,푐 ,푀e훾 푊 푊 푃◦ graph potential e훾,푐 (also 훾,푐 ) is the polytope 훾,푐 .
4.2. Analyzing the singularities. Let (훾 0, 푐) be the colored trivalent graph constructed from a trivalent graph 훾 with at most one half-edge. We wish to study the singularities of 푌 푋 the toric variety 푃훾 ,M훾 푃훾0,푐 ,푀훾0 . Before we discuss the general picture, let us discuss what happens for the toric degener- ations of M퐶 (2, L) and M퐶 (2, O퐶 ) for 푔 = 2. There are two graphs, and as will be clear
from Theorem 4.2.3 and Remark 4.2.10 , the distinction between the behavior of the toric degenerations for M퐶 (2, L) depending on the choice of graph is representative of what happens in general.
Example 4.2.1. If 훾 is the Theta graph and there is one colored vertex, then the toric 5 variety we obtain is the singular intersection of quadrics in P given by 푥0푥5 − 푥1푥2 = 0 and 푥3푥4 − 푥1푥2 = 0 [ 143 ]. The polar dual of moment polytope is a cube with vertices (±1, ±1, ±1). This variety has 6 ordinary double points and hence has a small resolution. In particular, the variety
also has terminal singularities. This case is studied in Galkin’s thesis [ 72 ] and also by
Nishinou–Nohara–Ueda [ 143 , 144 ]. The results in this section and the next can be seen as a generalization from 푔 = 2 to 푔 ≥ 3. On the other hand, if 훾 is the dumbbell graph with one colored vertex, the toric variety is 2 5 given by the equations 푥 = 푧푤 and 푧푤 = 푢푣 in P [ 143 ]. The dual of the moment polytope of the toric variety is the convex hull of the points (150) (1, 2, 0), (1, −2, 0), (−1, 0, 0), (−1, 0, −2), (1, 0, 0), (−1, 0, 2). The polytope contains both the lattice points (−1, 0, −2) and (−1, 0, 2), but then it also contains (−1, 0, 0), and hence it cannot be terminal. 46 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
The distinction between the toric varieties obtained from these two graphs is explained by Theorem 4.2.3 . The case of M퐶 (2, O퐶 ) for 푔 = 2 is a bit pathological, as we have an 3 isomorphism M퐶 (2, O퐶 ) P [ 139 , Theorem 1]. Moreover the locus of semistable bundles up to S-equivalence is identied with the Kummer surface in P3.
Example 4.2.2. If 훾 is the Theta graph and there are no colored vertices, the toric variety is just P3. Clearly it is smooth and hence terminal. If 훾 is the dumbbell graph and there are no colored vertices. The dual of the moment polytope is the convex hull of the points (151) (1, 2, 0), (1, −2, 0), (−1, 0, 0), (1, 0, 2), (1, 0, 0), (1, 0, −2). The polytope contains the lattice points (1, 0, 0) which is exactly the half sum of the lattice points (1, 0, 2) and (1, 0, −2). Hence the toric variety is non-terminal.
Terminal singularities. Recall from [ 51 , Proposition 11.4.12(ii)] that for a Q-Gorenstein toric variety, having terminal singularities is equivalent to the only non-zero lattice points
in the fan polytope being its vertices. We can use this criterion, as by [ 64 , Theorem 12] 푌 the homogeneous coordinate rings of the toric degenerations 푃훾 ,M훾 are all arithmetically Gorenstein and in particular the toric degenerations themselves are Gorenstein.
Then the dichotomy observed for 푔 = 2 from Example 4.2.1 in the case of 1 colored vertex is illustrative for what happens for 푔 ≥ 3.
Theorem 4.2.3. Let (훾, 푐) be a colored trivalent graph of genus 푔 ≥ 3 with no half-edges. 푋 훾 The toric variety 푃훾,푐 ,푀훾 has terminal singularities if and only if has no separating edges. Before we prove this statement we give some preliminary lemmas. The following two lemmas work for an arbitrary integer 푛, even though we are mostly interested in the case 푛 = 3푔 − 3.
푛 푛 푆 3 푛 Lemma 4.2.4. Let ≥ 3 and consider Z . Consider the set of 2 3 points of the form ±푒푖 ± 푒푗 ± 푒푘 for 1 ≤ 푖, 푗, 푘 ≤ 푛 distinct. Let 푝 = 푠푖푒푖 + 푠푗푒푗 + 푠푘푒푘 be an element of 푆, then 푝 does not lie in the convex hull of 푆 \{푝}.
ℎ = 푠 푒∨ + 푠 푒∨ + 푠 푒∨ ℎ(푠 푒 + 푠 푒 + 푠 푒 ) = Proof. Consider the function 푖 푖 푗 푗 푘 푘 . We have that 푖 푖 푗 푗 푘 푘 3, and ℎ(푞) < 3 for 푞 ∈ 푆 \{푝}.
We will denote Π푛 the polytope given by the convex hull of the set 푆 from Lemma 4.2.4 .
푛 Lemma 4.2.5. With respect to the standard lattice Z the polytope Π푛 contains precisely the following lattice points:
• 0 ∈ Π푛
•± 푒푖 ∈ Π푛 for 1 ≤ 푖 ≤ 푛
•± 푒푖 ± 푒푗 ∈ Π푛 for 1 ≤ 푖, 푗 ≤ 푛 distinct
•± 푒푖 ± 푒푗 ± 푒푘 ∈ Π푛 for 1 ≤ 푖, 푗, 푘 ≤ 푛 distinct. Moreover,
• for all 1 ≤ 푖 ≤ 푛 the lattice points ±푒푖 lie in the convex hull of ±푒푖 ± 푒푗 ± 푒푘 for 1 ≤ 푗, 푘 ≤ 푛 distinct and 푖 ≠ 푗, 푘;
• for all 1 ≤ 푖, 푗 ≤ 푛 distinct the lattice points ±푒푖 ± 푒푗 lie in the convex hull of ±푒푖 ± 푒푗 ± 푒푘 for 1 ≤ 푘 ≤ 푛 such that 푘 ≠ 푖, 푗. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 47
ℎ Í푛 푒∨ 2 ℎ 푒 푒 푒 Proof. The function = 푖=1 ( 푖 ) is convex, and (± 푖 ± 푗 ± 푘 ) = 3, hence all these lattice points lie on a sphere of radius 1 and they cannot be contained in each other convex hulls. This function takes on the values 0, 1, 2, 3 on the lattice points listed in the statement of the lemma. To see that there are no other, observe that for two vectors 푢,푢0 whose coe cients in the standard basis belong to {−1, 0, 1}, we have that (푢0,푢) ≤ (푢,푢) in the standard bilinear 23 (푛) 푢 Í 3 훼 푣 푣 훼 pairing. If we write = 푖=1 푖 푖 where 푖 are the vertices of Π푛, such that 푖 ≥ 0 for 23 (푛) 푖 Í 3 훼 푢0 푒 푖 all and 푖=1 푖 = 1, we can apply this observation to = 푖 for all and we get 23 (푛) 23 (푛) ∑︁3 ∑︁3 (152) (푢,푢) = 훼푖 (푣푖,푢) ≤ 훼푖 (푢,푢) = (푢,푢) 푖=1 푖=1 hence either 훼푖 = 0 or (푣푖,푢) = (푢,푢). The nal claim is immediate.
We need one more lemma, giving a homological interpretation to the notion of a separating
edge, using the notation of Section 2.1 .
Lemma 4.2.6. Let 훾 be a trivalent graph. Let 푒 ∈ 퐸 be an edge, and consider the associated ∨ 1 ∨ indicator function 푒 ∈ C (훾, Z) = 푁e훾 . Then 푒 ∈ 푁훾 ⊆ 푁e훾 if and only if 푒 is separating.
Proof. Assume that 푒 is separating, i.e. by removing it we get two connected components which we will denote 훾 0 and 훾 00. Choose an orientation of the edges of 훾, and assign 1 to an incoming half-edge, and −1 to an outgoing half-edge. For 푣 ∈ 훾 0 we consider the sublattice 푁푣, and in there consider the sum ∑︁ (153) 푝푣,푠 (푣,표) ∈ 푁훾 푣∈훾0 where 표 denotes the orientation scheme we choose on the graph 훾, 푠(푣, 0) ∈ C1 (훾푣, F2) dened by 표 and 푝푣,푠 (푣,0) be as in ( 148 ). By construction the sum is, up to a sign, equal ∨ ∨ to 푒 , hence 푒 ∈ 푁훾 . ∨ Assume that 푒 ∈ 푁훾 , and for the sake of contradiction assume that 푒 is non-separating. Then there exists a cycle 푍 in 훾 (without repetition of vertices), dening an element of C1 (훾, Z) by assigning 1 to all edges of the graph. On the other hand, there exists a chain of inclusions
(154) 2푁e훾 ⊆ 푁훾 ⊆ 푁e훾 induced by the identity ±2푥푖 = (±푥푖 ± 푥 푗 ± 푥푘 ) + (±푥푖 + ∓푥 푗 ∓ 푥푘 ), where the 푥푖 ’s as in
Section 2 . These induce an isomorphism 1 (155) 푁e훾 /2푁e훾 C (훾, Z/2Z) which then induces an isomorphism 0 (156) 푁훾 /2푁e훾 d(C (훾, Z/2Z)) ∨ For 푒 to be in 푁훾 (and not just 푁e훾 ) it must evaluate to zero on cycles since 0 (157) 푁훾 /2푁e훾 d(C (훾, Z/2Z)). But by construction 푒∨ (푍) is odd. This contradiction shows that 푒∨ cannot be an element of 푁훾 . 48 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY ...
...
Figure 19. Ladder graph
Proof of Theorem 4.2.3 . Assume that 훾 has no separating edges. Then in particular it does 푃◦ 푃 not contain loops, and hence the set of vertices of the polar dual 훾,푐 of the polytope 훾,푐 푋 푒 푒 푒 푒 푁 #퐸 dening 푃훾,푐 ,푀훾 is a subset of {± 푖 ± 푗 ± 푘 }, where 푖 are standard vectors in e훾 Z .
By Lemma 4.2.4 we have that all the points 푝(푣, 푠) given by ( 148 ) are vertices. It remains to show that there are no other lattice points on the boundary of this polytope.
For this, by Lemma 4.2.5 it su ces to show that no points of the form ±푒푖 and ±푒푖 ± 푒푗 is a 푃◦ 훾 lattice point in the convex hull of the vertices of 훾,푐 . Since the graph has no separating edges, it follows that from Lemma 4.2.6 that ±푒푖 cannot be a lattice point in 푁훾 .
Now say for example that 푒푖 + 푒푗 would be a lattice point of the polytope, and let 푣 be a vertex adjacent to 푒푖 and 푒푗 . Then we can write 1 (158) 푒 + 푒 = (푒 + 푒 + 푒 ) + (푒 + 푒 − 푒 ) 푖 푗 2 푖 푗 푘 푖 푗 푘 where 푒푘 is the third edge adjacent to 푣. However, the only way in which we can do this if 푒푖 , 푒푗 and 푒푘 have the same end points, but then we obtain a subgraph of genus 2 in 훾 which is a contradiction.
The converse follows from Lemma 4.2.6 and ( 153 ).
As an application of Theorem 4.2.3 , we can give an alternative proof of the following result
due to Kiem–Li [ 105 , Corollary 5.4] (albeit only for a generic curve, rather than for every curve). Their proof is based on understanding the discrepancy between the canonical bundle of M퐶 (2, O퐶 ) and its Kirwan desingularization. In our setup it rather follows from Kawamata’s inversion of adjunction for terminal singularities.
Corollary 4.2.7 (Kiem–Li). Let 퐶 be a generic smooth projective curve of genus 푔 ≥ 2. Then M퐶 (2, O퐶 ) has at most terminal singularities.
Proof. By Kawamata [ 104 , Theorem 1.5] we know that having at most terminal singularities in the central ber implies that a general ber has at most terminal singularities. Manon’s construction provides us with a family of varieties whose special ber is toric and whose general bers are of the form M퐶 (2, O퐶 ). By Theorem 4.2.3 it su ces to choose a trivalent graph 훾 of genus 푔 without separating edges, so that the toric special ber has at most terminal singularities. For this we can consider for instance the ladder graph as in
Figure 19 . Then in the family given by the toric degeneration the central ber has at most terminal singularities, hence so does M퐶 (2, O퐶 ) in a Zariski open neighbourhood of the special ber.
Small resolutions. Recall that a resolution of singularities for a variety 푋 is said to be small if there is no exceptional divisor. In particular the resolution is crepant. The existence of a small resolution is a hard question in general and it is a delicate condition on the singularities of the variety 푋. We restrict ourselves to a special class of Gorenstein terminal toric varieties. Recall that (torically) resolving the singularities of a toric variety 푋 amounts to subdividing each cone in the fan of the toric variety into a union of cones generated by a basis of the lattice. A GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 49 toric resolution is small if we can nd such a subdivision without adding new rays. This is in general a non-trivial combinatorial problem. Toric degenerations admitting a small resolution of singularities are useful in the study
of mirror symmetry, as e.g. showcased in [ 17 , 15 , 145 ]. We brie y discuss this and refer
to [ 145 , 170 ] for more details. We are mostly interested in the question of computing the quantum period of the Fano variety M퐶 (2, L) which are certain enumerative invariants
(see Section 5 ). As explained in the introduction mirror symmetry predicts the existence of a Laurent polynomial whose periods compute the quantum periods.
Bondal–Galkin [ 36 ] observed that the work of Nishinou–Nohara–Ueda [ 145 ] can be inter-
preted as a candidate for this mirror Laurent polynomial. In [ 145 ], the authors construct a Laurent polynomial for any smooth Fano variety 퐹 by counting holomorphic discs with a xed Lagrangian boundary in the Fano variety 퐹. They construct their Laurent polynomial by rst constructing a toric degeneration 푋 of the Fano variety 퐹 and consider
the associated toric complete integrable system (see Denition 5.1.1 ). Using symplectic parallel transport they construct an integrable system on the Fano variety 퐹. A key feature of a small resolution of toric varieties is the fact that given a torus-invariant curve in a toric variety 푋, there exists a unique torus invariant curve in the small res- olution 푋e that maps isomorphically to the original curve. In turn this guarantees that the second homotopy groups of the two integrable systems they construct agree via the
symplectic parallel transport (see [ 145 , Section 9]). The existence of a small resolution guarantees that torically transverse (in the sense of Nishinou–Siebert [ 146 , Denition 4.9]) discs in the resolution 푋e project to torically
transverse discs in 푋. These results are crucially used in the works of [ 145 , 144 ] to compare their Laurent polynomial that counts holomorphic discs to a natural polynomial associated to 푋.
Manon has constructed a toric degeneration of M퐶 (2, L) and we gave conditions on the graph 훾 for the toric degeneration to have terminal singularities. To put us in the
perspective of [ 145 ], we further need to analyze whether Manon’s toric varieties admit a small resolution of singularities. In particular, we need to put further restrictions on the class of trivalent graph of genus 푔. We consider the following class of trivalent graphs.
Denition 4.2.8. Let 훾 be a trivalent graph. We say it is maximally connected if one needs to remove at least 3 edges to make it disconnected.
Maximally connected trivalent graphs 훾 have a very interesting positivity property. Let 퐶0
be a nodal curve whose dual graph is a maximally connected trivalent graph, then by [ 19 , Proposition 2.5] the canonical bundle of the curve 퐶0 is very ample. We do not precisely know what role the maximally connectedness play in the combinatorial description of the 푋 cones of toric varieties 푃훾,푐 ,푀훾 , but motivated by the genus two case and experimental evidence, we conjecture
Conjecture 4.2.9. Let (훾, 푐) be a maximally connected trivalent graph. Then the graph 푋 potential toric variety 푃훾,푐 ,푀훾 admit a small resolution of singularities.
푋 Since 푃훾,푐 ,푀훾 is toric, if we consider toric resolutions in Conjecture 4.2.9, we want to nd 푋 a renement of the fan of 푃훾,푐,푀훾 into a simplicial fan such that the resulting toric variety is smooth.
Remark 4.2.10. For 푔 = 2 this follows from the discussion in Example 4.2.1 for the Theta graph, which is the unique maximally connected trivalent graph of genus 2. 50 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
(a) Tetrahedron graph
(b) Bipartite graph of 푔 = 4
Figure 20. Graphs giving small toric resolutions
Using computer algebra we have checked Conjecture 4.2.9 also in genus 3 for the tetrahe-
dron graph (depicted in Figure 20(a) ), and in genus 4 for the bipartite graph (see Figure 20(b) ) with an odd number of vertices colored, by explicitly constructing a simplicial subdivision of the fan polytope. Moreover, in the genus three, four and ve cases, we can nd non- maximally connected graphs with no separating edges such that the corresponding toric varieties do not admit a small resolution of singularities.
5. Enumerative mirror symmetry and symplectic geometry The goal of this section is to relate graph potentials (that we have associated to every colored trivalent graph) to the enumerative and symplectic geometry of the Fano vari- ety M퐶 (2, L). On the symplectic level we do this by considering the monotone Lagrangian
ber 퐿훾 of Nishinou–Nohara–Ueda’s integrable system [ 145 , 144 ] associated with Manon’s toric degeneration. This allows us to identify the graph potential 푊e훾,푐 (at least for some graphs) with the Floer potential of 퐿훾 . On the enumerative level we will show that the period sequence of the graph potential equals the quantum period of M퐶 (2, L).
The proof that enumerative mirror symmetry holds for graph potentials and M퐶 (2, L) fol- lows the approach of Bondal–Galkin and Nishinou–Nohara–Ueda. This gives the following outline for this section.
(1) We explain how to obtain an integrable system Φ훾 : M퐶 (2, L) → Δ훾,푐 for every
(colored) graph 훾 in Section 5.1 .
−1
퐿 , (2) In Proposition 5.3.6 we show that the Lagrangian torus 훾 := Φ훾 (0) ⊂ M퐶 (2 L) is monotone.
Now assume that the respective toric variety has a small resolution of singularities, as discussed in the previous section.
(3) In Corollary 5.2.3 we identify the Floer potential of the torus 퐿훾 with the graph
potential from Section 2 .
This is related to the fact that the enumerative invariants counting holomorphic spheres (discs) passing through a point (with boundary on 퐿훾 ) can be computed tropically: the respective curves can be degenerated to the regular locus of the singular toric variety and lifted to the complement of the exceptional locus on the small resolution. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 51
(4) This allows us to interpret the constant term of the 푑th power of the graph potential as the number of maps from tropical trees to Δ훾,푐 in tropical descendant Gromov–
Witten invariant enumeration [ 158 ], which in turn is equated with the descendant
Gromov–Witten invariant of M퐶 (2, L). This is the content of Section 5.4 . (5) In particular, this proves a relation between closed and open Gromov–Witten invariants found and proved in the context of small toric degenerations indepen- dently by Bondal–Galkin and Nishinou–Nohara–Ueda. Recently Tonkonog proved a generalization of this relation for the case of an abstract monotone Lagrangian
torus [ 170 ].
(6) From 4 we deduce the regularity of regularized quantum periods and determine
their radius of convergence in Proposition 5.5.1 . Finally,
(7) In Section 5.6 we argue that the birational transformations from ( 1 ) should corre- spond to a new type of Lagrangian mutations.
The construction of the toric degenerations from Section 4 will be used from the very beginning, but most of the section makes sense for other degenerations with similar
properties. Step (3) shows the symplectic (and tropical) origin of graph potentials. Steps (3)
and (4) are subject to Conjecture 4.2.9 . Steps (3) and (4) imply (5) , and also steps (2) , (3) and
(5) imply (4) . In Appendix A we will give further geometric evidence, and the homological
mirror symmetry picture from Section 6.4 gives further corroboration for the relationship between graph potentials and M퐶 (2, L). 5.1. Integrable systems. Let us recall the notion of an integrable system on a symplectic manifold (푀, 휔) of real dimension 2푁 .
