1. Introduction

This paper is devoted to some aspects of counting of BPS states in a system of D-branes, with even , in type IIA string compactifications. The problems of BPS counting span a vast area of research in supersymmetric gauge and string theories. Their important feature is a special, nonconstant character of BPS multiplicities: their values depend on various moduli and jump discontinuously along some special loci in the corresponding moduli space, so called walls of marginal stability. The pattern of these jumps follows wall-crossing formulas, found from physical perspective by Denef and Moore [1] and, in more general context, formulated mathematically by Kontsevich and Soibelman [2]. The regions of the moduli space in between walls of marginal stability, in which BPS multiplicities are (locally) constant, are called chambers.

The BPS states we are interested in, and which we will refer to as closed BPS states, arise as bound states of a single D6-brane with arbitrary number of D0 and D2-branes wrapping cycles of a toric Calabi-Yau space. More generally, we will also consider open BPS states, which arise when an additional D4-brane spans a Lagrangian submanifold inside the Calabi-Yau space and supports open D2-branes attached to it. The closed and open BPS states give rise, respectively, to single-particle states in the effective four-dimensional and two-dimensional theory (in remaining, space-time filling directions of, resp., D6 and D4-branes). In this context, the character of BPS multiplicities can be understood in much detail, and it relates to other interesting exactly solvable models: free fermions, crystal, and matrix models. In brief, these connections arise as follows. Firstly, BPS states we consider turn out to be in one-to-one correspondence with configurations of certain statistical models of melting crystals. The structure of these crystals depends on geometry of the underlying Calabi-Yau space, as well as on the chamber one is considering. In consequence, BPS counting functions, upon appropriate identification of parameters, coincide with generating functions of melting crystals. It turns out that the structure of these crystals can be given a free fermion representation. Furthermore, once such free fermion formulation is known, it can also be represented in terms of matrix models. Connection with vast theory of matrix models has many interesting mathematical and physical consequences and allows to shed new light on wall-crossing phenomena. The aim of this paper is to explain these connections.

The BPS generating functions which we consider are intimately related to topological string amplitudes on corresponding Calabi-Yau spaces. This relation is most transparent in the physical derivation discussed in Section 2, which relies on lifting the D-brane system to M-theory. The M-theory viewpoint makes contact with original formulation of closed topological strings by Gopakumar and Vafa [3, 4], and open topological strings by Ooguri and Vafa [5]. In particular, in one specific, so-called noncommutative chamber, the BPS-generating function is given as the modulus square of the topological string partition function. In all other chambers, BPS generating functions can be uniquely determined from that noncommutative result. There is also another special, so-called commutative chamber, in which BPS generating function coincides (up to the factor of MacMahon function) with the topological string partition function. For toric manifolds which we consider, such topological string amplitudes can be constructed, among the other, by means of the powerful topological vertex formalism [6]. Relation to crystal models was in fact first understood in this topological string chamber [7–9]. One advantage of the formalism presented in this paper is the fact that it allows to construct matrix model representation of all these generating functions (so, in particular, matrix model representation of topological string amplitudes).

In more detail, we will consider generating functions of D2 and D0-branes bound to a single D6-brane of the following form: where is D0-brane charge and is D2-brane charge. Multiplicities jump when central charges (which itself are functions of Kähler moduli) of building blocks of a bound state align, and therefore these generating functions are locally constant functions of Kähler moduli. Along the walls of marginal stability, the degeneracies change and indeed obey wall-crossing formulas of [1, 2] mentioned above.

If there is an additional D4-brane which spans a Lagrangian submanifold inside the Calabi-Yau space, in addition to the above closed BPS states, one can consider also open BPS states of D2-branes with boundaries ending on a one-cycle on this D4-brane. In this case, the BPS states arise on the remaining two-dimensional world-volume of the D4-brane. The holonomy of the gauge field along provides another generating parameter , so that open BPS-generating functions take form As we will show, generating functions of such open BPS states can be identified with integrands of matrix models mentioned above.

One more important aspect of BPS counting is referred to as refinement, and amounts to refining BPS counting by introducing one more parameter, customarily denoted . The refinement can be introduced from several perspectives which give rise to identical results; however, their fundamental common origin is still not fully understood. We will introduce refinement by distinct counting of states with different spins inside spacetime rotation group in the generating function (1.1). In [10], it was argued that this physical viewpoint should agree with the mathematical counterpart of motivic deformation [2], and also a refined version of a crystal model was constructed. Another notion of refinement arises in Nekrasov partition functions, which are defined in a nontrivial gravitational (so-called -) background parametrized by two parameters and [11]. Nekrasov partition functions can also be defined for five-dimensional gauge theories and then they agree with topological string amplitudes. In particular, the formalism of the topological vertex [6] has also been extended to the refined context in [12], and shown to reproduce relevant Nekrasov partition functions. Also BPS generating functions, in the limit of commutative chamber, are known to reproduce refined topological string amplitudes with [13]. However, the worldsheet definition of refined topological string amplitudes is not fully understood.

