Non-abelian quantum statistics on graphs

We show, that the non-abelian quantum statistics can be studied using certain topological invariants, which are the homology groups of configuration spaces. In particular, we formulate a general framework for describing quantum statistics of particles constrained to move in a topological space $X$. The framework involves the study of flat complex vector bundles over the configuration space of $X$, which can be achieved by determining its homology groups. We apply this methodology for configuration spaces of graphs. As a conclusion, we provide families of graphs, which are good candidates for simplified models in the further study of quantum statistical phenomena and obtain a number of new results concerning topology of graph configuration spaces.

unitary representations of the fundamental group π 1 (Cn(X)). When X = R 2 this group is known to be the braid group and when X = R k , where k ≥ 3, it is the permutation group Sn. QS corresponding to a one-dimensional unitary representation of π 1 (Cn(X)) is called abelian whereas QS corresponding to a higher dimensional nonabelian unitary representation is called non-abelian QS. In fact quantum statistics can be also viewed as a flat connection on the configuration space Cn(X) that modifies definition of the momentum operator according to minimal coupling principle. The flatness of the connection ensures that there are no classical forces associated with it and the resulting physical phenomena is purely quantum [4,6] (c.f. Aharonov-Bohm effect [5]) The first part of this paper (sections [1][2][3] contains general considerations about the connections between topology of configuration spaces and the existence of different types of quantum statistics. This is an attempt to collect and organise some of the results that are partially a folklore knowledge in mathematical physics. The underlying idea is to compare the approach that dates back to the works of Souriau [1] and Leinaas and Myrheim [2] with the modern mathematical methods concerning the classification of complex vector bundles. We especially emphasise the important role of nontrivial flat vector bundles, that can lead to spontaneously occurring non-abelian quantum statistics. This is motivated by the fact that in R 3 fermions and bosons correspond to two non-isomorphic vector bundles that admit flat connections. Our approach to classification of quantum statistics is connected to classification of possible quantum kinematics i.e. defining the space of wave functions and deriving momentum operators that satisfy the canonical commutation rules. Then our classification scheme for quantum kinematics of rank k on a topological space X is divided into two steps 1. Topological classification of wave functions. Classify isomorphism classes of flat hermitian vector bundles of rank k over Cn(X). Here we also point out that in fact physically meaningful is the classification of vector bundles with respect to the so-called stable equivalence, as nonisomorphic but stable equivalent vector bundles have identical Chern numbers. An important role is played by the reduced K-theory and (co)homology groups of Cn(X). Calculation of those groups for various graph configuration spaces is the main problem we solve in section 5. 2. Classification of statistical properties. If X is a manifold, for each flat hermitian vector bundle, classify the flat connections. The parallel transport around loops in Cn(X) determines the statistical properties. For general paracompact X, this point can be phrased as classification of the U (k) -representations of the fundamental group.
One of our main goals is to formulate our work in a general context, so that the considered topological spaces do not have to be differential manifolds. All the results formulated in this paper are suitable for topological spaces that have the homotopy type of a finite CW -complex. Our work results with a number of open problems indicating the possibility of the existence of some new quantum statistical phenomena. The general methods that we describe in the first three sections of this paper, are applied to a special class of configuration spaces of particles on graphs (treated as 1-dimensional CW complexes), which may serve as simple models for studying quantum statistical phenomena [22,8,9,10,21] and already proved useful in other branches of physics such as quantum chaos and scattering theory [23,24,25]. In particular, we compute the homology groups of graph configuration spaces beyond the first homology [22,26]. The methods we use for these calculations built on our recent work [22,18] that combines tools from algebraic topology and graph theory with a set of combinatorial relations derived from the analysis of certain small canonical graphs. The main advantage of this approach is that it allows to easily predict the results of many complicated key calculations. It was successfully applied to calculate H 1 (Cn(Γ )) and all homology groups of graph configuration spaces of tree graphs [18] (see also [33,34,35,36,37,38,39,40] for recent results in this area). In this paper we combine our approach with the model proposed by Świątkowski [20]. This allows us to calculate homology groups of certain families of graphs by writing down their generators explicitly. These families are i) wheel graphs (subsection 5.3), ii) graph K 3,3 (subsection 5.5), iii) graphs K 2,p (subsection 5.6). Moreover, we provide a large family of simple graphs, for which the second homology group of Cn(Γ ) has a simplified structure, i.e. is generated by the so-called product cycles.