Denition 5.1.1. An integrable system on (푀, 휔) is a collection of 푁 functionally inde- ∞ pendent real-valued 퐶 functions {퐻1, . . . , 퐻푁 } on the manifold 푀 which are pairwise Poisson-commutative, i.e. {퐻푖, 퐻푗 } = 0 for all 푖, 푗 = 1, . . . , 푁 , where {−, −} is the Poisson bracket induced by the symplectic form 휔. We will denote it as Φ: 푀 → R푁 .
By denition, an integrable system induces a Hamiltonian R푁 -action on 푀. By the Arnold– Liouville theorem we have that any regular, compact, connected orbit of this action gives a Lagrangian torus in 푀. We will work with integrable systems on singular toric varieties. In that set-up we want the real-valued 퐶∞ functions to Poisson commute on the smooth locus. Natural examples of integrable systems are found in the theory of toric varieties. Here × 푁 we let 푋0 be a toric variety of complex dimension 푁 , and consider the torus action (C ) on 푋0. The moment map for the torus action with respect to any torus-invariant Kähler form gives an integrable system.
Integrable systems from toric degenerations. Nishinou–Nohara–Ueda [ 145 ] and Harada–
Kaveh [ 87 ] construct integrable systems from a toric degeneration of a smooth projective variety satisfying some additional conditions. We brie y recall their construction. Let 휋 : 픛 → A1 be a at family of complex projective varieties of dimension 푁 , such −1 −1 that 푋0 := 휋 (0) is a toric variety, whilst the general ber 푋푡 := 휋 (푡) for 푡 ≠ 0 is smooth. Assume moreover that
• the singular locus of the total space 픛 is contained in the singular locus of 푋0; • the regular part of 픛 has a Kähler form which restrict to a torus-invariant Kähler 푋 reg 푋 form on the regular part 0 of 0. 52 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
The choice of a piecewise smooth curve 퐼 : [0, 1] → A1 from 0 to 1 in A1 gives by symplectic 퐼 퐼 푋 reg 푋 reg 푋 reg parallel transport along a symplectomorphism e: 0 → 1 , where 1 is an open subset of 푋1. 퐼 푋 reg 푁 Using the symplectomorphismewe can transport the toric integrable system Φ0 : 0 → R to the integrable system 퐼 −1 푋 reg 푁 (159) Φ := Φ0 ◦ e : 1 → R ∞ on 푋1. Observe that the same ow gives a 퐶 map from the compact manifold 푋1 to the compact space 푋0, however over the preimage of the singularities, one cannot extend the torus action. A suitable choice of the form 휔 and path 퐼 makes the map real-analytic,
cfr. [ 115 ].
Let 푃 denote the moment polytope of the toric variety 푋0. Let
(160) ℓ푖 (푢) := h푣푖,푢i − 휏푖 for 푖 ∈ {1, . . . ,푚} denote the a ne equations dening the polytope, i.e. 푁 (161) 푃 := Φ0 (푋0) = {푢 ∈ R | ∀푖 = 1, . . . ,푚 : ℓ푖 (푢) ≥ 0 }. ◦ The convex hull of the −푣푖 /휏푖 ’s as in ( 160 ) is the polar dual 푃 of the moment polytope 푃.
In Section 4 we have discussed the geometry of a toric degeneration of M퐶 (2, L). We will use the following notation.
Denition 5.1.2. Let (훾, 푐) be a trivalent graph with one colored vertex. Then 3푔−3 (162) Φ훾,푐 : M퐶 (2, L) → R is the integrable system obtained by applying the Nishinou–Nohara–Ueda construction to Manon’s toric degeneration. Its image will be denoted Δ훾,푐 .
5.2. Discs of Maslov index two and Floer potentials. We will now introduce the Floer potential, and recall some of the necessary ingredients. Let 퐿 be a Lagrangian submanifold 퐿 in a symplectic manifold (푀, 휔), and let 퐷 be a two-dimensional disc 퐷 with boundary a circle 휕퐷 ' 푆1. For a continuous map 푓 : (퐷, 휕퐷) → (푀, 퐿) its restriction 휕푓 : 휕퐷 → 퐿 is a loop on 퐿. Denote the class of 푓 in the relative homotopy or homology group as [퐷, 휕퐷] = [퐷] = [푓 ] ∈ 휋2 (푀, 퐿) → H2 (퐿, Z) and the boundary class as [휕퐷] = [휕푓 ] = 휕[푓 ] ∈ 휋1 (퐿)' H1 (퐿, Z).
dimR 푀 The bundle LGr(T푀) of Lagrangian Grassmannians LGr(T푚푀)' LGr(R ) =: LGr trivializes when pulled back to the contractible space 퐷. Via this trivialization the tan- ∗ gential section 푝 → T푓 (푝)퐿 ∈ LGr(T푀푓 (푝) ) of (휕푓 ) LGr(T푀)' 휕퐷 × LGr has a class in 2 휋1 (LGr)' H1 (LGr)' Z known as the Maslov index 휇(퐷) . For orientable 퐿 the lifting to the orientable Lagrangian Grassmannian LGr+ guarantees that the Maslov index takes + even values 휇(퐷) ∈ 휋1 (LGr ) = 2Z ⊂ Z = 휋1 (LGr). The next ingredient is an almost complex structure 퐽. We say that it is 휔-tame if 휔 (푣, 퐽푣) > 0 for all non-zero tangent vectors 푣. For xed 휔 the space of 휔-tame almost complex struc- tures is contractible, hence the Chern classes of 퐽 are symplectic invariants. In this situation the Maslov index of a 퐽-holomorphic map 푓 equals twice its rst Chern number. In partiucu- lar a 퐽-holomorphic disc with boundary on 퐿 that intersects an e ective anti-canonical di- visor transversally in a unique point has Maslov index 2. More generally, one can show that the Maslov index controls the dimension of the (virtual) fundamental cycle on the (Kuran- ishi) moduli space of 퐽-holomorphic maps 푓 : (퐶 3 푃1, . . . , 푃푛; 휕퐶 3 푄1, . . . , 푄푚) → (푀, 퐿)
2This notation hides the mapping 푓 from a pair (퐷, 휕퐷) to (푀, 퐿). We will further call such maps discs of Maslov index 휇. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 53 from a Riemann surface 퐶 with 푛 marked points in its interior and 푚 marked points on the boundary 휕퐶.
Denition 5.2.1. For 퐽 xed the Floer potential of 퐿 (denoted 픓픒 or 푚0 (퐿) or 푊퐿 ) is the generating function for the 퐽-holomorphic discs (퐷, 휕퐷, 푝) ⊂ (푀, 퐿, 푃) of Maslov index 2, whose boundary passes through a specied point 푃 on 퐿 (one can also x a interior point and ask it to lie on a xed e ective anti-canonical divisor).
In general, it is dened on the moduli space of weakly bounded cochains as ∫ ∑︁ 휔 (163) 픓픒(푥) := exp(h[휕퐷], 푥i)푇 퐷 (퐷 ⊃휕퐷 3푝) ⊂(푀 ⊃퐿3푃) 휇 (퐷)=2 where • ∫ 휔 (푀, 퐿 ) → ∗ : H2 ; Z R is the symplectic area;
• 휇 : H2 (푋, 퐿; Z) → Z is the Maslov index; 1 • 푥 ∈ H (퐿, Λ0) is a cohomology class valued in Novikov ring Λ0;
Í+∞ 푎 푇 휆푖 푎 , 휆 , 휆 • Λ := 푖=0 푖 | 푖 ∈ Q 푖 ∈ R lim푖→+∞ 푖 = +∞ and Λ0 is its subring of ele- ments with non-negative valuation ∑︁ 휆푖 픳 푎푖푇 := min {휆푖 | 푎푖 ≠ 0} ∈ R
∫ 휔 2 The Novikov ring is used to deal with symplectic areas푇 Σ when the class [휔] ∈ H (푀, R) of the symplectic form is not an integer, or when the surface Σ is not closed. However under assumptions of integrality and monotonicity, (which will be introduced and established later) the generating functions can be rewritten as Taylor series. For similar results on monotone toric manifolds and the passage between critical points in archimedean and
Novikov contexts, See the work of Ritter [ 160 , Section 7.9] and Judd–Rietsch [ 99 ]. For generic 휔-tame 퐽 the potential gives a well-dened element in the Novikov ring thanks to the Gromov compactness theorem. The symplectic parallel transport for integrable systems is used to show the following
[ 144 , Theorem 1]
Theorem 5.2.2 (Nishinou–Nohara–Ueda). Consider a toric degeneration 픛 → A1 as in
Section 5.1 . Assume that 푋0 is Fano, and admits a small resolution of singularities. Then for any 푢 ∈ 푃int ⊆ R푁 in the interior of the moment polytope, the Floer potential of the Lagrangian torus ber 퐿(푢) := Φ−1 (푢) is given by 푚 ∑︁ ℓ푖 (푢) (164) 픓픒(푥) = exp(h푣푖, 푥i)푇 , 푖=1 ◦ where 푣푖 are the generators of the polar dual 푃 and ℓ푖 are the equations dening the facets of the moment polytope 푃 and valuations of 푥 belong to the moment polytope 푃.
In particular it says that if the degenerating toric Fano variety admits a small resolution of singularities, then the holomorphic discs with Lagrangian boundary that contribute to the Floer potential of the smoothing are transversal to toric strata. In contrast, for degenerations with transversal A1 singularities there is an additional contribution from a disc that intersects it, in our situation this can be observed for e.g. on graphs with loops.
Applying Theorem 5.2.2 to our situation we obtain the following corollary. 54 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Corollary 5.2.3. Let (훾, 푐) be a trivalent graph with one colored vertex, such that the 푋 corresponding toric variety 푃훾,푐 ,푀훾 has a small resolution of singularities. Then the following are equal:
• the graph potential 푊훾,푐 ,
• the Floer potential of the Lagrangian torus 퐿훾 ⊂ M퐶 (2, L); 퐿 −1 , where 훾 = Φ훾 (0) is a ber of a Nishinou–Nohara–Ueda integrable system Φ훾 : M퐶 (2 L) → Δ훾,푐 , 푌 associated with Manon’s toric degeneration of M퐶 (2 L) to 푃훾 ,M훾 .
By Conjecture 4.2.9 it is expected that we can always nd such a graph. In genus 푔 = 2, 3, 4, for maximally connected graphs the corresponding toric variety has a small resolution. Let us conclude this section by giving two remarks on the case of even degree.
Remark 5.2.4. In the case of M퐶 (2, O퐶 ) symplectic transport gives a symplectomorphism from the non-singular locus of the toric ber to an open subset in non-singular locus of M퐶 (2, O퐶 ). The corresponding Nishinou–Nohara–Ueda integrable system seems to
be quite di erent from Goldman–Je rey–Weitsman integrable system from [ 96 ], despite sharing the base Δ훾 . In the former case the preimage of the interior of the polytope is the cotangent bundle of the torus and in the latter it is singular.
Remark 5.2.5. There are alternative points of view on M퐶 (2, O퐶 ), namely as a Poisson
variety (by the work of Huebschmann, see e.g. [ 95 ]) or as a stack, which could take care of the singularities present in M퐶 (2, O퐶 ) for 푔 ≥ 3, and the discrepancy between the period 3 sequence for 푔 = 2 and M퐶 (2, O퐶 ) P . • To x the discrepancy for 푔 = 2, it su ces to rescale the graph potential obtained
in ( 36 ) by 1/2 to get the correct period sequence.
• For 푔 ≥ 3 and a connected graph 훾 without separating edges the polynomial 푊훾 is supported in the vertices of its Newton polytope. There are 8푔 − 8 non-vanishing coe cients, each equal to one, which are in bijection with facets of Δ훾 and with irreducible torus-invariant Weil divisors on 푋0. A ber of a moment map bounds 8푔 − 8 holomorphic discs, each transversally intersecting exactly one of toric divisors, and if 푋0 has a small resolution of singularities this shows that there are no other holomorphic discs of Maslov index 2.
Hence the graph potential 푊훾,0 re ects some aspects of the symplectic geometry of the singular variety M퐶 (2, O퐶 ). Since Fukaya categories of singular varieties are not so well established, it seems reasonable to start with categories of matrix factorizations of the graph potentials 푊훾,0 instead of the not so well-dened categories Fuk M퐶 (2, O퐶 ), in order to gain further insight into their (conjecturally common) properties.
5.3. Monotone Lagrangian tori. Now to every trivalent graph 훾 of genus 푔 without leaves we will associate a monotone Lagrangian torus 퐿훾 on the moduli space M퐶 (2, L). In
Remark 5.3.7 we will comment on how the monotone Lagrangian tori for di erent choices of 훾 might be compared to each other. Recall the following denition.
Denition 5.3.1. Let 퐿 be a Lagrangian submanifold of a symplectic manifold (푀, 휔). We say that 퐿 is (positive) monotone if the symplectic area H2 (푀, 퐿; Z) → R and the Maslov index H2 (푀, 퐿; Z) → Z are (positively) proportional to each other.
We now discuss a general result which is implicit in the works of Nishinou–Nohara–Ueda
and Ritter [ 145 , 144 , 160 ]. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 55
Lemma 5.3.2. Let 푋1 be a Fano manifold admitting a degeneration to a normal toric variety 푋0. For any interior point푢 ∈ 푃, let 퐿 = 퐿(푢) be a Lagrangian torus in 푋1 constructed via symplectic parallel transport. Identify H1 (퐿, Z) with the cocharacter lattice 푁 of the ◦ toric variety 푋0. Assume that the vertices of 푃 generate 푁 . Then there is a isomorphism
(165) H2 (푋1, 퐿; Z) 푁 ⊕ H2 (푋1, Z).
Proof. Consider the long exact sequence for the homology of the pair (푋1, 퐿),
(166) H2 (퐿, Z) → H2 (푋1, Z) → H2 (푋1, 퐿; Z) → H1 (퐿, Z) → H1 (푋1, Z).
The last term vanishes, as 푋1 is simply connected. Moreover by Lemma 5.3.3 , the image of
H2 (퐿, Z) in H2 (푋1, Z) is zero. Hence we obtain a short exact sequence from ( 166 ), which allows us to identify
H2 (푋1, 퐿; Z) H1 (퐿; Z) ⊕ H2 (푋1; Z) (167) 푁 ⊕ H2 (푋1; Z).
The following lemma is used in the proof of Lemma 5.3.2 , however it does not require the assumption that 푋1 is projective.
Lemma 5.3.3. In the situation of in Lemma 5.3.2 the map H2 (퐿, Z) → H2 (푋1, Z) is zero.
푃reg 푋 reg Proof. Let be the image of the moment map from the smooth part 0 of the toric 푋 푋 1-cosk 푋 variety 0. Let 0 is the toric variety whose fan is the 1-skeleton of the fan of 0 and let 푃1-cosk be the image of the moment map. In particular 푃1-cosk is a union of open faces of codimension at most one. Let 푃int be the interior of 푃. 푋 int 푃int 푋 int 푋 int 푋 int Let 0 denote the inverse image of . Similarly dene 1 . Both 1 and 0 contain a torus isomorphic to 퐿. Consider the following diagram.
푋1 푋0 푃
푋 reg 푋 reg 푃reg 1 0
(168) 푋 1-cosk 푋 1-cosk 푃1-cosk 1 0
푋 int 푋 int 푃int 1 0
퐿 = 퐿(푢) 퐿 = 퐿(푢) 푢
By construction 푃reg = 푃\푃sing, where 푃sing is the singular locus of 푃. To show that the image of H2 (퐿, Z) in H2 (푋1, Z) is zero it would su ce to show that the image of H2 (퐿, Z) 푋 reg, is zero in H2 ( 0 Z) while going up along the middle column of the diagram. ◦ 1 Now let 푉 = {푣1, . . . , 푣푚 } be the vertices of 푃 . Each 푣푖 corresponds to a unit sphere 푆 in 퐿. 푇 Ð R 푣 퐿 ∪ Ð 퐷 The toric variety 푉 whose fan is 푣푖 ∈푉 ≥0 푖 is homotopy equivalent to 푣푖 ∈푉 푣푖 , 56 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
퐷 푣 where 푣푖 is a disc whose boundary is the one-cycle corresponding to 푖 . We have the following diagram: 푆1 퐿 (169) 퐷 푇 . 푣푖 푉
Hence H (푇 , Z) = 퐿/É Z푣 . But by assumption 푉 generates 퐿. Hence H (푇 , Z) is 1 푉 푣푖 ∈푉 푖 1 푉 zero. Moreover the natural embedding of 푆1 푆1 푝 퐿 퐷 ×푛−1 (170) × × ↩→ ↩→ 푣푖 × C 푆1 푆1, 퐷 × 푛−1 , induces the zero map between H2 ( × Z) → H2 ( 푣푖 × (C ) Z). Thus we are done.
Remark 5.3.4. The toric variety푇푉 and its homotopy retract to the union of the Lagrangian torus 퐿 with attached holomorphic discs 퐷푖 is borrowed from Galkin–Mikhalkin [ 78 ], where it is used to compute the (non-abelian) relative second homotopy group 휋2 (푇푉 , 퐿) responsible for a “quantization” of the Floer potential of surfaces.
The following proposition gives us a useful criterion to test whether the Lagrangian
torus 퐿(푢) in Lemma 5.3.2 is monotone.
Proposition 5.3.5. Let 푢 ∈ 푃 be an interior point as in Lemma 5.3.2 such that that
ℓ푖 (푢) = ℓ > 0, where ℓ푖 are the length functions ( 160 ) dening the polytope 푃. Further assume that (푋1, 휔) is monotone. Then the Lagrangian torus 퐿(푢) is monotone.
Proof. Consider the (holomorphic) discs 퐷푖 of Maslov index 2 enumerated by the vertices 푣푖 ◦
푃 of the polytope 훾,푐 . By [145, Theorem 10.1] or [45, Theorem 8.1] they have symplectic area ∫ (171) 휔 = ℓ푖 (푢) = ℓ > 0. 퐷푖 Non-negative linear combinations of the classes of their boundaries [휕퐷푖 ] ∈ H1 (퐿(푢), Z) = 푁 generate a nite index sublattice. Hence any element 훽 ∈ H2 (푋1, 퐿(푢); Q) can be presented Í as 훽 = 훽1 + 푏푖 [퐷푖 ], a linear combination of [퐷푖 ] and some 훽1 ∈ H2 (푋1, Q). Moreover Í Í there exists 푎푖 ≥ 0 such that 푎푖 > 0 such that 푎푖 [퐷푖 ] ∈ H2 (푋1, Z). Then we have that ∫ ∑︁ ∫ (172) 휔 = 푏푖 ℓ + 휔, 훽 훽1 and ∫ ∑︁ (173) 휔 = 푎푖 ℓ > 0 Í 푎푖 [퐷푖 ] Í Í where 휇( 푎푖 [퐷푖 ]) = 2 푎푖 > 0. Monotonicity of (푋1, 휔) implies that ∫ 휔 ∫ Í 푎 [퐷 ] (174) 휔 = 휇(훽 ) 푖 푖 = 휇(훽 )/2 · ℓ, 1 휇(Í 푎 [퐷 ]) 1 훽1 푖 푖 ∫ 휔 = (ℓ/ )휇(훽) hence 훽 2 .
Now we apply Proposition 5.3.5 to produce a Lagrangian torus in M퐶 (2, L), using Deni-
tion 5.1.2 .