As an exemplary and, hopefully, inspiring application of the entire formalism presented in this paper, in the final Section 6 we derive matrix model representation of the refined topological string partition function for the conifold. The refined matrix model which we find has a standard measure; however, its potential is deformed by -dependent terms. It is obtained by constructing appropriate refined crystal model and free fermion representation, and subsequently reformulating this representation in matrix model form. Finally, taking the limit of the commutative chamber, we obtain matrix model representation of the refined topological string amplitude. Even though we demonstrate this result in the conifold case, with some technical effort it can be generalized to other toric manifolds which we consider (As we recall in Section 6, refined topological string amplitudes were also postulated to be reproduced by another type of matrix models, so-called -deformed ones (whose Vandermonde measure is deformed by raising it to power ); however, explicit computations showed that this cannot be the correct representation of refined amplitudes.).

1.1. Short Literature Guide

The literature on the topics presented in this paper is extensive and still growing, and we unavoidably mention just a fraction of important developments. The relation between Donaldson-Thomas invariants for the noncommutative chamber of the conifold was first found by Szendrői [14]. It was generalized to orbifolds of , and related to free fermion formalism, by Bryan and Young [15]. The relation to free fermions and crystals was extended to a large class of toric manifolds without compact four-cycles [16, 17]. These developments were accompanied by other mathematical works [18, 19].

In parallel to the above-mentioned mathematical activity, wall-crossing phenomena for local Calabi-Yau manifolds were analyzed from physical viewpoint. The analysis of nontrivial BPS counting for the conifold was described by Jafferis and Moore in [20]. This and more general cases were related to quivers and crystal models in [21, 22]. Derivation of BPS degeneracies from M-theory viewpoint and relation to closed topological strings were discussed in [23], and generalized to open BPS counting in [24–27]. Relations to matrix models, discussed for plane partitions with some other motivation in [28], were extended to other crystal models relevant for BPS counting in [29], and also in [30]. Subsequently, it was related to open BPS counting in [27]. Refined BPS counting was related to crystal models in [10, 13], and corresponding matrix models were constructed in [31].

Let us also mention some other, related works devoted to crystals and free fermions. The fermionic construction of MacMahon function for was originally presented in [7], and its relation to open topological strings and more complicated Calabi-Yau manifolds were discussed in [32–34]. Newer ideas, analyzing more complicated systems involving D4-branes, were presented in [35, 36]. More expository presentations of various aspects described here can be found in [37, 38]. A general introduction to mathematical and physical aspects of mirror symmetry can be found in [39].

1.2. Plan

The plan of this paper is as follows. In Section 2, we introduce BPS generating functions and present one possible derivation of their form, which relies on the M-theory interpretation of a D-brane system, following [23–25, 27]. In Section 3, we provide a little mathematical background and introduce notation pertaining to toric Calabi-Yau manifolds, free fermion formalism, and matrix models. In Section 4, we introduce fermionic formalism for BPS generating functions and present corresponding crystal models, building on earlier ideas of [7, 15] and following [16]. In Section 5, we reformulate the problem of closed BPS counting in terms of matrix models and relate it to open BPS counting [27, 29]. In Section 6, we refine our analysis, present refined BPS generating functions and crystals [10], and construct corresponding refined matrix models [31].

2. BPS Generating Functions

In this section, we introduce generating functions of BPS states of D-branes in toric Calabi-Yau manifolds. Our task in the rest of this paper is to provide interpretation of these generating functions in terms of free fermions, melting crystals, and matrix models. These generating functions can be derived using wall-crossing formulas, as was done first in the unrefined [20] and refined [10] conifold case, and later generalized to arbitrary geometry without compact four-cycles in [18, 19]. On the other hand, we will focus on a simpler physical derivation of BPS generating functions which uses the lift of the D-brane system to M-theory [23]. This also makes contact with M-theory interpretation of topological string theory and allows to express BPS counting functions in terms of topological string amplitudes. Moreover, this M-theory derivation can be extended to the counting of open BPS states,that is, open D2-branes attached to additional D4-brane, which we are also interested in [24, 25, 27].