Quantum kinematics on smooth manifolds
A quantisation procedure for configuration spaces, where X is a smooth manifold, known under the name of Borel quantisation, has been formulated by H.D. Doebner et. al. and formalised in a series of papers [12,13,14,16,17]. Borel quantisation on smooth manifolds can be also viewed as a version of the geometric quantisation [15]. The main point of Borel quantisation is the fact that the possible quantum kinematics on Cn(X) are in a one-to-one correspondence with conjugacy classes of unitary representations of the fundamental group of the configuration space. We denote this fact by QKin k (Cn(X)) ∼ = Hom(π1(Cn(X)), U (k)) /U(k), where QKin k are the quantum kinematics of rank k. i.e. kinematics, where wave functions have values in C k and π 1 is the fundamental group. Let us next briefly describe the main ideas standing behind the Borel quantisation, which will be the starting point for building an analogous theory for indistinguishable particles on graphs.
In Borel quantisation on smooth manifolds, wave functions are viewed as squareintegrable sections of hermitian vector bundles. For a fixed hermitian vector bundle, the momentum operators are constructed by assigning a self-adjoint operatorp A acting on sections of E to a vector field A that is tangent to Cn(X) in the way that respects the Lie algebra structure of tangent vector fields. Namely, we require the standard commutation rule for momenta, i.e.
[p A ,q f ] =q Af .
(2) It turns out that such a requirement implies the form of the momentum operator, which is well-known form the minimal coupling principle, namelŷ where ∇ A is a covariant derivative in the direction of A, that is compatible with the hermitian structure. Moreover, commutation rule (1) implies, that ∇ A is necessarily the covariant derivative stemming from a flat connection. The component proportional to div(A) in formula (3) comes from the fact, that functor A →p A must be valid for an arbitrary complete vector field. Usually, one considers momentum operators coming from some specific vector fields, that form an orthonormal basis of local sections of T Cn(X). The divergence of such a basis sections usually vanishes and formula (3) describes the standard minimal coupling principle, see example 1 below. Flat hermitian connections of rank k are classified by conjugacy classes of U (k) representations of π 1 (Cn(X)) (see [60]). Representatives of these classes can be picked by specifying the holonomy on a fixed set of loops generating the fundamental group. In order to illustrate these concepts, consider the following example of one particle restricted to move on the plane and its scalar wave functions.
Example 1 Quantum kinematics of rank 1 for a single particle on the plane. The momentum has two components that are given by (3) for A = ∂x =: ∂ 1 and A = ∂y =: ∂ 2 .
By a straightforward calculation, one can check that commutation rule (2) is satisfied.
However, commutation rule (1) requires [p 1 ,p 2 ] = 0. The commutator reads Therefore, in order to satisfy the momentum commutation rule, we need ∂ 1 α 2 −∂ 2 α 1 = 0. This is precisely the condition for the connection form Γ := α 1 dx + α 2 dy to have zero curvature, i.e. dΓ = 0. The plane is a contractible space, hence the problem of classifying flat connections is trivial and there are no topological effects in the quantum kinematics. However, we can make the problem nontrivial by considering the situation, where a particle is moving on a plane without a point, i.e. X = R 2 − { * }. Then, π 1 (X) = Z generated by a circle around { * } travelled clockwise. Let us denote such a loop by γ. The parallel transport of Ψ around γ giveŝ Tγ Ψ = e ι γ Γ Ψ.