퐿 퐿 −1 , Proposition 5.3.6. The Lagrangian torus = (0) := Φ훾,푐 (0) ⊂ M퐶 (2 L) is monotone. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 57
Proof. For each vertex 푣 of the graph (훾, 푐) and 푠 ∈ 퐶푣,푐 ( 147 ), consider the equations ℓ푣,푠 ◦
푃 ℓ from (144) used in Section 4.1 to describe the polytope 훾,푐 . Rewrite 푣,푠 as in (160)
(175) h푝(푣, 푠),푢i − 휏푣,푠
where 푝(푣, 푠) is as in ( 148 ) and 휏푣,푠 = −1. Setting 푢 = 0 we obtain that ℓ푣,푠 (0) = 1. Observe 푝 푣, 푠 푃◦ that the lattice points ( ) of the polytope 훾,푐 , are non-zero, that the vertices form a subset by Theorem 4.2.3 , and furthermore span the lattice 푁훾 . The statement then follows
directly from Proposition 5.3.5 .
Remark 5.3.7. We expect that there exists a combinatorial non-abelian Torelli theorem for the toric degenerations. Combined with the computation of the Floer potential for graphs such that the respective toric variety has a small resolution, this would imply that all these tori are not Hamiltonian-isotopic to each other. We leave this for future work.
5.4. Descendant invariants. Let 푋 be a Fano variety. We will denote by 푋0,1,훽 the moduli space of stable maps 푓 : (퐶; 푥) → 푋 where 퐶 is a rational curve with a marked point 푥 ∈ 퐶 such that [푓 (퐶)] = 훽 ∈ H2 (푋, Z) is a xed homology class. It has a virtual fundamental ∫ (푋) − + 푋 class in degree 훽 c1 2 dim , and comes with the evaluation and the forgetful morphisms
ev: 푋0,1,훽 → 푋 = 푋0,1,0, (176) 휋 : 푋0,1,훽 → 푋0,0,훽 . Denote by 휓 the rst Chern class of the universal cotangent line bundle, which is also the relative dualizing sheaf 휔휋 . We can use it to dene the descendant Gromov–Witten invariants for 훽 ∈ H2 (푋, Z). These count count rational curves of degree 푑 passing through a xed generic point with some kind of tangency condition, and are dened as ∫ 푝 휓푑−2 푋 ∗ 휓푑−2, 훽 := h[pt]· i0,1,훽 = ev ([pt]) ∪ vir [푋0,1,훽 ] (177) ∑︁ 푝푑 (푋) := 푝훽, 푑 ≥ 2, 휇 (훽)=2푑 [ ] ∫ [ ] = where pt is the class of the point, i.e. a homogeneous class such that 푋 pt 1. We moreover set
(178) 푝0 (푋) := 1, 푝1 (푋) := 0, and
(179) 푐푑 := 푑! · 푝푑,
3 Remark 5.4.1. Under the assumption that 푋 has a smooth anti-canonical divisor 퐷 , 1 Mandel [ 125 , Theorem 1.2] shows that 푐푑 is the naive number of maps from P to 푋 passing through a xed generic point on 푋 such that the preimage of the divisor 퐷 is a 1 xed generic 푑-tuple of points on P . In particular the number 푐푑 is a non-negative integer.
The next result, subject to Conjecture 4.2.9 , demonstrates how the combinatorics of graph potentials controls the enumerative geometry of M퐶 (2, L). The equality in this corollary corresponds to the original condition of enumerative mirror symmetry for Fano
manifolds proposed by Eguchi–Hori–Xiong. In our setup it follows from Tonkonog’s [ 170 ,
Theorem 1.1], using the monotone Lagrangian torus constructed in Proposition 5.3.6 .
3Which by the Bertini theorem holds for the varieties we consider in this paper, but fails for some other Fano varieties such as the squares of degree one del Pezzo surfaces. 58 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Corollary 5.4.2. Let (훾, 푐) be a colored trivalent graph of genus 푔 ≥ 2 with one colored
vertex. Let 푑 ≥ 2. Assuming Conjecture 4.2.9 we have the equality 푝 , 푊 푑 (180) 푑 (M퐶 (2 L)) = [ 훾,푐 ]푧0 of the Gromov–Witten invariant 푐푑 and the constant term of 푑th power of the Laurent polynomial 푊훾,푐 .
5.5. Quantum periods. Let us brie y recall the notion of quantum periods for Fano
varieties, as discussed in Coates–Corti–Galkin–Kasprzyk and Galkin–Iritani [ 47 , 76 ]. The (unregularized and regularized) quantum periods are the generating power series for descendant Gromov–Witten invariants 푝푑 and 푐푑 = 푑! · 푝푑 dened in Equation ( 177 ): +∞ ∑︁ 푑 G푋 (푡) := 1 + 푝푑 (푋)푡 , 푑=2 (181) +∞ ∑︁ 1 ∫ ∞ (휅) := + 푐 (푋)휅푑 = (푡) (푡/휅) 푡 Gb푋 1 푑 휅 G푋 exp d 푑=2 0
The unregularized quantum period G푋 (푡) is equal to the fundamental term of Givental’s
퐽-series 퐽푋 (푡) from [ 82 ]. This implies that G푋 (푡) satises the quantum di erential equation 퐿(푡, 휕푡 ) with an irregular singularity. Its Fourier–Laplace transform Gb푋 (휅) satises the so-called regularized quantum di erential equation b퐿(휅, 휕휅 ) that is expected to be regular.
Under the assumption that all eigenvalues of the quantum multiplication operator ★0c1 (푋) are distinct the respective regularized connection is Fuchsian. However for 푋 = M퐶 (2, L)
only two eigenvalues have multiplicity one, as computed by Muñoz (see Proposition 6.4.2 for the statement). But under the identication of the period sequences we can show the regularized quantum connection is a summand of a Picard–Fuchs equation and, as a corollary, indeed has regular singularities.
푡푑 푑
휅 Summing (180) with weights 푑! or we get an equivalent reformulation of Corollary 5.4.2 in terms of the equality of the respective period series.
Proposition 5.5.1. Let (훾, 푐) be a colored trivalent graph of genus 푔 ≥ 2 with one colored
vertex. Assuming Conjecture 4.2.9 we have the equality ∫ d푧 G (푡) = exp(푡 · 푊 (푧)) , M퐶 (2,L) 훾,푐 푧 (182) |푧 |=1 휅 휋 휅 휅 1 , GbM퐶 (2,L) ( ) = ( ) for | | ≤ 푊f훾,푐 8푔 − 8 of the unregularized (resp. regularized) quantum period of M퐶 (2, L) and the oscillating integral (resp. period) of the graph potential 푊e훾,푐 . Corollary 5.5.2. The regularized quantum period is a solution of an ordinary di erential equation with regular singularities.
Proof. By Proposition 5.5.1 we have the identication of the regularized quantum period
with the period of the graph potential. As discussed in [ 47 , §3], the period of a Laurent polynomial 푓 satises an ordinary di erential equation, given by the Picard–Fuchs operator. The local system of solutions is a summand of a polarised variation of Hodge structures,
which by [ 54 , Theorem 4.5] has regular singularities.
Remark 5.5.3. This also allows us to use methods of convex geometry and random walks
to estimate the asymptotics of the coe cients, see e.g. [ 76 , Lemma 3.13]. In particular, the 4푘 1/4푘 푊 0 푔 푇 limit lim푘→∞ ([ ]푧 ) exists and is equal to 8( − 1) = M퐶 (2,L) . GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 59
In Tables 1 and 2 we give the period sequences for the even and odd graph potentials. In the odd case this in turn coincides with the quantum period sequence for M퐶 (2, L)
by Proposition 5.5.1 . These numbers were computed using the method introduced in
Section 3 for computing periods of graph potentials using the graph potential TQFT. This
e cient method for computing periods from Section 3 could be useful for algorithmically determining the quantum di erential equation for the quantum periods of M퐶 (2, L).
In Appendix A we will give some explicit calculations for the quantum period sequence, highlighting some of its patterns which one can read o from the tables.
5.6. On mutations and cluster charts. In this nal subsection we collect some obser- vations regarding mutations between the di erent graph potentials. Recall that in the context of Batyrev’s small toric degeneration approach to Fano mirror
symmetry cluster mutations in the sense of Berenstein–Fomin–Zelevinsky [ 31 ] and Fock–
Goncharov [ 67 ] were rediscovered independently
• for spherical varieties using representation theory, by Rusinko [ 162 ] • for del Pezzo surfaces and Fano threefolds using deformation theory, by Galkin
[ 72 ]. They found that the Laurent polynomials associated by a mirror construction to di erent toric degenerations of a Fano manifold are related to each other by volume-preserving birational changes of coordinates. In the del Pezzo case a recursive procedure to construct innitely many di erent Laurent polynomials from just one was constructed. An algebraic theory of mutations of mirror potentials was constructed in joint works
of Galkin with Usnich and Cruz Morales [ 52 , 79 ], and the respective excessive Laurent
phenomenon was proved in this context. In [ 79 ] from a more algebro-geometric point of
view (similar to that of Gross–Hacking–Keel–(Kontsevich) [ 85 , 86 ]), and then in [ 52 ] using the relation to cluster algebras and their upper bounds. In this way they constructed a framework free from (commutative) algebraic geometry, with a view towards quantization. Amongst others, the notions of exchange collections and upper bounds, and mutation rules for them are introduced, and the Laurent phenomenon proved by two methods (both geometric and algebraic).
Ten equivalence classes of exchange collections in dimension 2 were constructed in [ 79 ,
52 ]. It was conjectured that each respective upper bound written in all possible seeds is generated by a single Laurent polynomial with positive integer coe cients, which is equal to a Floer potential for a monotone Lagrangian torus in an integrable system associated by the Nishinou–Nohara–Ueda construction to the respective toric degeneration of a smooth del Pezzo surface. Moreover it was conjectured that these birational transformations coincide with Lagrangian wall-crossing formulae, and that Floer potentials for all embedded monotone Lagrangian tori are equal to one of the constructed functions. Most of these were settled positively by Vianna, Galkin–Mikhalkin, Pascale –Tonkonog
and Ekholm–Rizell–Tonkonog [ 174 , 77 , 153 , 161 ]. However in higher dimensions (apart from the holomorphic symplectic case) our knowledge is much less complete. Already at the most basic level, we do not know a recursive rule of generalized “mutations” that would enjoy Laurent phenomenon and the same time produce enough potentials.
Pascale –Tonkonog [ 153 ] introduced the idea that the birational mutation of potentials
of Cruz Moralez–Galkin [ 52 ] (which itself is a geometric lift in the sense of Berenstein– Zelevinsky of a piecewise-linear homeomorphism of polytopes) can be lifted further to a Lagrangian mutation, i.e. a symplectic surgery between the respective Lagrangian tori 60 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY with the help of auxilliary Lagrangian disc, such that class of its boundary belongs to the exchange collection.
Dimitroglou–Eckholm–Tonkonog [ 161 ] provide further insight into mutations of two- dimensional Lagrangian tori by interpreting it as two deformations of an immersed La- grangian sphere. For the latter they propose one unifying Floer-like potential with values in a (non-commutative) multiplicative preprojective algebra, out of which the Floer poten-
tials for Lagrangian tori are obtained by specialization. Finally, Hong–Kim–Lau [ 92 ] also study study Floer potentials for immersed Lagrangian subvarieties, which are isomorphic to a product of an embedded torus with an immersed two-sphere, from the perspective of
Abouzaid–Auroux–Katzarkov [ 2 ], based on Strominger–Yau–Zaslow [ 167 ]. This leads us to the following remark.
Remark 5.6.1. One naturally expects that the rational change of coordinates that we
construct and discuss in Section 2 is similarly associated with a (new kind of) Lagrangian mutations between monotone Lagrangian tori 퐿훾 in M퐶 (2, L).
One new and peculiar feature is that elementary transformations of graph potentials naturally involve three variables. This is in contrast to (various generalizations of) cluster algebras where a single mutations involves just two variables. Another is the fact that the birational transformations of this paper equate two binomi- als whereas cluster algebraic transformations are binomial-equals-monomial. A limiting
procedure, discussed in Remark 3.3.19 specializes these new transformations to the trans-
formations of Nohara–Ueda [ 149 , 147 , 148 ] between Floer potentials of Euclidean polygon
spaces. In [ 148 ] they show how the specialized transformations produce a covering for a
total space of Rietsch’s Lie-theoretic mirror potential from [ 159 ] for the Grassmannian of planes, which generally is an open subspace in Langlands dual homogeneous variety.
6. Semiorthogonal decompositions In this section we will discuss the structure of the bounded derived category of coherent b sheaves D (M퐶 (2, L)), by introducing a conjectural semiorthogonal decomposition, given
as Conjecture E in the introduction. Using properties of graph potentials, and assuming ho- mological mirror symmetry we will discuss evidence for this conjecture. Other supporting
evidence for the conjecture, not motivated by HMS considerations, is discussed in Section 7 . We also suggest two semiorthogonal decompositions with additional structure, namely a
Lefschetz decomposition (as introduced by Kuznetsov [ 116 ]) and we also introduce the notion of a Ringel–Samokhin decomposition, related to the structure of quasi-hereditary algebras and an informal analogy between moduli spaces of bundles with ag varieties.
6.1. The fundamental conjecture. Bondal–Kapranov [ 37 ] introduced the notion of a semiorthogonal decomposition, as a means to decompose a triangulated category. Later, Bondal–Orlov studied these semiorthogonal decompositions in algebraic geometry, see
[ 41 , 39 ] and references therein. A more recent overview is given in Kuznetsov’s ICM
address [ 118 ].
In Conjecture E we have formulated a conjecture giving a semiorthogonal decomposition into geometrically meaningful pieces, namely copies of Db (Sym푖 퐶) for 푖 = 0, . . . ,푔 − 1.
This was conjectured independently by Narasimhan (as communicated in [ 122 ]) and the
authors (see [ 25 , Conjecture 7]). In the conjecture we have not specied the precise form of the embedding functors. As discussed later, for the copy of Db (퐶) one natural candidate is given by the universal vector bundle W on 퐶 × M퐶 (2, L), but for higher symmetric powers there is no candidate yet. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 61
b 푖 Moreover, it was conjectured in [ 25 , Conjecture 2] that each piece D (Sym 퐶) is inde- composable, i.e. does not admit further semiorthogonal decompositions. This was proven 푔+3
푖 for ≤ b 2 c −1 in [13, Corollary D], with a weaker result being given in [34, Theorem 1.3]. Known results. There exist a partial semiorthogonal decomposition which agrees with
the conjectural decomposition. Narasimhan [ 137 , 138 ] and, independently, Fonarev–
Kuznetsov [ 69 ] have shown that the Fourier–Mukai functor ΦW, where W is the uni- versal vector bundle on 퐶 × M퐶 (2, L) is fully faithful. Together with the exceptional objects OM퐶 (2,L) and Θ this gives rise to 3 components in ( 22 ).
The rst and third author have shown in [ 30 ] that it is possible to twist the embedding ΦW
by Θ to obtain the embedding of the four low-dimensional components in ( 22 ). More precisely, Theorem B of op. cit gives the semiorthogonal decomposition b , , b 퐶 , , b 퐶 , (183) D (M퐶 (2 L)) = hOM퐶 (2,L) ΦW (D ( )) Θ ΦW (D ( )) ⊗ Θ Ai for 푔 ≥ 12, where A is dened as the orthogonal to the rst subcategories. The bound on 푔 arises from the need for su ciently strong vanishing results in op. cit., and conjecturally can be removed. Context. This conjecture forms a part of a greater program, which aims at nding natural semiorthogonal decompositions of the bounded derived categories of coherent sheaves on the moduli spaces of (stable) objects in the bounded derived categories of coherent sheaves on algebraic varieties. Another important class of moduli spaces of coherent sheaves are punctual Hilbert schemes, which for smooth curves coincide with their symmetric powers. The respective categories
are studied in [ 98 , §1.4] and [ 29 , Theorem D]. This result was originally obtained in [ 169 ], which also discusses semiorthogonal decompositions for moduli spaces of stable pairs on K3 surfaces. For Hilbert schemes of points on surfaces, the study of (partial) semiorthogonal decompo-
sitions was started in [ 113 ]. It was generalized to higher dimensions in [ 112 , 27 ]. Further
generalisations to nested Hilbert schemes were obtained in [ 98 , §3.1.4] and [ 29 , Theo- rem E]. Moreover, for moduli spaces of coherent sheaves on rational surfaces, wall-crossing
results give rise to semiorthogonal decompositions as in [ 10 ]. In a di erent direction, a semiorthogonal decomposition for the Hilbert square of a cubic hypersurface was obtained
in [ 26 ] using the Fano variety of lines.
6.2. Lefschetz decompositions. A more structured form of Conjecture E is given by
that of a Lefschetz decomposition, a notion introduced by Kuznetsov [ 116 ] in the context of his homological elaboration on projective duality.
Denition 6.2.1. Let 푋 be a smooth projective variety, and O푋 (1) be a line bundle on 푋. b A (right) Lefschetz decomposition of D (푋) with respect to the line bundle O푋 (1) is a semiorthogonal decomposition b (184) D (푋) = hA0, A1 ⊗ O푋 (1),..., A푛−1 ⊗ O푋 (푛 − 1)i, such that
(185) A푛−1 ⊆ A푛−2 ⊆ ... ⊆ A0.
We say that 푛 is the length of the decomposition, and A0 is the starting block.
The decomposition ( 184 ) is rectangular if A푛−1 = ... = A0.
By [ 117 , Lemma 2.18(i)] a Lefschetz decomposition is completely determined by its starting ⊥ block, as one can inductively dene A푖 := (A0 ⊗ O푋 (−푖)) ∩ A푖−1. This allows us to dene 62 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY the notion of a minimal Lefschetz collection, by considering the (partial) inclusion order b on the starting blocks for Lefschetz collections of D (푋) with respect to O푋 (1).
In [ 119 , Denition 1.3(iii)] the notion of residual category for a Lefschetz exceptional collection is introduced, to measure how far it is from being rectangular. We can generalize this in a straightforward fashion to an arbitrary Lefschetz decomposition as follows, see
also [ 119 , §2.2].
b Denition 6.2.2. Let D (푋) = hA0, A1 ⊗ O푋 (1),..., A푛−1 ⊗ O푋 (푛 − 1)i be a Lefschetz decomposition of length 푛 with respect to the line bundle O푋 (1). The residual category R of this decomposition is dened as ⊥ (186) R := hA푛−1, A푛−1 ⊗ O푋 (1),..., A푛−1 ⊗ O푋 (푛 − 1)i ;
Remark 6.2.3. As discussed in [ 120 ] (in the context of Fano varieties with vanishing odd cohomology), properties of the residual category are conjecturally related to the structure of ber over zero for the quantum cohomology of 푋. For Fano varieties with odd cohomology, one aspect of a generalization of this conjecture is the prediction that the dimension of the Hochschild homology of the residual category is equal to the length
of the eigenspace associated to the eigenvalue 0. By Proposition 6.4.2 this eigenspace is isomorphic to H• (Sym푔−1 퐶) (of dimension (2푔 + 1)!/푔!2), which by the Hochschild– 푔−1 Kostant–Rosenberg decomposition is isomorphic to HH• (Sym 퐶), hence the equality holds.
This brings us to the following more precise form of Conjecture E , incorporating a Lefschetz structure.
Conjecture 6.2.4. Let 퐶 be a smooth projective curve of genus 푔. Then there exists a b minimal Lefschetz decomposition of D (M퐶 (2, L)) of length 2 with respect to Θ, whose rst block has a semiorthogonal decomposition b b b 2 b 푔−1 (187) A0 = hD (pt), D (퐶), D (Sym 퐶) ..., D (Sym 퐶)i, so in particular the residual category is Db (Sym푔−1 퐶).