We start this section by reviewing the M-theory derivation of (unrefined) closed and open BPS generating functions. Then, to get acquainted with a crystal interpretation of these generating functions, we discuss their crystal interpretation in simple cases of and conifold. Later, using fermionic interpretation, we will generalize this crystal representation to a large class of toric manifolds without compact four cycles.

2.1. M-Theory Derivation

We start by considering a system of D2 and D0-branes bound to a single D6-brane in type IIA string theory. It can be reinterpreted in M-theory as follows [23]. When additional is introduced as the eleventh dimension transversely to the D6-brane, then this D6-brane transforms into a geometric background of a Taub-NUT space with unit charge [40]. The Taub-NUT space is a circle fibration over , with a circle attaining a fixed radius at infinity, and shrinking to a point in the location of the original D6-brane. From M-theory perspective, bound states involving D2 and D0-branes are interpreted as M2-branes with momentum on a circle. Therefore, the counting of original bound states to the D6-brane is reinterpreted as the counting of BPS states of M2-branes in the Taub-NUT space. While in general this is still a nontrivial problem, for the purpose of counting BPS degeneracies we can take advantage of their invariance under continuous deformations of the Taub-NUT space, in particular under deformations of the radius . We can therefore consider taking this radius to infinity, whereupon BPS counting is reinterpreted in terms of a gas of particles in . To make the problem fully tractable, we have to ensure that the particles are noninteracting, which would be the case if moduli of the Calabi-Yau would be tuned so that M2-branes wrapped in various ways would have aligned central charges. This can be achieved when Kähler parameters of the Calabi-Yau space are tuned to zero. However, to avoid generation of massless states, at the same time one has to include nontrivial fluxes of the M-theory three-form field through the two cycles of the Calabi-Yau and . In type IIA, this results in the -field flux through two cycles of Calabi-Yau. Finally, to avoid creation of the string states arising from M5-branes wrapping four cycles in Calabi-Yau, we simply restrict considerations to manifolds without compact four cycles. For a state arising from D2-brane wrapping a class , the central charge then reads where counts the D0-brane charge, which is taken positive to preserve the same supersymmetry.

Under the above conditions, the counting of D6-D2-D0 bound states is reinterpreted in terms of a gas of particles arising from M2-branes wrapped on cycles . The excitations of these particles in , parametrized by two complex variables , are accounted for by the modes of the holomorphic field Decomposing the isometry group of as , there are five-dimensional BPS states of intrinsic spin . We are interested in their net number arising from tracing over spins The total angular momentum of a given state contributing to the index is . Finally, in a chamber specified by the moduli and , the invariant degeneracies can be expressed as the trace over the corresponding Fock space where the subscript chamber denotes restriction to those factors in the above product, which represent states which are mutually BPS As usual, and above encode, respectively, the Kähler class and the string coupling (we wish to distinguish carefully which encodes string coupling, from a counting parameter which will arise in what follows in crystal interpretation). The above condition on central charges is crucial in determining a particular form of the BPS generating functions. If we would restrict products in the formula (2.4) to factors with only positive , we would get (up to possibly some factor of MacMahon function) the Gopakumar-Vafa representation of the topological string amplitude. With all negative and positive values of , we would get modulus square of the topological string partition function. Therefore, the upshot of [23] is that in general the above BPS generating function can be expressed in terms of the closed topological string partition function where chamber restriction is to be understood as picking up only those factors in Gopakumar-Vafa product representation of for which (2.5) is satisfied. In this context, we will often refer to the choice of a chamber as a closed BPS chamber. The (instanton part of the) closed topological string partition function entering the above expression is given by [3, 4] where is the MacMahon function and is the Euler characteristic of the Calabi-Yau manifold.

To be more precise, an identification as a topological string partition function or its square arises if in (2.1). Because arises just as a multiplicative factor in (2.1), degeneracies depend only on its sign. Therefore, another extreme case corresponds to negative and sufficiently small, when only a single D6-brane contributes to the partition function More generally, for , BPS generating functions often (but not always) take finite form.

In what follows we denote BPS generating functions in chambers with positive by , and in chambers with negative by (and often omit the subscript BPS). Topological string partition functions will be denoted by , while generating functions of melting crystals by ordinary .

The above structure can be generalized by including in the initial D6-D2-D0 configuration additional D4-branes wrapping Lagrangian cycles in the internal Calabi-Yau manifold and extending in two space-time dimensions [24, 25, 27]. For simplicity, we consider a system with a single D4-brane wrapping a Lagrangian cycle. There are now additional BPS states in two remaining spacetime dimensions arising from open D2-branes ending on these D4-branes. Their net degeneracies are characterized, firstly, by the spin whose origin is most clearly seen from the M-theory perspective [5, 41]. Secondly, they depend on two-cycles wrapped by open M2-branes, as well as one-cycles on which these M2-branes can end (In case of D4-branes wrapping the same Lagrangian cycle, these states would additionally arise in representations of [5]. In case of a single brane, this reduces to , and such a dependence can be reabsorbed into a parameter specifying a choice of .).