The phase factor e ι γ Γ does not depend on the choice of the circle. In order to see this, choose a different circle γ that contains γ. Denote by D the area between the circles. We have ∂D = γ − γ. Hence, by the Stokes theorem Hence, all U (1) representations of π 1 (X) are the representations that assign a phase factor e iφ to a chosen non-contractible loop. Physically, these representations can be realised as the Aharonov-Bohm effect and phase φ is the magnetic flux through point * that is perpendicular to the plane.
Let us next review two scenarios that originally appeared in the paper by Leinaas and Myrheim [2] and that led to a topological explanation of the existence of bosons, fermions and anyons [59]. These are the scenarios of two particles in R 2 and R 3 . In both cases, the configuration space can be parametrised by the centre of mass coordinate R and the relative position r. In terms of the positions of particles, we have Permutation of particles results with changing r to −r, while R remains unchanged, hence In the above formula, RP m−1 := S m−1 / ∼ is the real projective space that is constructed by identifying pairs of opposite points of the sphere. Space (R m − 0) / ∼ can be deformation retracted to RP m−1 by contracting all vectors so that they have length 1. In the case, when m = 2, RP 1 is topologically a circle. Equivalently, R 2 − 0 / ∼ is a cone. Hence, we have so similarly to Example 1, there is a continuum of U (1)-representations of the fundamental group that assign an arbitrary phase factor to the wave function when transported around a non-contractible loop. Note that a loop in the configuration space corresponds to an exchange of particles (see Fig. 1). The case of two particles mov- ing in R 3 has an important difference when compared to the other cases analysed in this paper so far. Namely, there are two non-isomorphic hermitian vector bundles of rank 1 that admit a flat connection. In all previous cases there was only one such vector bundle, which was isomorphic to the trivial vector bundle E 0 ∼ = Cn(X) × C. For m = 3, there is one more flat hermitian vector bundle, which we denote by E . Neglecting the R 3 -component of C 2 (R 3 ), which is contractible, bundles E 0 and E can be constructed from a trivial vector bundle on S 2 in the following way.
Intuitively, nontrivial bundle E is constructed from the trivial vector bundle on S 2 by twisting fibres over antipodal points. In order to determine the statistical properties corresponding to each bundle, we consider U (1) representations of the fundamental group for each vector bundle. The choice of statistical properties for each vector bundle is a consequence of a general construction of flat vector bundles, which we describe in more detail in section 3.3. The fundamental group reads There are two types of loops, the contractible ones and the non-contractible ones, which become contractible when composed twice (see Fig. 2). Bundle E 0 corresponds to the trivial representation of π 1 , while E corresponds to the alternating representation that acts with multiplication by a phase factor e iπ . Consequently, the holonomy group changes the sign of the wave function from E when transported along a non-contractible loop, while the transport of a wave function from the trivial bundle results with the identity transformation. Therefore, bundle E 0 is called bosonic bundle, whereas bundle E is called the fermionic bundle.
As we have seen in the above examples, there is a fundamental difference between anyons in R 2 and bosons and fermions in R 3 . Anyons emerge as different flat connections on the trivial line bundle over C 2 (R 2 ), while fermions and bosons emerge as flat connections on non isomorphic line bundles over C 2 (R 3 ). As we explain in section 3, these results generalise to arbitrary numbers of particles.
In this paper, we approach the problem of classifying complex vector bundles by computing the cohomology groups of configuration spaces over integers. Such strategy has also been used used in [12] to partially classify vector bundles over configuration spaces of distinguishable particles in R m . To this end, we combine the following methods concerning the structure of Vect C (B), the set of complex vector bundles over a paracompact base space B. 2. Classification of vector bundles of rank 1 by the second cohomology group (subsection 3.1).
3. Classification of stable equivalence classes of vector bundles using K-theory (subsections 3.2 and 3.3).
A possible source of new signatures of topology in quantum mechanics would be the existence of non-trivial vector bundles that admit a flat connection. These bundles can be detected by the Chern classes, which for flat bundles belong to torsion components of H 2i (B, Z). We explain this fact and its relation with quantum statistics in section 3.3.