For 푔 = 2 this conjecture is known by considering the semiorthogonal decomposition [ 41 ,
Theorem 2.9] and mutating the object OM퐶 (2,L) to the right. Proposition 6.2.5 (Bondal–Orlov). Let 퐶 be a curve of genus 2. Then b , , b 퐶 , . (188) D (M퐶 (2 L)) = hOM퐶 (2,L) ΦW (D ( )) Θi is a minimal Lefschetz decomposition, whose residual category is equivalent to Db (퐶).
Induced polarizations. An important feature for (non-rectangular) Lefschetz decom- positions and their residual categories R, is that we obtain an autoequivalence 휏R of R induced by − ⊗ O푋 (1) [ 119 , Theorem 2.8]. In this setting the index of the ambient variety is 2, so we get that 휏◦2 −1 , , (189) R = SR [dim M퐶 (2 L)]
where SR denotes the Serre functor of R. If we assume Conjecture 6.2.4 , so that the residual category R Db (Sym푔−1 퐶), this formula tells us that the induced polarization is a square root of (the inverse) of the Serre functor, shifted by 3푔 − 3. Hence 휏R is a square root of − ⊗ 휔∨ [2푔 − 2], which one could call a categorical theta characteristic. Sym푔−1 퐶 For 푔 = 2 we have that Sym1 퐶 = 퐶 is anti-Fano, so the categorical theta characteristic is necessarily of the form ∗ st b (190) (− ⊗ M, 휎 )[푔 − 1] ∈ Aut (D (퐶)) (Pic퐶 o Aut퐶) × Z GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 63 such that 휔∨ 휎∗ (191) 퐶 M ⊗ M for M a line bundle of degree 1 on 퐶 and 휎 an involution.
Remark 6.2.6. The description of autoequivalences of Db (Sym푔−1 퐶) is an open problem, as Sym푔−1 퐶 is of general type, yet does not have an ample canonical bundle for 푔 ≥ 3.
Hence one cannot apply [ 40 , Theorem 3.1], and not every autoequivalence is necessarily standard. Therefore it is not clear a priori what categorical theta characteristics can look like.
6.3. Ringel–Samokhin-type decompositions. There is another way in which we can b impose further conditions on a semiorthogonal decomposition for D (M퐶 (2, L)). This condition is inspired by the theory of hereditary algebras and Ringel duality, and by similar semiorthogonal decompositions for Db (퐺/퐵) obtained by Samokhin, and encodes special symmetries not found in most semiorthogonal decompositions. Let 푋 be a smooth projective variety, and consider a semiorthogonal decomposition b (192) D (푋) = hA0,..., A푛i. of length 푛 + 1 ≥ 2. Let 휎 be an antiequivalence of Db (푋) which is moreover an involution. Let B푘 be the image of A푛−푘 under this involution, for 푘 = 0, . . . , 푛. Then we have a semiorthogonal decomposition b (193) D (푋) = hB0,..., B푛i.
On the other hand consider the left dual decomposition in the sense of [ 164 , Denition 3.6] denoted by b (194) D (푋) = hC0,..., C푛i. This brings us to the following denition.
Denition 6.3.1. A decomposition ( 192 ) is of Ringel–Samokhin type if there exists an b autoequivalence 휎 of D (푋) such for all 푘 = 0, . . . , 푛 the subcategory C푘 is the image of B푘 under 휎.
Remark 6.3.2. In [ 164 ] Samokhin studied such exceptional collections for the derived category Db (퐺/퐵), where 퐺/퐵 is the full ag variety of a group of rank 2, and the anti- H (−, 휔1/2 ) equivalence is given by R om 퐺/퐵 .
We can rene Conjecture E into a semiorthogonal decomposition of Ringel–Samokhin-type as follows.
Conjecture 6.3.3. Let 퐶 be a smooth projective curve of genus 푔 ≥ 2. Then there exists a Ringel–Samokhin-type decomposition (195) b b b b 푔−2 b 푔−1 b 푔−2 D (M퐶 (2, L)) = hD (pt), D (퐶),..., D (Sym 퐶), D (Sym 퐶), D (Sym 퐶),..., ..., Db (퐶), Db (pt)i , with the anti-equivalence given by RHom(− OM퐶 (2,L) (1)).
This conjecture highlights how semiorthogonal decompositions of M퐶 (2, L) are expected to have strong symmetry properties, much stronger than found for most varieties exhibiting semiorthogonal decompositions. For 푔 = 2 we can check this conjecture using the following general result for Ringel– Samokhin-type decompositions of length 3. 64 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Proposition 6.3.4. Let 푋 be a smooth projective variety. Assume that we have a semiorthog- onal decomposition b (196) D (푋) = hL1, A, L2i where L1, L2 are exceptional line bundles. Then this decomposition is of Ringel–Samokhin- ∨ 푋 type if and only if L1 ⊗ L2 is a theta characteristic of .
Proof. A decomposition of the form ( 196 ) is determined by the line bundles L1 and L2, as the subcategory A consists of the objects 퐸 ∈ Db (푋) such that • • (197) Hom (RHom(퐸, O푋 ), L1) = Hom (L2, 퐸) = 0.
The left dual decomposition of ( 196 ) is given by b (푋) h , , ⊗ 휔∨ i. (198) D = L2 LL2 A L1 푋 On the other hand, we consider the anti-equivalence given by dualizing and tensoring with a line bundle M gives the semiorthogonal decomposition b 푋 ∨ , ∨ , ∨ . (199) D ( ) = hL2 ⊗ L A ⊗ L L1 ⊗ Li ⊗2 These two semiorthogonal decompositions agree if and only if L L2 , so that we ∨ ⊗2 휔∨ ∨ 푋 have (L1 ⊗ L2) 푋 and hence L1 ⊗ L2 is a theta characteristic of .
By Proposition 6.2.5 , using that M퐶 (2, L) is of index 2, we obtain the following corollary.
Corollary 6.3.5. Conjecture 6.3.3 holds for 푔 = 2.
6.4. Decompositions from the point of view of mirror symmetry. To understand the results in this section we will brie y recall homological mirror symmetry for Fano varieties. If 푋 is a (smooth projective) Fano variety, its mirror is expected to be a Landau– Ginzburg model 푓 : 푌 → A1 (where 푌 is a quasiprojective variety), such that we have equivalences of triangulated categories Fuk(푋) MF(푌, 푓 ), (200) Db (푋) FS(푌, 푓 ). On the rst line we have that the Fukaya category of 푋 has an orthogonal decomposi- tion, indexed by the eigenvalues of c1 (푋) ∗0 −, see [ 166 ]. For the matrix factorization
category we have an orthogonal decomposition, indexed by the critical values of 푓 [ 151 , Proposition 1.14]. Hence under the homological mirror symmetry conjecture these sets (of eigenvalues, and critical values) are the same. For the second line there is no such natural (semiorthogonal) decomposition. But the philosophy behind Dubrovin’s conjecture (which a priori is only formulated in the case of semisimple quantum cohomology) predicts how an orthogonal decomposition of the triangulated category Fuk(푋) gives rise to a semiorthogonal decomposition of Db (푋). Generalizations of this philosophy to not necessarily semisimple quantum cohomology
are discussed in [ 165 ].
b Hence to understand semiorthogonal decompositions for D (M퐶 (2, L)) using mirror symmetry, one can study either of the following • (1) the eigenvalues of the quantum multiplication c1 (M퐶 (2, L))∗0− on QH (M퐶 (2, L)); (2) the critical values of the potential 푓 : 푌 → A1. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 65
Remark 6.4.1. There is no worked out candidate construction for the Landau–Ginzburg mirror (푌, 푓 ) of M퐶 (2, L) yet, let alone a proof of homological mirror symmetry. But the
graph potentials we have constructed in Section 2 can be seen as open cluster charts of 푌 , and gluing these tori together along the birational transformations between (which we
can do because of the compatibilities from Theorems 2.2.3 and 2.2.4 ) them is a rst step one can take in the construction of the Landau–Ginzburg model mirror to M퐶 (2, L).
A better understanding of the critical loci would be the next step in understanding if and how these graph potentials can be glued together to obtain (part of) the Landau–Ginzburg
model, but we will not do this here. We only discuss in Remark 6.4.6 how the critical locus over 0 (if 푔 is even) is a least 푔 − 1-dimensional. One necessary feature of Landau–Ginzburg mirrors in homological mirror symmetry for Fano varieties is that the critical locus of the potential is compact and that there are no critical points at innity. Because the critical locus must contain higher-dimensional components (and a component of dimension at least 푔 − 1 is exhibited) we see the need for multiple cluster charts, as a positive-dimensional proper variety needs multiple a ne charts (so in this case at least 푔, and hence this grows to ∞ as the genus grows).
Muñoz’s description. It turns out that all eigenvalues of quantum multiplication with c1 have been described, which allows us to check property O for M퐶 (2, L), adding an another class of Fano varieties where the conjecture is satised. The following summarizes this description.
Proposition 6.4.2 (Muñoz). The eigenvalue decomposition for the quantum multiplication • by c1 (M퐶 (2, L)) on QH (M퐶 (2, L)) is 푔−1 • Ê (201) QH (M퐶 (2, L)) = 퐻푚 푚=1−푔 where √ √ (1) the eigenvalues are 8(1−푔), 8(2−푔) −1, 8(3−푔),..., 8(푔−3), 8(푔−2) −1, 8(푔−1); • 푔−1−|푚 | (2) 퐻푚 is isomorphic (as a vector space) to H (Sym 퐶).
This is a combination of various results of Muñoz. First we have [ 136 , Proposition 20], describing the eigenvalues of multiplication by the generator of the Picard group on the instanton Floer homology of the 3-manifold given by taking the product of the curve (seen 1 as a real manifold) and 푆 . The (conjectural) identication as rings from [ 136 , Theorem 1] with the quantum cohomology of M퐶 (2, L) is in turn given by [ 135 , Corollary 21] after an explicit description in terms of generators and relations for both rings. Hence we can interpret any result for instanton Floer homology as a result for quantum cohomology. In
[ 136 , Conjecture 24] the conjectural decomposition was given, and in [ 134 , Corollary 3.7] it was proved. Property O. An interesting symmetry property for the quantum cohomology of a Fano variety 푋 of index 푟, is obtained by considering the exceptional collection O푋 ,..., O푋 (푟 −1) in Db (푋). The existence of this exceptional collection can, extending the Dubrovin dictio-
nary, be encoded in quantum cohomology using [ 75 , Denition 3.1.1] as follows.
Denition 6.4.3. Let 푋 be a smooth projective Fano variety, of index 푟 ≥ 1. Dene 푇 푢 푢 푋 . (202) := max{| | | ∈ C is an eigenvalue of c1 ( ) ∗0 −} ∈ Q≥0 Then we say that 푋 has property O if
(1) 푇 is an eigenvalue of c1 (푋) ∗0 −, of multiplicity 1; 66 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
1 2 (a) Theta graph with marked edges
푎 푐 푥
푣1 푣2 푏 푑
(b) Local picture of étale cover of decorated Theta graph
(c) Decorated triple cover of Theta graph
Figure 21. Decorations for étale covers the Theta graph
(2) if 푢 is another eigenvalue of c1 (푋) ∗0 − such that |푢| = 푇 , then there exists a primitive 푟th root of unity 휁 such that 푢 = 휁푇 .
In [ 75 , Conjecture 3.1.2] it was conjectured that this property holds for all Fano varieties.
In [ 44 ] it was checked for all homogeneous varieties 퐺/푃, and in [ 165 , Corollary 7.7] the case of complete intersections of index 푟 ≥ 2 in P푛 was checked.
Hence by Proposition 6.4.2 we immediately obtain the following result.
Corollary 6.4.4. Property O holds for M퐶 (2, L), where 푇 = 8(푔 − 1).
Remark 6.4.5 (The conifold point). In [ 75 , Remark 3.1.6] a conjecture regarding the value
of 푇 was suggested, relating it to the conifold point as discussed in Proposition 2.4.2 . Dene 푇con := 푊e훾,푐 (푥con) as the value of the potential 푊e훾,푐 at the conifold point. The conjecture says that
(203) 푇 = 푇con,
From the description of the conifold point in Proposition 2.4.2 we can easily check this equality. Indeed, the graph potential푊e훾,푐 is the sum over the vertex potentials푊e푣 for 푣 ∈ 푉 , and the evaluation at the conifold point for each of these is equal to 4. There are 2(푔 − 1) vertices, so 푇con = 8(푔 − 1). This agrees with the value 푇 from Corollary 6.4.4 .
Critical values. As discussed in the introduction of this section, the eigenvalues of quan- tum multiplication should correspond to the critical values of a suitable Landau–Ginzburg model. There is a priori no reason why it would su ce to consider a single Laurent polynomial, which is the restriction of the potential to a Zariski-open torus inside 푌 . But it turns out that graph potentials already see all the critical values one expects. This is
immediate by comparing Proposition 6.4.2 to Proposition 2.4.1 . It should be remarked that we have not proven that there are no other critical values, but we conjecture that these are all. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 67
Remark 6.4.6. We can understand the dimension of the critical locus over 0 as follows. Let 푔 be even. We can obtain a colored trivalent graph (훾, 푐) of genus 푔 with 푔 − 1 colored vertices (so half of them are colored), by considering an étale cover of degree 푔 − 1 of the Theta graph, so that by Riemann–Hurwitz the graph (훾, 푐) is of the desired form. For bookkeeping reasons we introduce a decoration of the edges of the Theta graph using
three distinct styles as in Figure 21(a) . The induced local structure of the graph (훾, 푐) is
given in Figure 21(b) , which is a decorated version of Figure 5(b) . To illustrate the global
structure, in Figure 21(c) the decorated étale cover of the Theta graph is given. √ √ To estimate the size of the critical locus over 0, assign −1 to all dashed edges, − −1 to
all dotted edges, and (푔 − 1) arbitrary values to the remaining edges. By ( 83 ) every point with these coordinates is critical of value 0. Hence the critical locus has a component over 0 of dimension at least 푔 − 1 (we expect it to be precisely of dimension 푔 − 1).
• 푔−1 Remark 6.4.7. By Proposition 6.4.2 , we see that the H (Sym 퐶) appears as an eigenspace
of the quantum multiplication with eigenvalue 0. On the other hand, by Remark 6.4.6 , we see that the critical locus with critical value 0 is at least 푔 − 1 dimensional. This is in agreement with the appearance of Db (Sym푔−1 퐶) in our conjectural semiorthogonal b decomposition of D (M퐶 (2, L)).
7. Decomposition in the Grothendieck ring of varieties
In this section we compute the class of the moduli space M퐶 (2, L) in the Grothendieck ring of varieties K0 (Var/푘), and the main result in this section is a proof of Theorem F . This
identity gives further evidence for Conjecture E , by considering the appropriate motivic measure on K0 (Var/푘) as we will explain in Section 7.3 . Recall that the Grothendieck ring of varieties is generated by the isomorphism classes [푋] of algebraic varieties over 푘, modulo the relations [푋] = [푈 ] + [푍] for 푍 ↩→ 푋 a closed subvariety, and 푈 = 푋 \ 푍 its complement. The product of varieties induces the multiplica- tive structure, such that [pt] is the unit. An important element of this ring is the class of the a ne line L = [A1] = [P1] − [pt], also called the Lefschetz class. An alternative presentation for this ring in characteristic 0, relevant to our goal, is the
Bittner presentation from [ 35 ]. It says that K0 (Var/푘) is isomorphic to the ring generated by isomorphism classes [푋] of smooth and proper algebraic varieties over 푘, modulo the relations [Bl푍 푋] − [퐸] = [푋] − [푍] for 푍 ↩→ 푋 a smooth closed subvariety, and 퐸 → 푍 the exceptional divisor in the blowup Bl푍 푋 → 푋. Via the cut-and-paste relations one can obtain the standard identities (204) [푋] = [퐹][푌 ] if 푋 → 푌 is a Zariski-locally trivial bration with bers 퐹, and 푛 ∑︁ 1 − L푛+1 (205) [P푛] = L푖 = . 1 − L 푖=0
In order to use this alternative presentation, and because we depend on other results in the literature which are only phrased for algebraically closed elds (but likely hold more generally), we will work over an arbitrary algebraically closed eld 푘 of characteristic 0. Motivic measures. A motivic measure is a ring morphism whose domain is the Grothendieck ring of varieties. In our setting we consider the motivic measure
(206) 휇 : K0 (Var/푘) → K0 (dgCat/푘) 68 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
from [ 38 , §8], obtained by using the Bittner presentation and Orlov’s blowup formula, and sending [푋] to the class of the (unique) dg enhancement of K0 (dgCat/푘). Then the
identity ( 23 ) matches up with Conjecture E .
The Grothendieck ring of dg categories K0 (dgCat푘 ) precisely encodes semiorthogonal decompositions, as it is generated by the quasiequivalence classes [C] of smooth and proper pretriangulated dg categories, modulo the relations (207) [C] = [A] + [B] for every semiorthogonal decomposition C = hA, Bi.
Hence the image of the equality ( 23 ) is consistent with Conjecture E (up to 2-torsion, as L ↦→ 1), as explained by the following corollary.
Corollary 7.0.1. We have the equality 푔−2 b b 푔−1 ∑︁ b 푖 0 (208) [D (M퐶 (2, L))] = [D (Sym 퐶)] + 2[D (Sym 퐶)] + 푇 푖=0 0 0 in K0 (dgCat/푘), for some class 푇 such that 2 · 푇 = 0.
Again we expect that 푇 0 = 0, but our method of proof is not strong enough to remove this error term. This is precisely the equality induced by the semiorthogonal decomposition from Conjec-
ture E , and therefore can be seen as further evidence for it.
Comparison to other results. Motivic and cohomological invariants of M퐶 (2, L) have been an active topic of interest for a long time, and many tools are used for this. The
starting point for the computation of these invariants can be found in Newstead’s [ 142 ], where the Betti numbers were rst computed.
With a view towards providing evidence for Conjecture E , Lee has given in [ 122 , The-
orem 1.2] an isomorphism similar to ( 23 ) which holds in any semisimple category of
motives, such as the category of numerical motives. It is based on the isomorphism [ 12 , Theorem 2.7] due to del Baño, which also require this semisimplicity. This result is stronger in the sense that it is known that the error term 푇 vanishes. But it only works for (the Grothendieck group of) a semisimple category of motives. The motivic measure
(209) 휇 : K0 (Var/푘) → K0 (Chow/푘) which sends the class [푋] of a smooth projective variety to the class [픥(푋)] of its Chow
motive, allows one to obtain a strengthening (modulo vanishing of푇 ) of [ 122 , Theorem 1.2].
See also [ 12 , Remark 2.8].
From the identity ( 23 ) it is possible to deduce known results on motivic and cohomological invariants of M퐶 (2, L) by taking the appropriate motivic measures. Examples of this are given by the Betti polynomial (with values in Z[푡]), the Hodge–Poincaré polynomial (with values in Z[푥,푦]). It would be interesting to weaken the assumptions on the eld 푘, to also recover point-counting realizations.
7.1. The Thaddeus picture. We break the proof of Theorem F up into several steps. The main idea of the proof is to use Thaddeus’ variation of GIT for moduli of stable pairs and compare the class of the variety when we cross a wall using a telescopic sum. This is the
“Thaddeus picture” alluded to in the title of this section, and will be given in ( 212 ). GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 69
Setup. Let 푑 be an odd integer which is greater than 2푔 − 2. Let L be a line bundle of 푑 푀푑 1 퐶, ∨ 푖 ,..., 푑 degree and let 0 = P(H ( L )). For each ∈ {1 ( − 1)/2}, Thaddeus constructs 푑 푑 푑
푀 푀 푀 in [168] two smooth projective varieties 푖 and e푖 such that e푖 is: 푀푑 − 푖 퐶 • a blowup of 푖−1 along a projective bundle P(W푖 ) over Sym . 푀푑 + 푖 퐶 • a blowup of 푖 along a projective bundle P(W푖 ) over Sym ; − + 푖 퐶 푖 Here W푖 (respectively W푖 ) is vector bundle on Sym of rank (respectively of rank 푑 + 푔 − 2푖 − 1). This leads us to the following ip diagram: 퐸
푀푑 e푖 (210) − 푀푑 푀푑 + P(W푖 ) 푖−1 푖 P(W푖 )
Sym푖 퐶
Summarizing the situation from [ 168 ] for all 푖 we obtain the following theorem.