Lifting this system to M-theory, we obtain a background of the form , with the additional D4-brane promoted to M5-brane. This M5-brane wraps the Lagrangian submanifold inside Calabi-Yau, the time circle , and inside the Taub-NUT space. A part of this Lagrangian is a torus , which will lead to some modular properties of the BPS counting functions: this modularity will be manifest in one chamber, where the open topological string amplitude will be completed to the product of functions. This M5-brane also breaks the spatial symmetry down to . We denote the spins associated to both factors, respectively, by and , and the degeneracies of particles with such spins by . In addition to closed Kähler parameters , let us also introduce open ones related to discs wrapped by M2-branes . The real and imaginary parts of encode, respectively, the sizes of two-cycles and the value of the -field through them. The real and imaginary parts of encode, respectively, sizes of the discs and holonomies of the gauge fields around them. Similarly as in the closed string case, to get nontrivial ensemble of mutually supersymmetric states, we set the real parts of and to zero, and consider nontrivial imaginary parts.

From the M-theory perspective, we are interested in counting the net degeneracies of M2-branes ending on this M5-brane In the remaining three-dimensional space, in the limit, the M2-branes ending on the M5-brane are represented by a gas of free particles. These particles have excitations in which we identify with the -plane. To each such BPS particle, similarly as in the closed string case discussed above and in [23, 40], we can associate a holomorphic field The modes of this field create states with the intrinsic spin and the orbital momentum in the plane. The derivation of the BPS degeneracies relies on the identification of this total momentum in the limit, with the Kaluza-Klein modes associated to the rotations along for the finite , following the five-dimensional discussion in [40, 42].

The BPS generating functions we are after are given by a trace over the Fock space built by the oscillators of the second quantized field and restricted to the states which are mutually supersymmetric. In such a trace, each oscillator from (2.10) gives rise to one factor of the form , where the exponent corresponds to the bosonic or fermionic character of the top component of the BPS state, where the product is over either both positive or both negative . The parameters and specify the chamber structure: the restriction to a given chamber is implemented by imposing the condition on a central charge, analogous to (2.5), This condition in fact specifies a choice of both closed and open chambers. The walls of marginal stability between chambers correspond to subspaces where, for some oscillator, the above product becomes 1, and then the contribution from such an oscillator drops out from the BPS generating function.

Similarly as in the closed string case, the above degeneracies can be related to open topological string amplitudes, rewritten in [5] in the form with integer Ooguri-Vafa invariants (In case of D4-branes wrapping a Lagrangian cycle, this structure is again more complicated, because the states in arise in representations of [5]. This requires replacing the factor by the sum of traces in all possible representations of this of the matrix encoding holonomies of the gauge fields. For simplicity we restrict here to the simplest case.). This formula represents in fact a series of quantum dilogarithms and can be written in the product form Comparing with (2.11) we conclude that the BPS counting functions take form of the modulus square of the open topological string amplitude

Similarly as in the closed string case, there are also a few particularly interesting chambers to consider. For example, in the extreme chamber corresponding to Im , Im , the trace is performed over the full Fock space and yields the modulus square of the open topological string partition function. In this case, the quantum dilogarithms arise in pairs, which (using the Jacobi triple product identity) combine to the modular function ; in consequence, the total BPS generating function is modular and expressed as a product of such functions.

2.2. Crystal Interpretation

Closed BPS generating functions (2.4) turn out to be generating functions of statistical models of crystals, when parameters relevant for both interpretations are appropriately matched. Physical reasons for such relations have been given in [8, 21, 22], and mathematical interpretation arose from works [9, 14, 15]. Such crystal interpretation arises also from the fermionic formulation [16, 17], as we will review below. These crystals, in a more intricate way [27], encode also open BPS generating functions (2.11). However, before discussing details of all these constructions, in this introductory section we present crystal models for two simplest toric Calabi-Yau manifolds, that is, and conifold.

is the simplest Calabi-Yau manifold. It has no compact two-cycles, so relevant BPS states are bound states of arbitrary number of D0-branes with a single D6-brane wrapping entire . Their generating function is therefore expressed in terms of a single parameter . There is just a single nonzero Gopakumar-Vafa invariant , and as follows from (2.4) this generating function coincides with the so-called MacMahon function