Quantum kinematics on graphs
Configuration spaces of indistinguishable particles on graphs are defined as Example 2 Configuration space of two particles on graph Y . In Y × Y there are 9 two-cells. Six of them are products of distinct (but not disjoint) edges of Y . Their intersect with ∆ 2 is a single point, which we denote by (2,2). The three remaining two-cells are of the form e × e. They have the form of squares, which intersect ∆ 2 along the diagonal. Graph Y and space C 2 (Y ) are shown on Fig. 3. The fact that Cn(Γ ) is composed of pieces that are locally isomorphic to R n is the key property that allows one to define quantum kinematics as gluing the local quantum kinematics on R n . Namely, the momentum operator on (e 1 × e 2 × · · · × en − ∆n)/Sn has n components that are defined aŝ We may define orthonormal coordinates and connection coefficients on each n-cell separately. For each n-cell we require that the connection 1-form Γ = n i=1 α i is closed, hence locally the connection is flat. In order to impose global flatness of the considered bundle, we require that the parallel transport does not depend on the homotopic deformations of curves that cross different pieces of Cn(Γ ). This requirement imposes conditions on the parallel transport operators along certain edges (1-dimensional cells) of Cn(Γ ). To see this, we need the following theorem by Abrams [32].
Theorem 1 Fix n -the number of particles. If Γ has the following properties: i) each path between distinct vertices of degree not equal to 2 passes through at least n−1 edges, ii) each nontrivial loop passes through at least n + 1 edges, then Cn(Γ ) deformation retracts to a CW -complex Dn(Γ ), which is a subspace of Cn(Γ ) and consists of the n-fold products of disjoint cells of Γ .
Complex Dn(Γ ) is called Abram's discrete configuration space and we elaborate on its construction in section 4. For the construction of quantum kinematics, we only need the existence of the deformation retraction. This is because under this deformation, every loop in Cn(Γ ) can be deformed to a loop in Dn(Γ ) ⊂ Cn(Γ ), which has a nicer structure of a CW -complex. Therefore, we only need to consider the parallel transport along loops in Dn(Γ ). Furthermore, every loop in Dn(Γ ) can be deformed homotopically to a loop contained in the one-skeleton of Dn(Γ ). The problem of gluing connections between different pieces of Cn(Γ ) becomes now discretised. Namely, we require that the unitary operators that describe parallel transport along the edges of Dn(Γ ) compose to the identity operator whenever the corresponding edges form a contractible loop. In other words, By σ 1 → σ 2 → · · · → σ l we denote the path constructed by travelling along 1-cells σ i in Dn(Γ ). This is a closed path whenever σ l ∩ σ 1 = ∅.
More formally, we classify all homomorphisms ρ ∈ Hom(π 1 (Cn(Γ )), U (k)) and consider the vector bundles that are induced by the action of ρ on the trivial principal U (k)-bundle over the universal cover of Cn(Γ ). For more details, see section 3. Therefore, the classification quantum kinematics of rank k on Cn(Γ ) is equivalent to the classification of the U (k) representations of π 1 (Dn(Γ )). The described method of classification of quantum kinematics in the case of rank 1 becomes equivalent to the classification of discrete gauge potentials on Cn(Γ ) that were described in [21].
Example 3 Quantum kinematics of rank 1 of two particles on graph Y . The two-particle discrete configuration space of graph Y consists of 6 edges that form a circle (Fig. 4). Therefore, any non-contractible loop in C 2 (Y ) is homotopic with D 2 (Y ). The classification of kinematics of rank 1 boils down to writing down the consistency relation for U (1) operators arising from the parallel transport along the edges in D 2 (Y ).

These operators are just phase factors
The parallel transport of a wave function results witĥ This is reflected in the fact that π 1 (C 2 (Y )) = Z.

Methodology
All topological spaces that are considered in this paper have the homotopy type of finite CW complexes. This is due to the following two theorems.
Theorem 2 [32,20] The configuration space of any graph Γ can be deformation retracted to a finite CW complex, which is a cube cumplex.