Theorem 7.1.1 (Thaddeus). With the above notation: 푀푑 푀푑 푖 퐶 푖,푑 푔 푖 (1) There is a ip 푖−1 d 푖 with center Sym and type ( + − 2 − 1) for each 푖 ∈ {1,..., (푑 − 1)/2}. 푀푑 푑+푔−2 푀 푀푑 (2) We have that 0 P , and e1 1 . (3) There is a natural map 휋 푀푑 , (211) : (푑−1)/2 → M퐶 (2 L) 0 with ber P(H (퐶, E)) over a stable bundle E in M퐶 (2, L). Moreover, if 푑 ≥ 4푔 − 3, then 휋 is a projective bundle associated to a vector bundle of rank 푑 + 2(1 − 푔).
This yields the following picture, which we refer to as the “Thaddeus picture”: (212) 푀푑 푀푑 푀푑 e1 e2 e(푑−1)/2
푀푑 푀푑 푀푑 ... 푀푑 푀푑 0 1 2 (푑−3)/2 (푑−1)/2 휋
M퐶 (2, L) All morphisms in the diagram, except 휋, are blowups. To minimize the dimension of the moduli spaces and the number of steps involved in the construction, we will take 푑 = 4푔 − 3. Then
(1) dim 푀푖 = 5푔 − 5; 푀4푔−3, . . . , 푀4푔−3 (2) we are considering moduli spaces 0 2푔−2 ; 2푔−2 (3) the morphism 푀2푔−2 is a P -bration. 70 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
Setup for the proof. We now give some easy lemmas, as a setup for the proof in Sec-
tion 7.2 . The following lemma follows directly from Theorem 7.1.1 and ( 204 ).
Lemma 7.1.2. Let 푑 be an odd integer greater than 4푔 − 4, then we have that 1 − L푑+2(1−푔) (213) [푀푑 ] = [M (2, L)] (푑−1)/2 1 − L 퐶 in K0 (Var/푘).
Using the description from Lemma 7.1.2 for 푑 = 4푔 − 3 and 푑 = 4푔 − 1 and computing the di erence gives the following description. One could take a less optimal choice, at the cost
of obtaining a larger coe cient in Theorem F .
Lemma 7.1.3. We have that , 푀4푔−1 2 푀4푔−3 (214) (1 + L)[M퐶 (2 L)] = [ 2푔−1 ] − L [ 2푔−2 ] in K0 (Var/푘).
For every ip diagram as in ( 212 ) one can prove the following, where only standard bookkeeping techniques are required.
Proposition 7.1.4. Let 푑 be an odd integer greater than 4푔−4. Then for 푖 = 1,..., (푑 −1)/2 푀푑 푀푑 the di erence of the classes [ 푖 ] and [ 푖−1] satises (215) L [푀푑 ] − [푀푑 ] = (1 − L푑+푔−2푖−2)(1 − L푖 ) − (1 − L푖−1)(1 − L푑+푔−2푖−1) [Sym푖 퐶] 푖 푖−1 (1 − L)2 in K0 (Var/푘).
Sketch of proof. By applying the blowup formula and the projective bundle formula twice with the appropriate codimension and relative dimension we get the equalities 푖 푑 푑 + 1 − L + [푀e ] = [푀 ] − [P(W )] + [P(W )] 푖 푖 푖 1 − L 푖 L(1 − L푖−1) 1 − L푑+푔−2푖−2 = [푀푑 ] + [Sym푖 퐶] 푖 − L − L (216) 1 1 − L푑+푔−2푖−1 푀푑 푀푑 − 1 − [ e ] = [ − ] − [P(W )] + [P(W )] 푖 푖 1 푖 1 − L 푖 L(1 − L푑+푔−2푖−2) 1 − L푖 = [푀푑 ] + [Sym푖 퐶] 푖−1 1 − L 1 − L and subtracting them gives the result.
Finally, let us recall the following fundamental identity encoding the behavior of Abel–
Jacobi morphisms, as e.g. discussed in [ 91 , §3].
Proposition 7.1.5. Let 퐶 be a curve of genus 푔. Let 푒 ≥ 0 and denote 푎 = (푔 − 1) + 푒. Then we have an equality (217) [Pic푎 퐶][P푒−1] = [Sym푔−1+푒 퐶] − L푒 [Sym푔−1−푒 퐶] in K0 (Var/푘).
Observe that under our assumptions we have isomorphisms Pic푖 퐶 Jac퐶 for all 푖 ∈ Z, by the existence of a rational point. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 71
7.2. Proof of the decomposition. We will restrict ourselves to the cases where 푑 = 4푔−3
and 4푔−1, i.e. the rst two degrees for which the morphism 휋 in ( 212 ) is an equidimensional projective bration. We will compare di erent stages of the Thaddeus picture, using the fol- lowing notation for the di erence of classes of moduli of stable pairs, where 푖 = 0,..., 2푔−2. 훿푀 푀4푔−1 2 푀4푔−3 . (218) 푖 := [ 푖 ] − L [ 푖 ] 훿푀 For notational convenience, we also set −1 := 0 Now for 푖 = 0,..., 2푔 − 2, we dene
(219) 푋푖 := 훿푀푖 − 훿푀푖−1.
We can describe the classes 푋푖 in the following way.
Proposition 7.2.1. For 푖 = 0,..., 2푔 − 2 we have that 푖 푖 (220) 푋푖 = L (1 + L)[Sym 퐶]
Proof. The proof follows from the denition of 푋푖 and by applying Proposition 7.1.4 , via the following chain of equalities.
푋푖 = 훿푀푖 − 훿푀푖−1 푀4푔−1 2 푀4푔−3 푀4푔−1 2 푀4푔−3 (221) = [ 푖 ] − L [ 푖 ] − [ 푖−1 ] + L [ 푖−1 ] 푀4푔−1 푀4푔−1 2 푀4푔−3 푀4푔−3 = ([ 푖 ] − [ 푖−1 ]) − L ([ 푖 ] − [ 푖−1 ])
We now use Proposition 7.1.4 and rewrite 푋푖 as L 푋 = (1 − L5푔−3−2푖 )(1 − L푖 ) − (1 − L5푔−2−2푖 )(1 − L푖−1) 푖 (1 − L)2 − L2 (1 − L5푔−5−2푖 )(1 − L푖 ) − (1 − L5푔−4−2푖 )(1 − L푖−1) [Sym푖 퐶]
(222) L(1 − L푖 )(1 − L2) L(1 − L2)(1 − L푖−1) = − [Sym푖 퐶] (1 − L)2 (1 − L)2 L(1 − L2) = (L푖−1 − L푖 ) [Sym푖 퐶] (1 − L)2 = L푖 (1 + L)[Sym푖 퐶].
Using Propositions 7.2.1 and 7.1.5 we obtain the following result, where we use the following polynomial in L: (223) P(푖) := L2푔−2−푖 (1 + L)(1 + L + ... + L푔−2−푖 ).
Proposition 7.2.2. For 푖 = 0, . . . ,푔 − 2, we have that 푖 푖 3푔−3−2푖 (224) 푋푖 + 푋2푔−2−푖 = [Sym 퐶](L + L )(1 + L) + P(푖)[Jac(퐶)],
Proof. By Proposition 7.1.5 we get (225) [Sym2푔−2−푖 퐶] = L푔−푖−1 [Sym푖 퐶] + [P푔−2−푖 ][Jac(퐶)]
We apply Proposition 7.2.1 for 푖 ∈ {0, 1, . . . ,푔 − 2} to ( 225 ) to get 2푔−2−푖 2푔−2−푖 푋2푔−2−푖 = L (1 + L)[Sym 퐶] (226) = L3푔−3−2푖 (1 + L)[Sym푖 퐶] + L2푔−2−푖 (1 + L)[P푔−2−푖 ][Jac(퐶)] = L3푔−3−2푖 (1 + L)[Sym푖 퐶] + P(푖)[Jac(퐶)] 72 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
푖 푖 Now by Proposition 7.2.1 for 푖 ≤ 푔 − 1, we get 푋푖 = L (1 + L)[Sym 퐶]. Thus the proof follows.
The following proposition is an important step in the proof of Theorem F . It shows that there are no contributions of the Jacobian of 퐶 to the class of M퐶 (2, L).
Proposition 7.2.3. We have that 푔−2 ∑︁ 푖 퐶 푀4푔−1 푀4푔−1 (227) P( ) [Jac( )] = [ 2푔−2 ] − [ 2푔−1 ] 푖=0 in K0 (Var/푘).
Proof. First let us simplify the left hand side of the above equation. We have that 푔−2 푔−2 ∑︁ 1 + L ∑︁ P(푖) = L2푔−2−푖 (1 − L푔−1−푖 ) 1 − L 푖=0 푖=0 푔−2 (1 + L) ∑︁ = L2푔−2−푖 − L3푔−3−2푖 1 − L 푖=0 푔−2 (228) L푔 (1 + L) ∑︁ = (L푖 − L2푖+1) 1 − L 푖=0 L푔 (1 + L) 1 − L푔−1 L(1 − (L2)푔−1) = − 1 − L 1 − L 1 − L2 L푔 = (1 − L푔−1)(1 − L푔) (1 − L)2 We will be done if we can show that the above expression multiplied by the class of Jac(퐶) 4푔−1 4푔−1
푀 푀 푒 푔 is equal to [ 2푔−2 ] − [ 2푔−1 ]. First by Proposition 7.1.5 where we take = we get (229) [Sym2푔−1 퐶] = [P푔−1][Jac(퐶)].
Combining this with Proposition 7.1.4 we get (230) − − 퐿 [푀4푔 1] − [푀4푔 1] = (1−L푔) (1−L푔−1)(1−L2푔−1) − (1−L푔)(1−L2푔−2) [Jac(퐶)]. 2푔−1 2푔−2 (1 − L)3 Now we use that (231) (1 − L푔−1)(1 − L2푔−1) − (1 − L푔)(1 − L2푔−2) = −L푔−1 (1 − L) + L2푔−1 (1 − L)
Thus from ( 231 ) we get 푔 − − L(1 − L ) (232) [푀4푔 1] − [푀4푔 1] = − L푔−1 (1 − L푔−1)(1 − L)[Jac(퐶)], 2푔−1 2푔−2 (1 − L)3 and we are done.
Now we are ready to complete the proof of Theorem F .
Proof of Theorem F . First we write the class of M퐶 (2, L) in terms of the di erence of classes
of the smooth projective varieties considered by Thaddeus. We apply Lemma 7.1.3 to get , 훿푀 푀4푔−1 푀4푔−1 (233) (1 + L)[M퐶 (2 L)] = [ 2푔−2] + [ 2푔−1 ] − [ 2푔−2 ] Now we can write 2푔−2 ∑︁ (234) [훿푀2푔−2] = 푋푖, 푖=0 GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 73
where 푋푖 ’s are as used in Proposition 7.2.2 . Thus applying Proposition 7.2.2 we get
(1 + L)[M퐶 (2, L)] 훿푀4푔−3 푀4푔−1 푀4푔−1 = [ 2푔−2 ] + [ 2푔−1 ] − [ 2푔−2 ] 2푔−2 ∑︁ 푋 푀4푔−1 푀4푔−1 = 푖 + [ 2푔−1 ] − [ 2푔−2 ] 푖=0 푔−2 ∑︁ 푋 푋 푋 푀4푔−1 푀4푔−1 = ( 푖 + 2푔−2−푖 ) + 푔−1 + [ 2푔−1 ] − [ 2푔−2 ] 푖=0 (235) 푔−2 ∑︁ = L푔−1 (1 + L)[Sym푔−1 퐶] + (L푖 + L3푔−3−2푖 )(1 + L)[Sym푖 퐶] 푖=0 푔−2 ∑︁ 푖 퐶 푀4푔−1 푀4푔−1 + P( ) [Jac( )] + [ 2푔−1 ] − [ 2푔−2 ] 푖=0 푔−2 ∑︁ = L푔−1 (1 + L)[Sym푔−1 퐶] + (L푖 + L3푔−3−2푖 )(1 + L)[Sym푖 퐶] 푖=0
where the last step is by Proposition 7.2.3 .
7.3. Motivic zeta functions and a Harder-type formula. As an application of the
results above we can give an analogue of Harder’s point counting formula from [ 88 ] (see
also [ 12 , Corollary 2.11]). We do this by exhibiting an identity in the Grothendieck ring of varieties. Note however that the ground eld cannot be chosen to be F푞, hence Harder’s formula cannot be obtained by applying the motivic measure
(236) # : K0 (Var/F푞) → Z : [푋] ↦→ #푋 (F푞).
푋 푛 푛 푋 푋 푛 For any variety , we denote the th symmetric Sym to be /Sym푛, where Sym푛 is the symmetric group of 푛 letters. All the symmetric powers can be put together to give
Kapranov’s motivic zeta function, as introduced in [ 100 , §1.3]: ∑︁ 푛 푛 (237) ZKap (푋, 푡) := [Sym 푋]푡 ∈ K0 (Var/푘)[[푡]]. 푛≥0 This is a universal version of the Hasse–Weil zeta function, valid for arbitrary ground elds, where the counting measure for 푘 = F푞 gives the usual Hasse–Weil zeta function.
The following theorem is due to Kapranov [ 100 , Theorem 1.1.9], and shows how the motivic zeta function has properties similar to the usual Hasse–Weil zeta function.
Theorem 7.3.1 (Kapranov). Let 퐶 be a smooth curve. The motivic zeta function of 퐶 is a rational function of the following form:
퐹2푔 (푡) (238) Z (퐶, 푡) = , Kap (1 − 푡)(1 − L푡) where 퐹2푔 (푡) is a polynomial of degree 2푔. Moreover ZKap (퐶, 푡) (resp. 퐹2푔 (푡)) satises the functional equation 1 (239) Z (퐶, 푡) = L푔−1푡 2푔−2 Z 퐶, Kap Kap L푡 resp. 1 (240) 퐹 (푡) = L푔푡 2푔퐹 . 2푔 2푔 L푡 74 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
We can rearrange and reinterpret the terms in the motivic zeta function as follows.
Proposition 7.3.2. We have the identity 푔−2 ! ∑︁ 푖 푖 3푔−2푖−3 푔−1 푔−1 ZKap (퐶, L) = [Sym 퐶](L + L ) + [Sym 퐶]L =0 (241) 푖 푔−2 ∑︁ [Jac(퐶)]L2푔−1 1 L푔 + [Jac(퐶)][P푔−푖−2]L2푔−2−푖 + − 1 − L 1 − L (1 − L2) 푖=0 in K0 (Var/푘).
4 Proof. Using Proposition 7.1.5 we have the identity (242) 푔−2 ! 푔−2 ! ∑︁ 푖 푖 푔−1 푔−1 ∑︁ 2푔−푖−2 2푔−푖−2 ZKap (퐶, L) = [Sym ]L + [Sym 퐶]L + [Sym 퐶]L 푖=0 푖=0 ! ∑︁ + [Sym푖 퐶]L푖 푖 ≥2푔−1 푔−2 ! 푔−2 ! ∑︁ ∑︁ = [Sym푖 ]L푖 + [Sym푔−1 퐶]L푔−1 + [Sym푖 퐶]L3푔−2푖−3 + [Jac퐶][P푔−푖−2]L2푔−푖−2 푖=0 푖=0 ! ∑︁ + [Jac퐶][P푖−푔]L푖 . 푖 ≥2푔−1 It now su ces to see that ! ∑︁ [Jac퐶]L2푔−1 1 L푔 (243) [Jac퐶][P푖−푔]L푖 = − 1 − L 1 − L 1 − L2 푖 ≥2푔−1 which follows from an immediate verication.
The following lemma is also an immediate verication.
Lemma 7.3.3. We have the identity 푔−2 ∑︁ (244) L2푔−푖−2 (1 − L푔−푖−1)(1 − L2) + L2푔−1 (1 + L − L푔) = L푔 푖=0 in K0 (Var/푘).
Using Proposition 7.3.2 and Lemma 7.3.3 we can then obtain the following corollary to
Theorem F . The right-hand side is an element of K0 (Var/푘) (and not some completion) by
Theorem 7.3.1 , so that the rst term of the right-hand side is in fact 퐹2푔 (L).
Corollary 7.3.4. We have the identity 2 2 푔 (245) (1 − L)(1 − L )[M퐶 (2, L)] = (1 − L)(1 − L ) ZKap (퐶, L) − L [Jac퐶] in K0 (Var/푘).
푔−푖−2 1−L푔−푖−1 Proof. It su ces to rewrite [P ] as 1−L and then multiply both sides of ( 241 ) 2 with (1 − L)(1 − L ), and apply Lemma 7.3.3 .
4 One avoids manipulating innite sums of L by observing that ZKap (퐶, 푡) is a rational function in 푡, and only nitely many copies of powers of L are contributing to each power of 푡 before evaluation. GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 75
This identity is then the analogue of Harder’s formula from [ 12 , Corollary 2.11] using the functional equation for the zeta function.
Appendix A. Some explicit calculations for periods In this section we collect some observations about the (non-)vanishing and values of certain quantum periods and periods. These explain some of the patterns which one can
observe in Tables 1 and 2 , and conrm further the (conjectural) agreement of the quantum periods of M퐶 (2, L) and the periods of the graph potentials. 76 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY 01012 587002910880 13760135544283128000 0 22961207808000 0 95981760 0 1728 0 1 10 푝 푔 5607950 574762007983708676751808000 0 4248683870158728000 22317642547200 2006716647119184000 15571407667200 133905855283200 9963493478400 5892216422400 0 2131746903000 795162277629720000 0 237075779068128000 299893321728000 0 74995200 295708763750400 3112327680000 1984905402480000 56589120 40424237568 41211916193448000 1378578432000 232243200 0 40763520 115772770713600 11300889600000 928972800 0 27518400 1536 788889024 0 16853760 445839609600 68263641600 1344 0 0 1175731200 1152 0 0 1 435456000 16003008 8769600 960 0 1 9 3265920 11520 768 0 343000 1 8 23040 576 0 1 7 8000 384 0 1 6 216 0 1 5 8 1 4 1 3 2 0 푝 01012 587002910880 36154232000 13760135544283128000 0 22961207808000 0 95981760 0 1728 0 1 10 푝 푔 2 5607950 5747620079414146000 0 7994421145174464000 0 0 4323671149117320000 0 0 0 2211592605702480000 0 0 0 1062973988196120000 15571407667200 0 401643111149280000 4132896768000 9963493478400 80306439693480000 0 0 3976941969000000 5953528627200 0 0 74995200 3418888704000 0 0 0 0 56589120 0 1915060224000 0 805929062400 40763520 0 1536 0 136604160000 27518400 1513881600 0 1344 0 0 18144000 0 1152 0 0 1 12640320 0 0 960 0 1 6350400 9 0 768 0 645120 1 8 0 576 0 1 7 0 576 0 1 6 384 0 1 5 0 1 4 1 3 2 푝 0 4 푝 2 푝 6 푝 4 al 2. Table al 1. Table 푝 6 푝 8 푝 eidsqec o h vngahpotential graph even the for sequence Period eidsqec o h d rp potential graph odd the for sequence Period 8 unu eid for periods Quantum 푝 10 푝 10 푝 12 푝 12 M 퐶 ( 2 , L 푝 ) 14 푝 14 푝 16 푝 푝 18 16 GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 77
First we will consider the second and fourth (quantum) period, in Appendix A.3 we discuss some patterns which hold more generally.