On the other hand, the MacMahon function is a generating function of plane partitions, that is, three-dimensional generalization of Young diagrams. These plane partitions represent the simplest three-dimensional crystal model, namely, they can be identified with stacks of unit cubes filling the positive octant of space, as shown in Figure 1. A unit cube located in position can evaporate from this crystal only if all other cubes with coordinates are already missing. A plane partition is weighted by the number of boxes it consists of , with a weight associated to a single box, so indeed where is the number of plane partitions which consist of cubes. Therefore, plane partition generating function coincides with the BPS counting function when a simple identification is made. From (2.6), it follows that the topological string partition function for is given by the square root of the MacMahon function which is indeed true. The relevance of the MacMahon function for geometry was noticed for the first time in [3], and a statistical model interpretation of this result was proposed in [7].

Figure 1: Plane partitions represent melting crystal configurations of .

The conifold provides another simple, yet nontrivial example of toric Calabi-Yau manifold. It consists of two patches glued int, and it has one Kähler class representing , parametrized by . This class can be wrapped by D2-branes, which bind with D0-branes to an underlying D6-brane and give rise to BPS states in low energy theory. In this case, there is already a nontrivial structure of chambers and walls, which was analyzed in [14, 18, 20, 21]. This structure is consistent with M-theory derivation discussed in Section 2.1. The generating functions of D6-D2-D0 bound states are parametrized by and , and therefore corresponding crystal models consist of two-colored three-dimensional partitions. The Kähler moduli space consists of several infinite countable sets of chambers, and in each chamber relevant crystal configurations take form of so-called pyramid partitions. These partitions are infinite or finite (resp. for positive and negative in (2.1)) and their size depends on the value of the -field. This size changes discretely and the pyramid is enlarged when the value of the -field crosses integer numbers, which changes the chamber in the moduli space, as explained in Section 2.1. Examples of such infinite pyramid partitions are given in Figure 2, and finite ones in Figure 3.

Figure 2: Infinite pyramids with one and four balls in the top row, with generating functions given, respectively, by and .

Figure 3: Finite pyramids with stones in the top row (resp., (a), (b), and (c)), whose generating functions are given by (note that corresponds to an empty pyramid corresponding to the pure D6-brane).

To write down explicitly BPS generating functions for the conifold in various chambers, we can take advantage of their relation to the topological string amplitude (2.6). The topological string partition function in this case reads with the MacMahon function defined in (2.17). From this topological string partition function we can read off Gopakumar-Vafa invariants [3, 4]

Using the relation (2.6), we can now present conifold closed BPS generating functions in several sets of chambers. In the first set of chambers, we consider and positive (for ). Firstly, for small , there is so-called noncommutative chamber discussed first by Szendrői [14], which corresponds to . In this case, the pyramid crystal has just a single ball in the top row, as in Figure 2(a), and the BPS generating function is given by the square of the topological amplitude. On the other hand, for large , that is , we reach commutative chamber in which the length of the top row extends to infinity. In this case, the BPS generating function agrees, up to a single factor of MacMahon function, with the topological string amplitude. In between, there are chambers with balls in the top row, for which These BPS generating functions are related to pyramid generating functions with two colors and upon the identification (which generalizes (2.19) in case) Indeed, with this identification, the above counting functions agree with those of two-colored pyramid crystals with yellow balls in its top row

In the second set of chambers, we have and positive (for ). It extends between the core region with a single D6-brane (2.8) and the chamber characterized by so-called Pandharipande-Thomas invariants (for the flopped geometry, or equivalently for anti-M2-branes). The BPS generating functions read The corresponding statistical models were shown in [16, 18, 21] to correspond to finite pyramids with stones in the top row, as shown in Figure 3. In this case, the generating functions of such partitions are equal to The equality arises upon an identification There are two other sets of chambers characterized by the negative value of the -field, for which BPS generating functions are completely analogous to those given above.

Above, we presented just the simplest examples of crystal models. Using fermionic formulation presented below, one can find other crystal models for arbitrary toric geometry without compact four cycles. Let us also mention that those models can be equivalently expressed in terms dimers. In particular, the operation of enlarging the crystal, as in the conifold pyramids, corresponds to so-called dimer shuffling [15]. Dimers are also closely related to a formulation using quivers and associated potentials, which underlies physical derivations in [21, 22].

3. A Little Background—Free Fermions and Matrix Models

In this section, we introduce some mathematical background on which the main results presented in this paper rely. In Section 3.1, we start with a brief presentation of toric Calabi-Yau manifolds and introduce the notation which we use in what follows. In Section 3.2, we introduce free fermion formalism. In Section 3.3, we introduce basics of matrix model formalism. Our presentation is necessarily brief, and for more detailed introduction we recommend many excellent reviews on each of those topics.