Theorem 3 [51,52] The configuration space of n particles in R k has the homotopy type of a finite CW -complex.
One of the central notions in the description of quantum statistics is the notion of the fundamental group. Importantly, the fundamental group of a finite CW complex is finitely generated [53]. This means, that in all scenarios that are relevant in this paper, the fundamental group can be described by choosing a finite set of generators a 1 , . . . , ar and considering all combinations of generators and their inverses, subject to certain relations π 1 (X) = a 1 , a 2 , . . . , ar : W 1 (a 1 , . . . , ar) = e, . . . , W R (a 1 , . . . , ar) = e .
Relations {W i } have the form of words in a 1 , . . . , ar. Some relevant fundamental groups groups are the fundamental group for the n-particle configuration space of R 3 , which is the permutation group Sn, the fundamental group for the n-particle configuration space of R 2 , which is the braid group on n strands Brn, the fundamental group for the n-particle configuration space of a graph, which is a graph braid group [28,29].
Moreover, graph configuration spaces and Cn(R 2 ) are Eilenberg-MacLane spaces of type K(G, 1), i.e. the fundamental group is their only non-trivial homotopy group.
Such spaces are also called the aspherical spaces.
In this paper we focus on calculation of cellular homology of graph configuration spaces. This groups are finitely generated abelian groups, i.e. have the following form are natural numbers such, that p i divides p i+1 for all i. Number K is called the rank of H d (C, Z), and is equal to the dth Betti number of complex X.
The cyclic part of H d (C, Z) is called the torsion part and denoted by T (H d (C, Z)) or T d (C, Z). An important theorem that we will often use reads: The ranks of H k (X, Z) and H k (X, Z) are equal and the torsion of H k (X, Z) is equal to the torsion of H k−1 (X, Z).

Vector bundles and their classification
The main motivation for studying (co)homology groups of configuration spaces comes from the fact that they give information about the isomorphism classes of vector bundles over configuration spaces. In the following section, we review the main strategies of classifying vector bundles and make the role of homology groups more precise. Throughout, we do not assume that the configuration space is a differentiable manifold, as the configuration spaces of graphs are not differentiable manifolds. We only assume that Cn(X) has the homotopy type of a finite CW -complex. This means that Cn(X) can be deformation retracted to a finite CW -complex. As we explain in section 4, configuration spaces of graphs are such spaces. The lack of differentiable structure means, that the flat vector bundles have to be defined without referring the notion of a connection and all the methods that are used have to be purely algebraic. We provide such an algebraic definition of flat bundles in section 3.3.
In this paper, we consider only complex vector bundles π : E → B, where E is a total space and B is the base. Two vector bundles are isomorphic iff there exists a homeomorphism between their total spaces, which preserves the fibres. If two vector bundles belong to different isomorphism classes, there is no continuous function, which transforms the total spaces to each other, while preserving the fibres. Hence, the wave functions stemming from sections of such bundles must describe particles with different topological properties. The classification of vector bundles is the task of classifying isomorphism classes of vector bundles. The set of isomorphism classes of vector bundles of rank k will be denoted by E K k (B).
Before we proceed to the specific methods of classification of vector bundles, we introduce an equivalent way of describing vector bundles, which involves principal bundles (principal G-bundles). A principal G-bundle ξ : P → B is a generalisation of the concept of vector bundle, where the total space is equipped with a free action of group G 1 and the base space has the structure of the orbit space B ∼ = P/G. Fibre π −1 (p) is isomorphic to G is the sense that map π : P → B is G-invariant, i.e. π(ge) = π(e). Moreover, all relevant morphisms are required to be G-equivariant. The set of isomorphism classes of principal G-bundles over base space B will be denoted by While interpreting sections of vector bundles as wave functions, we need the notion of a hermitian product on E. This means, that we consider hermitian vector bundles, i.e. bundles with hermitian product · , · p on fibres π −1 (p), p ∈ B, that depends on the base point and varies between the fibres in a continuous way. Choosing sets of unitary frames, we obtain a correspondence between hermitian vector bundles and principal U (k)-bundles. If the base space is paracompact, any complex vector bundle can be given a hermitian metric [41]. Using the fact that principal U (k)-bundles corresponding to different choices of the hermitian structure are isomorphic [41], we have the following bijection From now on, we will focus only on the problem of classification of principal U (k)bundles.