A.1. Second quantum period. The rst interesting quantum period of M퐶 (2, L) is 푝2. We will show that it is always zero, as soon as 푔 ≥ 3. This is suggested by the calculation
in Proposition A.1.1 on the level of periods.
In Appendix A.3 we will extend this vanishing result for the period 휋2 to the vanishing of 휋4푛+2 for all 푛 ≥ 0 and 푔 0.
Proposition A.1.1. Let 훾 be a trivalent graph of genus 푔. The second period of 푊e훾,1 associated to M퐶 (2, L) is given by ( 8 푔 = 2 (246) 휋2 (푊e훾,1) = 0 푔 ≥ 3.
Proof. For 푔 = 2 this is given in Example 2.3.2 , but it is also immediate from taking the
square of ( 37 ) and pairing inverse monomials together. Similarly, for 푔 ≥ 3 we are pairing inverse monomials together by squaring and taking the constant term. But because we assume that the graph 훾 is connected, the monomials contributed by the unique colored vertex have no corresponding inverse monomial in the graph potential.
The second quantum period on the other hand counts the number of lines (in the polar- ization given by the ample generator Θ of Pic M퐶 (2, L)) through a generic point of the variety M퐶 (2, L). We wish to explain the vanishing for 푔 ≥ 3 geometrically, by using the explicit geometry in the hyperelliptic case.
Remark A.1.2. For 푔 = 2 we have that M퐶 (2, L) is a Fano 3-fold, and its second (regular-
ized) quantum period can be read o in [ 48 , §6]. More generally, in Corollary 3.3.10 we have obtained an expression of (the inverse Laplace transform of) the period sequence which agrees with that of the quantum period sequence from §6 of op. cit.
Geometric setup. In [ 57 ] the relationship between the geometry of the intersection of 2 quadrics (given by quadratic forms 푞1 and 푞2, determined by the hyperelliptic curve 퐶) in P2푔+1 and hyperelliptic curves of genus 푔 is discussed. We will use this to sketch a geometric argument for the vanishing of the quantum periods for 푔 ≥ 3.
By Theorem 1 of op. cit. M퐶 (2, L) is isomorphic to the closed subvariety of Gr(푔 −1, 2푔 +2) parametrizing P푔−2 ⊆ P2푔+1 contained in the quadrics. Likewise, by Theorem 2 of op. cit. the Jacobian of퐶 (or rather, the torsor Pic푔 퐶) is isomorphic to a closed subvariety of Gr(푔, 2푔+2) parametrizing P푔−1 ⊆ P2푔+1 contained in the quadrics. These two closed subvarieties are given as the intersection OGr(푘, 2푔 + 2;푞1) ∩ OGr(푘, 2푔 + 2;푞2) for OGr(푘, 2푔 + 2;푞푖 ) the orthogonal Grassmannian embedded in Gr(푘, 2푔 + 2), for 푘 = 푔 − 1,푔. Now consider the incidence correspondence
Fl(푔 − 1,푔, 2푔 + 2) Gr(푔 − 1, 2푔 + 2) (247) Gr(푔, 2푔 + 2)
The vertical map is a P푔−1-bundle, the horizontal map is a P푔+3-bundle. 78 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
The restriction of the vertical map to Pic푔 (퐶) corresponds to the interpretation of the
incidence correspondence from [ 57 , Remark 5.10(I)]. For this interpretation we will con- sider M퐶 (2, L) for a line bundle L of degree 2푔 + 1, and consider non-split extensions ∨ (248) 0 → 푗 → V푗 → 푗 ⊗ L → 0 for 푗 ∈ Pic푔 퐶, which are always stable. 1 2 ∨ The 푔 − 1-dimensional projective space P푗 := P(H (퐶, 푗 ⊗ L) ) parametrizing these extensions gives rise to a P푔−1-bundle P. If we let M be the Poincaré bundle on 퐶 ×Pic푔 (퐶), 1 ⊗2 then P P퐶 (E) where E := R 푝2,∗ (M (−1)). Then [ 57 , Remark 5.10(I)] gives the following
Lemma A.1.3. We have that P ts into the ber product diagram P Fl(푔 − 1,푔, 2푔 + 2) (249) Pic푔 (퐶) Gr(푔, 2푔 + 2)
The horizontal map in the incidence correspondence gives a morphism 푓 : P → M퐶 (2, L). We can consider other incidence correspondences. Before proving the interpretation
in Proposition A.1.5 , we will give an easier description for the Fano variety of lines on M퐶 (2, L). This is a natural continuation of the discussion in [ 57 , Remark 5.10].
Proposition A.1.4. The Fano variety of lines F(M퐶 (2, L)) ts in the ber product diagram
F(M퐶 (2, L)) Fl(푔 − 2,푔, 2푔 + 2) (250) . Pic푔 (퐶) Gr(푔, 2푔 + 2) In particular, it is a P푔−1-bundle over Pic푔 (퐶).
Proof. The image of Pic푔 (퐶) in Gr(푔, 2푔 + 2) consists of 푔-dimensional subspaces which are isotropic for the quadratic forms 푞1 and 푞2 dened on the 2푔 + 2-dimensional vector space 푉 . Hence we can describe the ber product as
(251) {퐸2 ⊂ 퐸1 ⊂ 푉 | dim 퐸2 = 푔 − 2, dim 퐸1 = 푔, 푞1 (퐸1) = 푞2 (퐸1) = 0}.
Recall that F(OGr(푔 − 1, 2푔 + 2;푞푖 )) is again a homogeneous variety, namely the quo- tient SO2푔+2/P푔−2,푔, where P푔−2,푔 denotes the parabolic subgroup
푔 (252) . 1 푔 − 2 푔 + 1 Now observe that F(M퐶 (2, L)) = F(OGr(푔 − 1, 2푔 + 2;푞1) ∩ OGr(푔 − 1, 2푔 + 2;푞2)) (253) ⊂ F(Gr(푔 − 1, 2푔 + 2)) Fl(푔 − 2,푔, 2푔 + 2) allows us to describe the Fano variety of lines using the homogeneous spaces SO2푔+2/P푔−2,푔,
giving the description from ( 251 ).
From Section 5.5 we recall that to compute the second quantum period, we are interested in the moduli space M퐶 (2, L)0,1,훽 for 훽 = 2 (because the Fano index of M퐶 (2, L) is 2). Again we wish to use the geometry of the ambient partial ag varieties, now using the isomorphism Gr(푔 − 1, 2푔 + 2)0,1,2 Fl(푔 − 2,푔 − 1,푔, 2푔 + 2). This allows us to obtain the
following identication, whose proof is similar to that of Proposition A.1.4 . GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 79
Proposition A.1.5. The variety M퐶 (2, L)0,1,훽 ts in the ber product diagram
M퐶 (2, L)0,1,2 Fl(푔 − 2,푔 − 1,푔, 2푔 + 2)
(254) ev . Pic푔 (퐶) Gr(푔, 2푔 + 2)
Summarizing, we have the following picture (255) 푓1 M퐶 (2, L)0,1,2 Fl(푔 − 2,푔 − 1,푔, 2푔 + 2) 푓2
ev P Fl(푔 − 1,푔, 2푔 + 2) 푓
M퐶 (2, L) Gr(푔 − 1, 2푔 + 2)
Pic푔 (퐶) Gr(푔, 2푔 + 2).
To prove the vanishing of the second quantum period for 푔 ≥ 3, we need the following property of the evaluation morphism.
Proposition A.1.6. The morphism ev : M퐶 (2, L)0,1,2 → M퐶 (2, L) factors through the natural map 푓 : P → M퐶 (2, L) from Lemma A.1.3 .
Proof. It su ces to observe that the image of 푓2 ◦ 푓1 is contained in P ⊂ Fl(푔 − 1,푔, 2푔 + 2),
which can be proven along the lines of Proposition A.1.4 .
Hence we obtain the following corollary, which is an independent proof of the vanishing
result from Proposition A.1.1 .
Corollary A.1.7. Let 푔 ≥ 3. Then 푝2 = 0.
Proof. Because the image of M퐶 (2, L)0,1,2 → M퐶 (2, L) under the evaluation morphism is contained in the image of P, we have that the codimension of the image is at least 푔 − 2. Hence the second quantum period vanishes for 푔 ≥ 3, by the denition of the quantum period.
A.2. Fourth period. The next interesting period is 휋4. In the case we can give a closed formula for it.
Proposition A.2.1. The fourth coe cient of the period of the potential 푊e훾,1 associated to M퐶 (2, L) is given by ( 216 푔 = 2 (256) 휋4 (푊e훾,1) = 192(푔 − 1) 푔 ≥ 3.
Proof. The total degree of a monomial in the graph potential is either 3 or −1, except for two monomials which are contributed by the (by assumption) unique colored vertex, whose total degrees are 1 and −3. Let us moreover consider a graph potential associated to a graph without loops, so that all coe cients are 1. 80 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
To understand the constant term of the fourth power of the graph potential, we consider the monomials of degree ±3. To obtain a constant term, we need to multiply this monomial with either a monomial of degree ∓3, or with 3 monomials of degree ∓1. If 푔 = 2, then the colored vertex and uncolored vertex share the same variables, so it is possible to multiply monomials of degree ±3 and get a constant term. But this is impossible if 푔 ≥ 3, and hence we assume from now on that we only multiply monomials of degree ±3 with monomials of degree ∓1. There are 2푔 − 2 monomials of degree ±3, one for each vertex. There are 4 ways to cancel this monomial: one way is by using the 3 distinct monomials contributed by the same
vertex in the vertex potentials ( 35 ). The other three ways are by considering the edge
potential as in ( 82 ) and ( 83 ) for the 3 edges going out of the vertex and multiplying the chosen monomial with one monomial from the same vertex and two monomials from the neighbor. There are 4! choices for doing this, hence the statement follows.
A.3. General patterns. The following result is completely general, and immediate from the denition ( 177 ). Because M퐶 (2, L) has index 2, we have that h훽, −K푋 i is always divisible by 2.
Lemma A.3.1. The odd-indexed coe cients of the quantum period of M퐶 (2, L) vanish: 푝2푘+1 (M퐶 (2, L)) = 푐2푘+1 (M퐶 (2, L)) = 0.
On the other hand, for the periods of graph potentials we have the following result. It follows immediately from the observation that the total degrees of the monomials in the graph potentials are ±1, ±3, so that no product of an odd number of them can have degree 0.
Lemma A.3.2. For any odd or even coloring of a graph 훾 without leaves the odd-indexed 2푘+1 coe cients of the period of the graph potential vanish: [푊훾,푐 (푧) ]푧0 = 0.
This explains why in Tables 1 and 2 we have not listed the odd periods. For the case of
an odd coloring this is consistent with Proposition 5.5.1 which gives the equality of the
periods from Lemmas A.3.1 and A.3.2 , but without the dependence on Conjecture C . We also have the following vanishing result.
Lemma A.3.3. For a constant coloring of a graph훾 without leaves non-vanishing of the 푘th 푘 coe cient of the period of the graph potential [푊훾,0 (푧) ]푧0 ≠ 0 implies the divisibility of 푘 by 4.
Proof. All monomials of the even graph potential 푊 := 푊훾,0 have total degree 3 and −1, so 푊 푘 is homogeneous of degree −푘 ∈ Z/4. On the other hand, the constant term 푧0 is homogeneous of degree 0 ∈ Z/4.
Finally we have the following general result, explaining why the lower triangular parts of the tables agree.
Proposition A.3.4. Let 푔 ≥ 2 and 푘 < 2푔 − 2. Then the 푘th coe cient of the period of the graph potential for the even coloring equals that of the odd coloring.
Observe that if 푘 is odd then we already knew that the 푘th period is zero in both cases.
Proof. By Corollary 2.3.4 we can consider a single graph 훾 of genus 푔, with an even
(resp. odd) coloring. This gives the graph potentials푊0 (resp.푊1). Moreover by Lemma 2.1.8 GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 81 we have that the action of C1 (훾, Z/2Z) preserves the graph potential up to biregular automorphism of the torus. By assumption we have that 푘 is strictly less than the number of vertices in the graph 훾. If 푚1 ···푚푘 is a product of monomials (which can also be considered as a sum of vectors 푁 푊 푘 푊 푘 in 훾 ) contributing to the constant term of 1 , taken from the product expansion of 1 , it cannot use monomials from every vertex of the graph 훾.
Let 푣 be an unused vertex. By the action of C1 (훾, Z/2Z) we can assume that the colored vertex is 푣, without changing the value of 푚1 ···푚푘 , as by assumption this is a constant and hence invariant under biregular automorphisms of the torus. Therefore this is the 푊 푘 same contribution as obtained from 0 , and we are done.
Combining this with Lemma A.3.3 we get
Corollary A.3.5. Let 푘 ≥ 0. Let 푔 ≥ 2. If 4푘 + 2 < 2푔 − 2 then the 4푘 + 2th period of the graph potential in the odd case vanishes.
Appendix B. Random walks on lattices
In Section 2 we dened a class of Laurent polynomials and subsequently discussed their properties. In this appendix we reinterpret our results in the language of random walks on lattices. We rst visualize the situation of random walks and the questions we discuss. Consider 푑 a lattice 푁 Z .A random walk is given by a nite collections of directions 푣푖 ∈ 푁 and probabilities 푝푖 ∈ [0, 1] whose sum is 1. On this lattice every agent makes independent moves. In the origin you only see the moments when an agent is at the origin. So mathe- matically for each agent you know only a sequence of numbers 휋푚 which is the probability to be back at the origin at moment 푚. We have the following natural question:
Question B.0.1. Can you guess a shape of the random walk if you only know the num- bers 휋푚?
We now formalize some denitions and relate it to the graph potentials we considered
in Section 2 . We refer the reader to [ 121 , §1.1] for denitions of random walks and their moments. 푑 푑 Consider a lattice 푁 Z in a 푑-dimensional real vector space 푁R ' R . Consider the following element in the group ring of the lattice 푁 . ∑︁ 푣 (257) 푊 (푧1, . . . , 푧푑 ) = 푝푣푧 , 푣∈Z푑 Assume the following: Í 푝 푝 푣 푑 • 푣∈Z푑 푣 = 1 and 푣 ≥ 0 for all ∈ Z . ◦ • the convex hull of 푣’s such that 푝푣 ≠ 0 is a bounded polytope 푃 (푊 ) in 푁R. • the origin is an interior point of 푃. With the assumptions as above, 푊 can be considered as a random walk on a lattice 푁 . The probability of the walk returning to the origin after 푚 steps is given by the 푚th period 휋푚 (푊 ) of the Laurent polynomial 푊 . Recall from Section 2.3 that this is just the 푚 constant term [푊 ]0 of the 푚th power of 푊 . We need to put conditions on the probability distribution such that an agent on a random walk returns back to the origin. The following condition is just a reworking of the condition that the expectation of 푊 is zero. 82 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
푑 푑 × 푑 Let 푢1, . . . ,푢푑 be coordinates on C . The exponential map C → (C ) is an étale cov- 푑 푑 푑 × 푑 푑 ering and identies R ⊂ C with R+ ⊂ (C ) . Let 푧푖 = exp(푢푖 ) and we identify R with R푑 via the exponential map. Observe that the partial derivative 휕 coincides with + 휕푢푖 the pullback 푧 휕 . 푖 휕푧푖 Denition B.0.2. A random walk 푊 on a lattice 푁 is balanced if 푐 = (1, 1,..., 1) ∈ 푁 is 푑 푧 휕 푊 푐 the unique point of R+ such that all the logarithmic derivatives 푖 푧푖 of vanish at and 2 further 푐 is also a Morse point, i.e. the Hessian matrix 휕 푊 of 푊 is non-degenerate. The 휕푢푖 휕푢푗 polytope 푃◦ (푊 ) of 푊 is called the Newton polytope.
Let Λ and Λ0 be two lattices and we have two balanced random walks 푊 and 푊 0 in Λ and Λ0 respectively. We dene the following:
Denition B.0.3. Two random walks 푊 and 푊 0 0 • are trivially equivalent if we can nd a linear map 휙 : Λ → Λ ⊗Z Q such that the 0 random walks are related by 휙∗푊 = 푊 ; • have the same shape if the map 휙 matches the Newton polytope associated to 푊 to the polytope associated to 푊 0. 0 • have equal moments if 휋푚 (푊 ) = 휋푚 (푊 ) for all 푚 ∈ N.
We ask the following questions:
Question B.0.4. (1) Can two balanced random walks have equal moments, whilst not being trivially equivalent or be of di erent shape?
(2) Can we nd a collection of balanced random walks {푊1, . . . ,푊푘 } such that all give the same period sequence but pairwise they are of di erent shape? What is a lower bound for such 푘?
We now reinterpret our results from Sections 4 and 5 to answer part (1) of Question B.0.4 a rmatively.
Construction B.0.5. We rst give a construction of such balanced random walks from
the graph potentials we considered in Section 2 . Let (훾, 푐) be a colored trivalent graph of 3푔−3 genus 푔. Consider lattice 푁훾 Z as in Section 2 and the graph potential푊e훾,푐 associated
to (훾, 푐) as in ( 33 ) recalled below:
∑︁ ∑︁ 푠푖 푠푗 푠푘 푊 := 푥 (−1) 푥 (−1) 푥 (−1) (258) e훾,푐 푖 푗 푘 푣∈푉 ⊕3 (푠푖,푠푗 ,푠푘 ) ∈F2 푠푖 +푠푗 +푠푘 =푐 (푣) where 푥푖, 푥 푗, 푥푘 are variable attached to the edges 푒푖, 푒푗, 푒푘 adjacent to the vertex 푣. If 푒푖 = 푒푗 , this implies a loop at vertex 푣. Now consider 1 (259) 푊 훾,푐 := 푊e훾,푐 8(푔 − 1)
◦ Since by construction the Newton polytope 푃 (푊e훾,푐 ) of the Laurent polynomial 푊e훾,푐 is a 푋 re exive polytope and also the dual of the moment polytope of the toric variety 푃훾,푐,푀훾 .
Moreover, by [ 74 ], we know that the 푊 훾,푐 is balanced. Hence we get
Proposition B.0.6. For each colored trivalent graph (훾, 푐), the Laurent polynomial 푊 훾,푐 denes a balanced random walk in the lattice 푁훾 . GRAPH POTENTIALS AND MODULI SPACES OF BUNDLES ON CURVES 83
Now we investigate the question of comparing periods of di erent graph potentials in
Section 2 . We can reinterpret Corollary 2.3.4 as
Proposition B.0.7. Let (훾, 푐) and (훾 0, 푐 0) be two colored trivalent graphs with the same parity of the number of colored vertices, then for 푚 ∈ N, their probability of the random walks 푊 훾,푐 and 푊 훾0,푐0 returning to the origin after 푚 steps are the same. √ Remark B.0.8. Observe that there exists approximately exp(퐶 푔) trivalent graphs of genus 푔, where 퐶 is a constant. And each one of them along with a choice of coloring produces balanced random walks 푊 훾,푐 . Moreover depending on the parity they have the same 휋푚’s.
Next we ask the question whether 푊 훾,푐 and 푊 훾,푐 having equal moments implies that
they are trivially equivalent or have the same shape, as in part (1) of Question B.0.4 . We
reinterpret Theorem 4.2.3 as follows to address this.
Theorem B.0.9. Let (훾, 푐) be a colored trivalent graph with a separating edge, and let (훾 0, 푐 0) be a another colored trivalent graph with no separating edge, having the same number of colored vertices as 훾. Then 푊 훾,푐 and 푊 훾0,푐0 do not have the same shape.