3.1. Toric Calabi-Yau Threefolds

Some introductory material on toric Calabi-Yau manifolds, from the perspective relevant for mirror symmetry and topological string thoery, can be found, for example, in [39]. In this section, our presentation is brief and mainly sets up the notation. Toric Calabi-Yau threefolds arise as the quotient of , possibly with a discrete set of points deleted, by the action of with certain weights. The simplest toric threefold is , which corresponds to the trivial choice . The resolved conifold, which we already discussed in Section 2.2, corresponds to and a choice of weights , which represent a local bundle . The structure of each toric three-fold can be encoded in a two-dimensional diagram built from trivalent vertices. Finite intervals joining two adjacent vertices represent local neighborhood inside the manifold. Equivalently, one can consider dual graphs. Examples of toric diagrams and their duals for , conifold and resolution of singularity are given in Figure 4 (the notation at each vertex will be explained in what follows).

Figure 4: Toric graphs for , conifold and resolution of .

A closed loop in a toric diagram represents a compact four cycles in the geometry. As follows from the reasoning in Section 2.1, in the context of BPS counting, we are forced to restrict considerations to manifolds which do not have such four cycles. Apart from a few special cases, there is an infinite class of such geometries whose dual diagrams arise from a triangulation, into triangles of area , of a long rectangle or a strip of height 1. A toric diagram arises as a dual graph to such a triangulation. From each vertex in such a toric diagram, one vertical line extends to infinity and crosses either the upper or the lower edge of the strip. Two such consecutive lines can emanate either in the same or in the opposite direction, respectively, when they are the endpoints of an interval representing with local or neighborhood. An example of a generic diagram of this kind is shown in Figure 8.

Let us denote independent 's, starting from the left end of the strip, from 1 to , and introduce corresponding Kähler parameters , . Moreover, to each toric vertex we associate a type , so that if the local neighborhood of (represented by an interval between vertices and ) is ; if this neighborhood is of type, then . The type of the first vertex we fix as . In Figures 4 and 8, these types are denoted by and . The types will be used much in the construction of fermionic states in Section 4.2.

As explained in Section 2.1, the BPS generating functions can be expressed in terms (the instanton part) of topological string amplitudes. For the above class of geometries, arising from a triangulation of a strip, these amplitudes read

3.2. Free Fermion Formalism

Formalism of free fermions in two dimensions is well known [43, 44] and ubiquitous in literature on topological strings and crystal melting [7, 15, 15, 37, 45]. The main purpose of this section is therefore to set up the notation which we will follow in the remaining parts of this paper.

The states in the free fermion Fock space are created by the (anticommuting) modes of the fermion field on the vacuum state . There is one-to-one map between such fermionic states and two-dimensional partitions , as shown in Figure 5. The modes of the bosonized field satisfy the Heisenberg algebra .

Figure 5: Relation between Young diagrams and states in the Fermi sea.

We introduce vertex operators which act on fermionic states corresponding to partitions as [15, 43, 44] The interlacing relation between partitions is defined as

The operator is the inverse of with negative argument. These operators satisfy commutation relations

We also introduce various colors and the corresponding operators (a hat is to distinguish them from Kähler parameters ) These operators commute with vertex operators up to rescaling of their arguments

3.3. Matrix Models

In matrix model theory, or theory of random matrices, one is interested in properties of various ensembles of matrices. Excellent reviews of random matrix theory can be found for example in [46] or, in particular in the context of topological string theory, in [47]. In matrix model theory, one typically considers partition functions of the form where is a matrix potential and is a measure over a set of matrices of interest of size . Typically it is not possible to perform the above integral; however, special techniques allow to determine its formal expansion. These techniques culminated with the formalism of the topological expansion of Eynard and Orantin [48] which, in principle, allows to determine entire expansion of the partition function recursively. This solution is determined by the behavior of matrix eigenvalues, whose distribution among the minima of the potential, in the continuum limit, determines one-dimensional complex curve, so-called spectral curve. The spectral curve is also encoded in the leading expansion of the so-called resolvent, which is defined as the expectation value computed with respect to the measure (3.15).

In the context of BPS counting and topological strings, unitary ensembles of matrices of infinite size arise. In this case, the matrix model simplifies to the integral over eigenvalues , with a measure which takes form of the unitary Vandermonde determinant The issue of infinite matrices is a little subtle; however, it can be taken care of by considering matrices of large but finite size , and subsequently taking limit. For finite , one can find the resolvent, and in consequence the spectral curve, using a standard technique of so-called Migdal integral. This requires redefining to the standard Vandermonde form [29, 47], as well as introducing 't Hooft coupling The form of the Migdal integral depends on the number of cuts into which eigenvalues condense in large limit, and this number of cuts determines the genus of the spectral curve. In our context, only single-cut situations will arise, for which the spectral curve has genus zero. In this case, the Migdal integral determines the resolvent as so that the integration contour encircles counter-clockwise the endpoints of the cut and . A proper asymptotic behavior of the resolvent is imposed by the condition Then the spectral curve is determined as a surface on which the resolvent is unambiguously defined, that is, it is given by an (exponential) rational equation automatically satisfied by and . There is also an important consistency condition for the resolvent: when computed on the opposite sides of the cut , it is related to the potential as On the other hand, a difference of these values of the resolvent on both sides of the cut provides eigenvalue density

It has been observed in several contexts that topological strings on toric manifolds can be related to matrix models, whose spectral curves take form of the so-called mirror curves. Mirror curves arise for manifolds which are mirror to toric Calabi-Yau manifolds [39, 45]. For toric manifolds, their mirror manifolds are determined by the following equation embedded in four-dimensional complex space: The mirror curve is the zero locus of , that is, it is given as . More precisely, are variables, and it is often convenient to represent them in the exponential form , , with . For example, for and the conifold they take the following form: where encodes the Kähler parameter of the conifold. Schematically mirror curves arise from thickening edges of the toric graphs, as shown in Figure 6.

Figure 6: Toric diagrams for and conifold and corresponding mirror curves.

One of the first relations between topological strings for toric manifolds and matrix models was encountered in [49, 50], where it was shown that the spectral curve of a unitary matrix model with a Gaussian (i.e., quadratic) potential agrees with the above mirror curve in (3.23), with ’t Hooft coupling encoded in . At the same time, it was shown that the matrix model partition function reproduces the topological string partition function. More recently these ideas became important in view of the remodeling conjecture [51], which states that the solution to loop equations in the form found by Eynard and Orantin [48], applied to the mirror curve, reproduces topological string partition functions. The method of [48] works for arbitrary curves, not necessarily originating from matrix models. Nonetheless, it is indeed possible to construct matrix models whose partition functions do reproduce topological string amplitudes, and whose spectral curves coincide with appropriate mirror curves [29, 30, 52–56].

One of our aims is to provide matrix model interpretation of BPS counting. It is natural to expect such an interpretation in view of an intimate relation between BPS counting and topological string theory discussed in Section 2.1, and the above-mentioned relations between topological strings and matrix models. As we will see in what follows, there are indeed unitary matrix models which naturally arise in the context of BPS counting and its fermionic formulation. Among the others, our task will be to analyze them using the above-mentioned Migdal method.

4. Fermionic Formulation of BPS Counting Functions

Having introduced all the ingredients above, we are now ready to present fermionic formulation of BPS counting. To start with, in Section 4.1 we present the idea of such a formulation in the simplest example of . In Section 4.2, we introduce a general fermionic formalism, and in Section 4.3 we provide its crystal interpretation. We illustrate the use of our formalism in Section 4.4 revisiting example, as well as in explicit case of , and conifold geometry.

4.1. The Idea and Example

As explained in Section 2.2, the generating function of bound states of D0-branes to a single D6-brane is given by the MacMahon function, and the corresponding crystal model takes form of the counting of plane partitions [7]. Let us slice each such plane partition by a set of parallel planes, as shown in Figure 7. In this way on each slice, we obtain a two-dimensional partition , and it is not hard to see that each two neighboring partitions satisfy the interlacing condition (3.7). Recalling that such a condition arises if we apply operators (3.5) to partition states, we conclude that a set of all plane partitions can be built, slice by slice, by acting with infinite sequence of on the vacuum. To count each slice with appropriate weight we also need to apply weight operator defined in (3.12). Therefore, the generating function of plane partitions can be represented as follows In the first line, we implicitly introduced two states and , defined by an infinite sequence of (resp. ) operators, interlaced with weight operators and acting on the vacuum. To confirm that this correlator indeed reproduces the MacMahon function, the second line can be reduced to the final infinite product using commutation relations (3.8) and (3.14). We can also represent insertions of operators graphically by arrows, so that the above computation can be represented as in Figure 7(b).

Figure 7: Slicing of a plane partition (a) into a sequence of interlacing two-dimensional partitions (b). A sequence of operators in (4.1) which create two-dimensional partitions is represented by arrows inserted along two axes. Directions of arrows represent interlacing condition on partitions. We reconsider this example from a new viewpoint in Figure 10.