푊 푊 0 0 푋 Proof. If 훾,푐 and 훾 ,푐 have the same shape, this implies that the toric varieties 푃훾,푐 ,푀훾
푋 and 푃훾0,푐0,푀훾0 that we considered in Section 4 are isomorphic. However, by Theorem 4.2.3 푋 푋 gives a contradiction, since 푃훾0,푐0,푀훾0 has terminal singularities whereas 푃훾,푐 ,푀훾 is non- terminal. 84 PIETER BELMANS, SERGEY GALKIN, AND SWARNAVA MUKHOPADHYAY
References [1] Casim Abbas. An introduction to compactness results in symplectic eld theory. Springer, Heidelberg, 2014, pp. viii+252. isbn: 978-3-642-31543-5. [2] Mohammed Abouzaid, Denis Auroux, and Ludmil Katzarkov. “Lagrangian brations on blowups of toric varieties and mirror symmetry for hypersurfaces.” In: Publ. Math. Inst.
Hautes Études Sci. 123 (2016), pp. 199–282. arXiv: 1205.0053v4 . [3] Lowell Abrams. “Two-dimensional topological quantum eld theories and Frobenius alge- bras.” In: J. Knot Theory Ramications 5.5 (1996), pp. 569–587. [4] Sharad Agnihotri and Christopher Woodward. “Eigenvalues of products of unitary matrices
and quantum Schubert calculus.” In: Math. Res. Lett. 5.6 (1998), pp. 817–836. arXiv: alg-
geom/9712013v1 . [5] Mohammad Akhtar, Tom Coates, Sergey Galkin, and Alexander M. Kasprzyk. “Minkowski polynomials and mutations.” In: SIGMA Symmetry Integrability Geom. Methods Appl. 8 (94
2012). with an appendix (690 pages), 17 pages. arXiv: 1212.1785v1 . [6] M. F. Atiyah and R. Bott. “The Yang-Mills equations over Riemann surfaces.” In: Philos. Trans. Roy. Soc. London Ser. A 308.1505 (1983), pp. 523–615. [7] Michael Atiyah. “New invariants of 3- and 4-dimensional manifolds.” In: The mathematical heritage of Hermann Weyl (Durham, NC, 1987). Vol. 48. Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, RI, 1988, pp. 285–299. [8] Michael Atiyah. “Topological quantum eld theories.” In: Inst. Hautes Études Sci. Publ. Math. 68 (1988), 175–186 (1989). [9] Denis Auroux, Ludmil Katzarkov, and Dmitri O. Orlov. “Mirror symmetry for del Pezzo surfaces: vanishing cycles and coherent sheaves.” In: Invent. Math. 166.3 (2006), pp. 537–582.
arXiv: math/0506166v1 . [10] Matthew Robert Ballard. “Wall crossing for derived categories of moduli spaces of sheaves
on rational surfaces.” In: Algebr. Geom. 4.3 (2017), pp. 263–280. arXiv: 1412.4424v1 . [11] Sebastian del Baño. “On the Chow motive of some moduli spaces.” In: J. Reine Angew. Math. 532 (2001), pp. 105–132. [12] Sebastian del Baño. “On the motive of moduli spaces of rank two vector bundles over a curve.” In: Compositio Math. 131.1 (2002), pp. 1–30. [13] Francesco Bastianelli, Pieter Belmans, Shinnosuke Okawa, and Andrea T. Ricol. Indecom-
posability of derived categories in families. arXiv: 2007.00994v1 . [14] Victor Batyrev. “Dual polyhedra and mirror symmetry for Calabi-Yau hypersurfaces in
toric varieties.” In: J. Algebraic Geom. 3.3 (1994), pp. 493–535. arXiv: alg-geom/9310003 . [15] Victor Batyrev. “Toric degenerations of Fano varieties and constructing mirror manifolds.” In: The Fano Conference. Univ. Torino, Turin, 2004, pp. 109–122. [16] Victor Batyrev, Ionuţ Ciocan-Fontanine, Bumsig Kim, and Duco van Straten. “Conifold transitions and mirror symmetry for Calabi-Yau complete intersections in Grassmannians.”
In: Nuclear Phys. B 514.3 (1998), pp. 640–666. arXiv: alg-geom/9710022 . [17] Victor Batyrev, Ionuţ Ciocan-Fontanine, Bumsig Kim, and Duco van Straten. “Mirror sym- metry and toric degenerations of partial ag manifolds.” In: Acta Math. 184.1 (2000), pp. 1–39.
arXiv: math/9803108v1 . [18] Victor Batyrev and Maximilian Kreuzer. “Conifold degenerations of Fano 3-folds as hyper- surfaces in toric varieties.” In: Strings, gauge elds, and the geometry behind. World Sci. Publ.,
Hackensack, NJ, 2013, pp. 187–212. arXiv: 1203.6058 . [19] Dave Bayer and David Eisenbud. “Graph curves.”In: Adv. Math. 86.1 (1991). With an appendix by Sung Won Park, pp. 1–40. [20] Arnaud Beauville. “Vector bundles on curves and generalized theta functions: recent results and open problems.” In: Current topics in complex algebraic geometry (Berkeley, CA, 1992/93). Vol. 28. Math. Sci. Res. Inst. Publ. Cambridge Univ. Press, Cambridge, 1995, pp. 17–33. [21] Arnaud Beauville and Yves Laszlo. “Conformal blocks and generalized theta functions.” In:
Comm. Math. Phys. 164.2 (1994), pp. 385–419. arXiv: alg-geom/9309003v1 . [22] Kai Behrend and Ajneet Dhillon. “On the motivic class of the stack of bundles.” In: Adv.
Math. 212.2 (2007), pp. 617–644. arXiv: math/0512640v1 . [23] Alexander A. Belavin, Alexander M. Polyakov, and Alexander B. Zamolodchikov. “Innite conformal symmetry in two-dimensional quantum eld theory.” In: Nuclear Phys. B 241.2 (1984), pp. 333–380. REFERENCES 85
[24] Prakash Belkale. “Quantum generalization of the Horn conjecture.” In: J. Amer. Math. Soc.
21.2 (2008), pp. 365–408. arXiv: math/0303013 . [25] Pieter Belmans. “Semiorthogonal decompositions for moduli of sheaves on curves.” In: Oberwolfach Workshop 1822, Interactions between Algebraic Geometry and Noncommutative Algebra. Oberwolfach Reports. Vol. 15. 2018, pp. 1473–1476. [26] Pieter Belmans, Lie Fu, and Theo Raedschelders. Derived categories of ips and cubic hyper-
surfaces. arXiv: 2002.04940v1 . [27] Pieter Belmans, Lie Fu, and Theo Raedschelders. “Hilbert squares: derived categories and
deformations.” In: Selecta Math. (N.S.) 25.3 (2019), Paper No. 37, 32. arXiv: 1810.11873v2 . [28] Pieter Belmans, Sergey Galkin, and Swarnava Mukhopadhyay. “Examples violating Goly- shev’s canonical strip hypotheses.” In: Exp. Math. (2020). published online ahead of print.
arXiv: 1806.07648v1 . [29] Pieter Belmans and Andreas Krug. Derived categories of (nested) Hilbert schemes. arXiv:
1909.04321v2 . [30] Pieter Belmans and Swarnava Mukhopadhyay. “Admissible subcategories in derived cate- gories of moduli of vector bundles on curves.” In: Adv. Math. 351 (2019), pp. 653–675. arXiv:
1807.00216v1 . [31] Arkady Berenstein, Sergey Fomin, and Andrei Zelevinsky. “Cluster algebras. III. Upper
bounds and double Bruhat cells.” In: Duke Math. J. 126.1 (2005), pp. 1–52. arXiv: math/
0305434v3 . [32] Arkady Berenstein and Andrei Zelevinsky. “Tensor product multiplicities, canonical bases
and totally positive varieties.” In: Invent. Math. 143.1 (2001), pp. 77–128. arXiv: math /
9912012 . [33] Aaron Bertram and András Szenes. “Hilbert polynomials of moduli spaces of rank 2. Vector bundles. II.” In: Topology 32.3 (1993), pp. 599–609. [34] Indranil Biswas, Tomas L. Gomez, and Kyoung-Seog Lee. Semi-orthogonal decomposition of
symmetric products of curves and canonical system. arXiv: 1807.10702v2 . [35] Franziska Bittner. “The universal Euler characteristic for varieties of characteristic zero.” In:
Compos. Math. 140.4 (2004), pp. 1011–1032. arXiv: math/0111062v1 . [36] Alexey I. Bondal and Sergey Galkin. “Mirror symmetry for minuscule varieties.” Unpublished
preprint. Some parts presented as [ 71 ]. 2011. IPMU: IPMU11-0101 . [37] Alexey I. Bondal and Mikhail M. Kapranov. “Representable functors, Serre functors, and reconstructions.” In: Izv. Akad. Nauk SSSR Ser. Mat. 53.6 (1989). translation in: Math. USSR-Izv. 35 (1990), no. 3, pp. 519–541, pp. 1183–1205, 1337. [38] Alexey I. Bondal, Michael Larsen, and Valery Lunts. “Grothendieck ring of pretriangulated
categories.” In: Int. Math. Res. Not. 29 (2004), pp. 1461–1495. arXiv: math/0401009v1 . [39] Alexey I. Bondal and Dmitri O. Orlov. “Derived categories of coherent sheaves.” In: Proceed- ings of the International Congress of Mathematicians, Vol. II (Beijing, 2002). Higher Ed. Press,
Beijing, 2002, pp. 47–56. arXiv: math/0206295v1 . [40] Alexey I. Bondal and Dmitri O. Orlov. “Reconstruction of a variety from the derived category
and groups of autoequivalences.” In: Compositio Math. 125.3 (2001), pp. 327–344. arXiv: alg-
geom/9712029v1 . [41] Alexey I. Bondal and Dmitri O. Orlov. Semiorthogonal decomposition for algebraic varieties.
1995. arXiv: alg-geom/9506012 . [42] Frédéric Bourgeois, Yakov Eliashberg, Helmut Hofer, Krzysztof Wysocki, and Eduard Zehn- der. “Compactness results in symplectic eld theory.” In: Geom. Topol. 7 (2003), pp. 799–888.
arXiv: math/0308183v3 . [43] Ragnar Buchweitz. “Maximal Cohen-Macaulay modules and Tate-cohomology over Goren- stein rings.” unpublished manuscript. 1987. [44] Daewoong Cheong and Changzheng Li. “On the conjecture O of GGI for 퐺/푃.” In: Adv.
Math. 306 (2017), pp. 704–721. arXiv: 1412.2475v5 . [45] Cheol-Hyun Cho and Yong-Geun Oh. “Floer cohomology and disc instantons of Lagrangian torus bers in Fano toric manifolds.” In: Asian J. Math. 10.4 (2006), pp. 773–814. arXiv:
math/0308225v1 . [46] Ionuţ Ciocan-Fontanine, Bumsig Kim, and Claude Sabbah. “The abelian/nonabelian corre- spondence and Frobenius manifolds.” In: Invent. Math. 171.2 (2008), pp. 301–343. 86 REFERENCES
[47] Tom Coates, Alessio Corti, Sergey Galkin, Vasily Golyshev, and Alexander Kasprzyk. “Mirror symmetry and Fano manifolds.” In: European Congress of Mathematics. Eur. Math. Soc.,
Zürich, 2013, pp. 285–300. arXiv: 1212.1722v1 . [48] Tom Coates, Alessio Corti, Sergey Galkin, and Alexander Kasprzyk. “Quantum periods for 3-
dimensional Fano manifolds.” In: Geom. Topol. 20.1 (2016), pp. 103–256. arXiv: 1303.3288v3 . [49] Tom Coates, Sergey Galkin, Alexander Kasprzyk, and Andrew Strangeway. “Quantum periods for certain four-dimensional Fano manifolds.” In: Exp. Math. 29.2 (2020), pp. 183–221.
arXiv: 1406.4891v3 . url: https://doi.org/10.1080/10586458.2018.1448018 . [50] David Cox. “Mirror symmetry and polar duality of polytopes.” In: Symmetry 7.3 (2015), pp. 1633–1645. [51] David Cox, John Little, and Henry Schenck. Toric varieties. Vol. 124. Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2011, pp. xxiv+841. isbn: 978-0-8218-4819-7. [52] John Alexander Cruz Morales and Sergey Galkin. “Upper bounds for mutations of potentials.” In: SIGMA Symmetry Integrability Geom. Methods Appl. 9 (2013), Paper 005, 13 pages. arXiv:
1301.4541v1 . [53] Aliakbar Daemi and Kenji Fukaya. “Atiyah-Floer conjecture: a formulation, a strategy of proof and generalizations.” In: Modern geometry: a celebration of the work of Simon Donaldson. Vol. 99. Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, RI, 2018, pp. 23–57. [54] Pierre Deligne. “Travaux de Shimura.” In: Séminaire Bourbaki, 23ème année (1970/71), Exp. No. 389. 1971, 123–165. Lecture Notes in Math., Vol. 244. [55] Christopher Deninger. “Deligne periods of mixed motives, 퐾-theory and the entropy of certain Z푛-actions.” In: J. Amer. Math. Soc. 10.2 (1997), pp. 259–281. [56] Christopher Deninger and Klaus Schmidt. “Expansive algebraic actions of discrete residually nite amenable groups and their entropy.” In: Ergodic Theory Dynam. Systems 27.3 (2007),
pp. 769–786. arXiv: math/0605723v2 . [57] U. V. Desale and S. Ramanan. “Classication of vector bundles of rank 2 on hyperelliptic curves.” In: Invent. Math. 38.2 (1976), pp. 161–185. [58] Robertus Dijkgraaf. “A geometrical approach to two-dimensional Conformal Field Theory.”
doctoralthesis. Utrecht University, 1989. url: https://dspace.library.uu.nl/handle/
1874/210872 . [59] S. K. Donaldson. “A new proof of a theorem of Narasimhan and Seshadri.” In: J. Di er-
ential Geom. 18.2 (1983), pp. 269–277. url: http://projecteuclid.org/euclid.jdg/
1214437664 . [60] Jean-Marie Drezet and M. S. Narasimhan. “Groupe de Picard des variétés de modules de brés semi-stables sur les courbes algébriques.” In: Invent. Math. 97.1 (1989), pp. 53–94. [61] David Eisenbud. “Homological algebra on a complete intersection, with an application to
group representations.” In: Trans. Amer. Math. Soc. 260.1 (1980), pp. 35–64. url: https:
//doi.org/10.2307/1999875 . [62] Yakov Eliashberg, Alexander Givental, and Helmut Hofer. “Introduction to symplectic eld theory.” In: Special Volume, Part II. GAFA 2000 (Tel Aviv, 1999). Oct. 2000, pp. 560–673.
arXiv: math/0010059v1 . [63] Gerd Faltings. “A proof for the Verlinde formula.” In: J. Algebraic Geom. 3.2 (1994), pp. 347– 374. [64] Theodore Faust and Christopher Manon. “The Gorenstein property for projective coordinate rings of the moduli of parabolic S퐿2-principal bundles on a smooth curve.” In: Electron. J.
Combin. 26.4 (2019), Paper No. 4.25, 12. arXiv: 1606.00507v2 . [65] Andreas Floer. “An instanton-invariant for 3-manifolds.” In: Comm. Math. Phys. 118.2 (1988), pp. 215–240. [66] Andreas Floer. “Morse theory for Lagrangian intersections.” In: J. Di erential Geom. 28.3 (1988), pp. 513–547. [67] Vladimir V. Fock and Alexander B. Goncharov. “Cluster ensembles, quantization and the
dilogarithm.” In: Ann. Sci. Éc. Norm. Supér. (4) 42.6 (2009), pp. 865–930. arXiv: math/0311245 . [68] Vladimir Fock and Alexander Goncharov. “Moduli spaces of local systems and higher Teichmüller theory.” In: Publ. Math. Inst. Hautes Études Sci. 103 (2006), pp. 1–211. arXiv:
math/0311149v4 . REFERENCES 87
[69] Anton Fonarev and Alexander Kuznetsov. “Derived categories of curves as components of
Fano manifolds.” In: J. Lond. Math. Soc. (2) 97.1 (2018), pp. 24–46. arXiv: 1612.02241v2 . [70] William Fulton and Rahul Pandharipande. “Notes on stable maps and quantum cohomology.” In: Algebraic geometry—Santa Cruz 1995. Vol. 62. Proc. Sympos. Pure Math. Amer. Math.
Soc., Providence, RI, 1997, pp. 45–96. arXiv: alg-geom/9608011v2 . [71] Sergey Galkin. Polynomial relations between holomorphic curves and discs. Nov. 26, 2014.
url: http://cms.unige.ch/math/tggroup/doku.php?id=tropgeofinal . [72] Sergey Galkin. “Торические вырождения многообразий Фано.” (Toric degenerations of
Fano manifolds). PhD thesis. Steklov Institute of Mathematics, 2008. url: http://www.mi.
ras.ru/~galkin/papers/disser.pdf . [73] Sergey Galkin. Small toric degenerations of Fano threefolds. IPMU 12-0121. Apr. 2008. arXiv:
1809.02705v1 . url: https://www.mi.ras.ru/~galkin .
[74] Sergey Galkin. The conifold point. arXiv: 1404.7388v1 . [75] Sergey Galkin, Vasily Golyshev, and Hiroshi Iritani. “Gamma classes and quantum cohomol- ogy of Fano manifolds: Gamma conjectures.” In: Duke Math. J. 165.11 (2016), pp. 2005–2077.
arXiv: 1404.6407v4 . [76] Sergey Galkin and Hiroshi Iritani. Gamma conjecture via mirror symmetry. Primitive Forms and Related Subjects – Kavli IPMU 2014, 55–115, Mathematical Society of Japan, Tokyo,
Japan, 2019. 2019. arXiv: 1508.00719v3 . [77] Sergey Galkin and Grigory Mikhalkin. unpublished. [78] Sergey Galkin and Grigory Mikhalkin. “Homotopy quantization of real plane curve enu- meration.” 2015. [79] Sergey Galkin and Alexandr Usnich. Laurent phenomenon for Ginzburg–Landau potential. IPMU 10-0100. Institute for the Physics and Mathematics of the Universe. 2010. IPMU:
10-0100 . url: http://db.ipmu.jp/ipmu/sysimg/ipmu/417.pdf . [80] Israel Gelfand, Mikhail M. Kapranov, and Andrei Zelevinsky. Discriminants, resultants and multidimensional determinants. Modern Birkhäuser Classics. Reprint of the 1994 edition. Birkhäuser Boston, Inc., Boston, MA, 2008, pp. x+523. isbn: 978-0-8176-4770-4. [81] Alexander Givental. “A mirror theorem for toric complete intersections.” In: Topological eld theory, primitive forms and related topics (Kyoto, 1996). Vol. 160. Progr. Math. Birkhäuser
Boston, Boston, MA, 1998, pp. 141–175. arXiv: alg-geom/9701016v2 . [82] Alexander B. Givental. “Equivariant Gromov-Witten invariants.” In: Internat. Math. Res.