Figure 8: Toric Calabi-Yau manifolds represented by a triangulation of a strip. There are independent 's with Kähler parameters , and vertices to which we associate and operators represented respectively by and signs. Yellow intervals, which connect vertices with opposite signs, represent local neighborhoods. Red intervals, which connect vertices with the same signs, represent local neighborhoods. The first vertex on the left is chosen to be .

In what follows, we present a formalism which allows to generalize this computation to a large class of chambers, for arbitrary toric geometry without compact four cycles.

4.2. Toric Geometry and Quantization

We wish to reformulate BPS counting in the fermionic language in a way in which we associate to each toric manifold a fermionic state, such that the BPS generating function can be expressed as an overlap of two such states, generalizing case (4.1). At the same time, the construction of such a fermionic state is supposed to encode the structure of the underlying crystal model (generalizing plane partitions in Figure 7). An important difference between and other geometries is the existence of many Kähler moduli and correspondingly many chambers, for which BPS generating functions change according to wall-crossing formulas. To take care of these changes in the fermionic formalism, we need to introduce special wall-crossing operators.

4.2.1. Toric Geometry and Fermionic Operators

In what follows we use the notation introduced in Section 3.1; in particular to each vertex of the toric diagram we associate its type , see also Figure 8. We start with a construction of fermionic states associated to a given toric Calabi-Yau manifold (without compact four-cycles). First we need to introduce several operators which are building blocks of such states. The structure of these operators is encoded in the toric diagram of a given manifold. Namely, these operators are given by a string of vertex operators (defined in (3.4)) which are associated to the vertices of the toric diagram; the type determines the type of a vertex operator as In addition the string of operators is interlaced with operators representing colors , for . Operators are associated to in the toric diagram, and there is an additional . We also define Therefore, the upper indices of and a choice of colors of the operators which we introduce below are specified by the data of a given toric manifold. As we will see, a sequence of lower indices ± is determined by the chamber we are going to consider.

Now we can associate several operators to a given toric manifold. Firstly, we define Commuting all 's using (3.14), we also define the following operators:

In addition, we define the above-mentioned wall-crossing operators Here the order of and is the same as for operators, and the difference is that now there are subscripts on first operators and ± on the remaining ones.

We often use a simplified notation when the argument of the above operators is

4.2.2. Fermionic Formulation and Quantization

Above we associated operators to each toric geometry with a strip-like toric diagram. From these operators, we can build the following states in the Hilbert space of a free fermion : which we define as follows: These states encode the full instanton part of the topological string amplitudes. Namely, as shown in [16], is equal to the BPS partition function in the noncommutative chamber where is given in (3.1). The above equality holds under the following identification between parameters (which enter the definition of ) and physical parameters and : We will provide a proof of (4.10) in Section 6.1.1 in a more general setting of refined invariants.

The states have nontrivial structure and encode the information about the noncommutative chamber. It turns out that the fermionic vacuum itself also encodes some interesting information. We recall that there is another extreme chamber representing just a single BPS state represented by the D6-brane with no other branes bound to it. This multiplicity 1 can be understood as and as we will see below, starting from this expression we can use wall-crossing operators to construct BPS generating functions in an infinite family of other chambers.

4.2.3. Other Chambers and Wall-Crossing Operators

In the previous, section we associated to toric manifolds the states , whose overlap reproduces the BPS generating function in the noncommutative chamber (4.10). Now we wish to extend this formalism to other chambers. As discussed in Section 2.1, in a given chamber, the allowed bound states we wish to count must have positive central charge (2.1) Firstly, the information about and must be encoded in the fermionic states which we wish to construct. It turns out that the choice of positive or negative is encoded in the choice of the ground state which generalizes the extreme cases (4.10) and (4.13).

On the other hand, the value of the field is encoded in the insertion of additional wall-crossing operators, such as those defined in (4.6). In particular, these two types of operators are sufficient if we wish to consider only these chambers, which correspond to a flux of the -field through only one, but arbitrary in the manifold. For simplicity below, we consider only this set of chambers. Denoting this as , it can be shown that insertion of copies of operators or creates, respectively, positive or negative quanta of the flux through 'th .

Therefore, schematically, the generating functions in chambers with read and those with read with appropriate form of wall-crossing operators. More precisely, depending on the signs of and , we need to consider four possible situations, which we present below. The proofs of all statements below, corresponding to these four situations, can be found in [16].

(i) Chambers with ,
Consider a chamber characterized by positive and positive -field through 'th two-cycle The BPS partition function in this chamber contains only those factors which include and it reads This can be expressed as the expectation value of wall-crossing operators under the following identification of variables: A special case of this result is the trivial generating function (4.13) representing a single D6-brane.