Notices 13 (1996), pp. 613–663. arXiv: alg-geom/9603021 . [83] Tomás L. Gómez and Kyoung-Seog Lee. Motivic decompositions of moduli spaces of vector
bundles on curves. arXiv: 2007.06067v1 . [84] Ángel González-Prieto, Marina Logares, and Vicente Muñoz. “A lax monoidal topological quantum eld theory for representation varieties.” In: Bull. Sci. Math. 161 (2020), pp. 102871,
34. arXiv: 1709.05724v2 . [85] Mark Gross, Paul Hacking, and Sean Keel. “Birational geometry of cluster algebras.” In:
Algebr. Geom. 2.2 (2015), pp. 137–175. arXiv: 1309.2573v2 . [86] Mark Gross, Paul Hacking, Sean Keel, and Maxim Kontsevich. “Canonical bases for cluster
algebras.” In: J. Amer. Math. Soc. 31.2 (2018), pp. 497–608. arXiv: 1411.1394v2 . [87] Megumi Harada and Kiumars Kaveh. “Integrable systems, toric degenerations and Okounkov
bodies.” In: Invent. Math. 202.3 (2015), pp. 927–985. arXiv: 1205.5249v5 . [88] Günter Harder. “Eine Bemerkung zu einer Arbeit von P. E. Newstead.” In: J. Reine Angew. Math. 242 (1970), pp. 16–25. [89] Joe Harris and David Mumford. “On the Kodaira dimension of the moduli space of curves.” In: Invent. Math. 67.1 (1982). With an appendix by William Fulton, pp. 23–88. [90] Allen Hatcher and Willian Thurston. “A presentation for the mapping class group of a closed orientable surface.” In: Topology 19.3 (1980), pp. 221–237. [91] Franziska Heinloth. “A note on functional equations for zeta functions with values in Chow
motives.” In: Ann. Inst. Fourier (Grenoble) 57.6 (2007), pp. 1927–1945. arXiv: math/0512237v1 . [92] Hansol Hong, Yoosik Kim, and Siu-Cheong Lau. “Immersed two-spheres and SYZ with
Application to Grassmannians.” In: (May 2018). arXiv: 1805.11738v4 . [93] Kentaro Hori. “Mirror symmetry and quantum geometry.” In: Proceedings of the International Congress of Mathematicians, Vol. III (Beijing, 2002). Higher Ed. Press, Beijing, 2002, pp. 431–
443. arXiv: hep-th/0207068v1 . 88 REFERENCES
[94] Kentaro Hori and Cumrun Vafa. Mirror Symmetry. 2000. arXiv: hep-th/0002222v3 . [95] Johannes Huebschmann. “Poisson structures on certain moduli spaces for bundles on a
surface.” In: Ann. Inst. Fourier (Grenoble) 45.1 (1995), pp. 65–91. arXiv: dg-ga/9411009 . [96] Lisa C. Je rey and Jonathan Weitsman. “Bohr-Sommerfeld orbits in the moduli space of at connections and the Verlinde dimension formula.” In: Comm. Math. Phys. 150.3 (1992), pp. 593–630. [97] Lisa C. Je rey and Jonathan Weitsman. “Toric structures on the moduli space of at con- nections on a Riemann surface: volumes and the moment map.” In: Adv. Math. 106.2 (1994), pp. 151–168. [98] Qingyuan Jiang and Naichung Conan Leung. Derived category of projectivization and ops.
arXiv: 1811.12525v4 . [99] Jamie Judd and Konstanze Rietsch. The tropical critical point and mirror symmetry. arXiv:
1911.04463v2 . [100] Mikhail M. Kapranov. The elliptic curve in the S-duality theory and Eisenstein series for
Kac-Moody groups. arXiv: math/0001005v2 . [101] Anton Kapustin and Yi Li. “D-branes in Landau-Ginzburg models and algebraic geometry.”
In: J. High Energy Phys. 12 (2003), pp. 005, 44. arXiv: hep - th / 0210296 . url: https :
//doi.org/10.1088/1126-6708/2003/12/005 . [102] Ludmil Katzarkov. “Homological mirror symmetry and algebraic cycles.” In: Homological mirror symmetry. Vol. 757. Lecture Notes in Phys. Springer, Berlin, 2009, pp. 125–152. [103] Ludmil Katzarkov, Maxim Kontsevich, and Tony Pantev. “Bogomolov-Tian-Todorov theo- rems for Landau-Ginzburg models.” In: J. Di erential Geom. 105.1 (2017), pp. 55–117. arXiv:
1409.5996v2 . [104] Yujiro Kawamata. “On the extension problem of pluricanonical forms.” In: Algebraic geome- try: Hirzebruch 70 (Warsaw, 1998). Vol. 241. Contemp. Math. Amer. Math. Soc., Providence,
RI, 1999, pp. 193–207. arXiv: math/9809091 . [105] Young-Hoon Kiem and Jun Li. “Desingularizations of the moduli space of rank 2 bundles
over a curve.” In: Math. Ann. 330.3 (2004), pp. 491–518. arXiv: math/0402447 . [106] Bumsig Kim and Claude Sabbah. “Quantum cohomology of the Grassmannian and alternate
Thom-Sebastiani.” In: Compos. Math. 144.1 (2008), pp. 221–246. arXiv: math/0611475 . [107] Maxim Kontsevich. “Homological algebra of mirror symmetry.” In: Proceedings of the In- ternational Congress of Mathematicians, Vol. 1, 2 (Zürich, 1994). Birkhäuser, Basel, 1995,
pp. 120–139. arXiv: alg-geom/9411018v1 .
[108] Maxim Kontsevich. “Lectures at ENS Paris.” http://math.uchicago.edu/~drinfeld/
langlands/kontsevich.ps . 1998. [109] Maxim Kontsevich and Yuri I. Manin. “Gromov-Witten classes, quantum cohomology,
and enumerative geometry.” In: Comm. Math. Phys. 164.3 (1994), pp. 525–562. arXiv: hep-
th/9402147v2 . [110] Alexis Kouvidakis and Tony Pantev. “The automorphism group of the moduli space of
semistable vector bundles.” In: Math. Ann. 302.2 (1995), pp. 225–268. arXiv: alg- geom/
9306001v1 . [111] Maximilian Kreuzer and Harald Skarke. PALP: a Package for Analyzing Lattice Polytopes.
url: http://hep.itp.tuwien.ac.at/~kreuzer/CY/CYpalp.html . [112] Andreas Krug, David Ploog, and Pawel Sosna. “Derived categories of resolutions of cyclic
quotient singularities.” In: Q. J. Math. 69.2 (2018), pp. 509–548. arXiv: 1701.01331v2 . [113] Andreas Krug and Pawel Sosna. “On the derived category of the Hilbert scheme of points on
an Enriques surface.” In: Selecta Math. (N.S.) 21.4 (2015), pp. 1339–1360. arXiv: 1404.2105v1 . [114] Shrawan Kumar, M. S. Narasimhan, and A. Ramanathan. “Innite Grassmannians and moduli spaces of 퐺-bundles.” In: Math. Ann. 300.1 (1994), pp. 41–75. [115] Frank Kutzschebauch and Frank Loose. “Real analytic structures on a symplectic manifold.” In: Proc. Amer. Math. Soc. 128.10 (2000), pp. 3009–3016. [116] Alexander Kuznetsov. “Homological projective duality.” In: Publ. Math. Inst. Hautes Études
Sci. 105 (2007), pp. 157–220. arXiv: math/0507292v1 . [117] Alexander Kuznetsov. “Lefschetz decompositions and categorical resolutions of singulari-
ties.” In: Selecta Math. (N.S.) 13.4 (2008), pp. 661–696. arXiv: math/0609240v2 . REFERENCES 89
[118] Alexander Kuznetsov. “Semiorthogonal decompositions in algebraic geometry.” In: Proceed- ings of the International Congress of Mathematicians—Seoul 2014. Vol. II. Kyung Moon Sa,
Seoul, 2014, pp. 635–660. arXiv: 1404.3143v3 . [119] Alexander Kuznetsov and Maxim Smirnov. “On residual categories for Grassmannians.” In:
Proc. Lond. Math. Soc. (3) 120.5 (2020), pp. 617–641. arXiv: 1802.08097v3 . [120] Alexander Kuznetsov and Maxim Smirnov. Residual categories for (co)adjoint Grassmannians
in classical types. arXiv: 2001.04148v1 . [121] Gregory F. Lawler and Vlada Limic. Random walk: a modern introduction. Vol. 123. Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2010, pp. xii+364. isbn: 978-0-521-51918-2. [122] Kyoung-Seog Lee. Remarks on motives of moduli spaces of rank 2 vector bundles on curves.
arXiv: 1806.11101v2 . [123] Valery Lunts and Victor Przyjalkowski. “Landau-Ginzburg Hodge numbers for mirrors of
del Pezzo surfaces.” In: Adv. Math. 329 (2018), pp. 189–216. arXiv: 1607.08880v2 . [124] Jacob Lurie. “On the classication of topological eld theories.” In: Current developments in
mathematics, 2008. Int. Press, Somerville, MA, 2009, pp. 129–280. arXiv: 0905.0465v1 . [125] Travis Mandel. Fano mirror periods from the Frobenius structure conjecture. 2019. arXiv:
1903.12014v1 . [126] Travis Mandel and Helge Ruddat. “Descendant log Gromov-Witten invariants for toric varieties and tropical curves.” In: Trans. Amer. Math. Soc. 373.2 (2020), pp. 1109–1152. arXiv:
1612.02402 . [127] Travis Mandel and Helge Ruddat. Tropical quantum eld theory, mirror polyvector elds, and
multiplicities of tropical curves. 2019. arXiv: 1902.07183v1 . [128] Christopher Manon. “Coordinate rings for the moduli stack of 푆퐿2 (C) quasi-parabolic principal bundles on a curve and toric ber products.” In: J. Algebra 365 (2012), pp. 163–183.
arXiv: 1105.2045v2 . [129] Christopher Manon. “The algebra of conformal blocks.” In: J. Eur. Math. Soc. (JEMS) 20.11
(2018), pp. 2685–2715. arXiv: 0910.0577v7 . [130] Hannah Markwig and Johannes Rau. “Tropical descendant Gromov-Witten invariants.” In:
manuscripta math. 129.3 (2009), pp. 293–335. arXiv: 0809.1102 . [131] Grigory Mikhalkin. “Enumerative tropical algebraic geometry in R2.” In: J. Amer. Math. Soc.
18.2 (2005), pp. 313–377. arXiv: math/0312530 . [132] Grigory Mikhalkin. “Moduli spaces of rational tropical curves.” In: Proceedings of Gökova Geometry-Topology Conference 2006. Gökova Geometry/Topology Conference (GGT), Gökova,
2007, pp. 39–51. ærxiv: 0704.0839 . [133] Swarnava Mukhopadhyay and Richard Wentworth. “Generalized theta functions, strange duality, and odd orthogonal bundles on curves.” In: Comm. Math. Phys. 370.1 (2019), pp. 325–
376. arXiv: 1608.04990v1 . [134] Vicente Muñoz. “Higher type adjunction inequalities for Donaldson invariants.” In: Trans.
Amer. Math. Soc. 353.7 (2001), pp. 2635–2654. arXiv: math/9901046v2 . [135] Vicente Muñoz. “Quantum cohomology of the moduli space of stable bundles over a Riemann
surface.” In: Duke Math. J. 98.3 (1999), pp. 525–540. arXiv: alg-geom/9711030v1 . [136] Vicente Muñoz. “Ring structure of the Floer cohomology of Σ × S1.” In: Topology 38.3 (1999),
pp. 517–528. arXiv: dg-ga/9710029v1 . [137] M. S. Narasimhan. “Derived categories of moduli spaces of vector bundles on curves.” In: J. Geom. Phys. 122 (2017), pp. 53–58. [138] M. S. Narasimhan. “Derived categories of moduli spaces of vector bundles on curves II.” In: Geometry, algebra, number theory, and their information technology applications. Vol. 251. Springer Proc. Math. Stat. Springer, Cham, 2018, pp. 375–382. [139] M. S. Narasimhan and S. Ramanan. “Moduli of vector bundles on a compact Riemann surface.” In: Ann. of Math. (2) 89 (1969), pp. 14–51. [140] M. S. Narasimhan and C. S. Seshadri. “Stable and unitary vector bundles on a compact
Riemann surface.” In: Ann. of Math. (2) 82 (1965), pp. 540–567. url: https://doi.org/10.
2307/1970710 . [141] Peter Newstead. “Stable bundles of rank 2 and odd degree over a curve of genus 2.” In: Topology 7 (1968), pp. 205–215. 90 REFERENCES
[142] Peter Newstead. “Topological properties of some spaces of stable bundles.” In: Topology 6 (1967), pp. 241–262. [143] Takeo Nishinou. “Toric degenerations, tropical curve, and Gromov-Witten invariants of
Fano manifolds.” In: Canad. J. Math. 67.3 (2015), pp. 667–695. arXiv: 1001.2884v1 . [144] Takeo Nishinou, Yuichi Nohara, and Kazushi Ueda. “Potential functions via toric degenera-
tions.” In: Proc. Japan Acad. Ser. A Math. Sci. 88.2 (2012), pp. 31–33. arXiv: 0812.0066v2 . [145] Takeo Nishinou, Yuichi Nohara, and Kazushi Ueda. “Toric degenerations of Gelfand-Cetlin
systems and potential functions.” In: Adv. Math. 224.2 (2010), pp. 648–706. arXiv: 0810.
3470v2 . [146] Takeo Nishinou and Bernd Siebert. “Toric degenerations of toric varieties and tropical
curves.” In: Duke Math. J. 135.1 (2006), pp. 1–51. arXiv: math/0409060v3 . [147] Yuichi Nohara and Kazushi Ueda. “Goldman systems and bending systems.” In: Canad. J.
Math. 67.5 (2015), pp. 1109–1143. arXiv: 1405.1058 . [148] Yuichi Nohara and Kazushi Ueda. “Potential functions on Grassmannians of planes and
cluster transformations.” In: J. Symplectic Geom. 18.2 (2020), pp. 559–612. arXiv: 1711 .
04456v2 . [149] Yuichi Nohara and Kazushi Ueda. “Toric degenerations of integrable systems on Grassman-
nians and polygon spaces.” In: Nagoya Math. J. 214 (2014), pp. 125–168. arXiv: 1111.4809v2 . [150] Sebastian Novak and Ingo Runkel. “State sum construction of two-dimensional topologi- cal quantum eld theories on spin surfaces.” In: J. Knot Theory Ramications 24.5 (2015),
pp. 1550028, 84. arXiv: 1402.2839v2 . [151] Dmitri O. Orlov. “Triangulated categories of singularities and D-branes in Landau-Ginzburg models.” In: Tr. Mat. Inst. Steklova 246.Algebr. Geom. Metody, Svyazi i Prilozh. (2004). trans-
lation in: Proc. Steklov Inst. Math. 2004, no. 3(246), 227–248, pp. 240–262. arXiv: math /
0302304v2 .
[152] PARI/GP version 2.14.0. available from http : / / pari . math . u - bordeaux . fr/ . The
PARI Group. 2020. url: http://pari.math.u-bordeaux.fr . [153] James Pascale and Dmitry Tonkonog. The wall-crossing formula and Lagrangian mutations.
arXiv: 1711.03209v3 . [154] Christian Pauly. “Espaces de modules de brés paraboliques et blocs conformes.” In: Duke Math. J. 84.1 (1996), pp. 217–235. [155] Victor Przyjalkowski. “Toric Landau-Ginzburg models.” In: Uspekhi Mat. Nauk 73.6(444)
(2018), pp. 95–190. arXiv: 1809.09230v2 . [156] Victor Przyjalkowski. “Weak Landau-Ginzburg models of smooth Fano threefolds.” In: Izv. Ross. Akad. Nauk Ser. Mat. 77.4 (2013). translation in: Izv. Math. 77 (2013), no. 4, 772–794,
pp. 135–160. arXiv: 0902.4668v3 . [157] Victor Przyjalkowski and Constantin Shramov. “On weak Landau-Ginzburg models for complete intersections in Grassmannians.”In: Uspekhi Mat. Nauk 69.6(420) (2014). translation in: Russian Math. Surveys 69 (2014), no. 6, 1129–1131, pp. 181–182. [158] Johannes Rau. “Tropical intersection theory and gravitational descendants.” doctoralthesis.
Technische Universität Kaiserslautern, Aug. 2009. url: https://nbn-resolving.org/urn:
nbn:de:hbz:386-kluedo-23706 . [159] K. Rietsch and L. Williams. “Newton-Okounkov bodies, cluster duality, and mirror symmetry for Grassmannians.” In: Duke Math. J. 168.18 (2019), pp. 3437–3527. [160] Alexander F. Ritter. “Circle actions, quantum cohomology, and the Fukaya category of Fano
toric varieties.” In: Geom. Topol. 20.4 (2016), pp. 1941–2052. arXiv: 1406.7174v2 . [161] Georgios Dimitroglou Rizell, Tobias Ekholm, and Dmitry Tonkonog. Rened disk potentials
for immersed Lagrangian surfaces. arXiv: 1806.03722v2 . [162] Joseph Rusinko. “Equivalence of mirror families constructed from toric degenerations of
ag varieties.” In: Transform. Groups 13.1 (2008), pp. 173–194. arXiv: math/0610563 .
[163] The Sage Developers. SageMath, the Sage Mathematics Software System (Version 9.0). sagemath.
org . 2020.
[164] Alexander Samokhin. The Frobenius morphism on ag varieties, I. arXiv: 1410.3742v3 . [165] Fumihiko Sanda and Yota Shamoto. “An analogue of Dubrovin’s conjecture.” In: Ann. Inst.
Fourier (Grenoble) 70.2 (2020), pp. 621–682. arXiv: 1705.05989v2 . [166] Ivan Smith. “Floer cohomology and pencils of quadrics.” In: Invent. Math. 189.1 (2012),
pp. 149–250. arXiv: 1006.1099v3 . REFERENCES 91
[167] Andrew Strominger, Shing-Tung Yau, and Eric Zaslow. “Mirror symmetry is 푇 -duality.” In: Nuclear Phys. B 479.1-2 (1996), pp. 243–259. [168] Michael Thaddeus. “Stable pairs, linear systems and the Verlinde formula.” In: Invent. Math.
117.2 (1994), pp. 317–353. arXiv: alg-geom/9210007v1 . [169] Yukinobu Toda. Semiorthogonal decompositions of stable pair moduli spaces via d-critical
ips. arXiv: 1805.00183v2 . [170] Dmitry Tonkonog. String topology with gravitational descendants, and periods of Landau-
Ginzburg potentials. arXiv: 1801.06921v3 . [171] Akihiro Tsuchiya, Kenji Ueno, and Yasuhiko Yamada. “Conformal eld theory on universal family of stable curves with gauge symmetries.” In: Integrable systems in quantum eld theory and statistical mechanics. Vol. 19. Adv. Stud. Pure Math. Academic Press, Boston, MA, 1989, pp. 459–566. [172] Andrei Tyurin. Quantization, classical and quantum eld theory and theta functions. Vol. 21. CRM Monograph Series. With a foreword by Alexei Kokotov. American Mathematical
Society, Providence, RI, 2003, pp. xii+136. isbn: 0-8218-3240-9. arXiv: math/0210466v1 . [173] Erik Verlinde. “Fusion rules and modular transformations in 2D conformal eld theory.” In: Nuclear Phys. B 300.3 (1988), pp. 360–376. [174] Renato Ferreira de Velloso Vianna. “Innitely many exotic monotone Lagrangian tori in 2 CP .” In: J. Topol. 9.2 (2016), pp. 535–551. arXiv: 1409.2850v4 .
(1) Universität Bonn, Endenicher Allee 60, 53115 Bonn, Germany, (2) Universiteit Antwerpen, Middel- heimlaan 1, 2020 Antwerpen, Belgium, (3) Universiteit Hasselt, Agoralaan, 3590 Diepenbeek, Belgium
(1) PUC-Rio, Departamento de Matemática, Rua Marês de São Vicente 225, Gávea, Rio de Janeiro, (2) HSE University, Russian Federation
School of Mathematics, Tata Institute of Fundamental Research, 1 Homi Bhabha Road, Navy Nagar, Colaba, Mumbai 400005