Universal bundles and Chern classes
Recall, that all vector bundles of rank k over a paracompact topological space can be obtained from a vector bundle, which is universal for all base spaces. This is done in and the corresponding universal bundle is denoted by γ k C . Therefore, any principal U (k)-bundle over a paracompact Hausdorff space B can be written as More applications of Chern classes and cohomology ring H * (B, Z) follow in the remaining parts of this section. In particular, they appear in K-theory and while studying characteristic classes of flat vector bundles.

Reduced K-theory
We start with recalling the definition of stable equivalence of vector bundles.
Definition 1 Vector bundles ξ and ξ are stably equivalent ξ ∼s ξ iff The set of stable equivalence classes of vector bundles over a compact Hausdorff space has the structure of an abelian group, which is called the reduced Grothendieck group K(B). If the base space has the homotopy type of a finite CW -complex, groupK(B) fully describes isomorphism classes of vector bundles that have a sufficiently high rank [49]. This concerns vector bundles, whose rank is in the stable range, i.e. is greater than or equal to where x denotes the smallest integer that is greater than or equal to x. The set of for all k ≥ ks and equal to E C ks (B). Therefore, The relation between reduced K-theory and cohomology is phrased via the Chern character, which induces isomorphism fromK(B) to H * (B, Q), when B has the homotopy type of a finite CW -complex. As a consequence, the classification of vector bundles in the stable range asserts that on condition that the even integral cohomology groups of B are torsion-free. In the case, when there is non-trivial torsion in H * (B, Z), torsion ofK(B) is determined by the Atiyah-Hirzebruch spectral sequence [19]. However, the correspondence between torsion of even cohomology andK(B) is not an isomorphism. In particular, torsion iñ K(B) can vanish, despite the existence of nonzero torsion in H 2i (B, Z). Finally, we note that stable equivalence of vector bundles is physically important in situations, when one has access only to Chern classes or other topological invariants stemming from Chern classes, e.g. the Chern numbers. This is because Chern classes of stably equivalent vector bundles are equal.

Flat bundles and quantum statistics
In this section, we describe the structure of the set of flat principal G-bundles over base space B. More precisely, we consider the set of pairs (ξ, A), where ξ is a principal G-bundle, and A is a connection 1-form on ξ. We divide the set of such pairs into equivalence classes [ξ, A] that consist of vector bundles isomorphic to ξ and the set We use this relation to explain some key properties of quantum statistics that were sketched in the introduction of this paper. Recall the description of the moduli space of flat connections in the case, when B is a smooth manifold. Having fixed a principal connection H on P , we consider parallel transport of elements of P around loops in B. Parallel transport around loop γ ⊂ B is a morphism of fibres Γγ : π −1 (b) → π −1 (b), which assigns the end point of the horizontal lift of γ (denote it byγ) to its initial point Because fibres are homogeneous spaces for the action of G, for every choice of the initial point p =γ(0) there is a unique group element g ∈ G such thatγ(1) = gp. We denote this element by holp(H, γ) and call the holonomy of connection H around loop γ at point p. Moreover, by the G-equivariance of the connection, we get that This means that holgp(H, γ) = g −1 holp(H, γ)g. If connection H is flat, the parallel transport depends only on the topology of the base space [56], i.e. i) Γγ depends only on the homotopy class of γ, ii) parallel transport around a contractible loop is trivial, iii) parallel transport around two loops that have the same base point is the composition of parallel transports along the two loops holonomies at different points from the same fibre differ only by conjugation in G, it is not necessary to specify the choice of the initial point. Instead, we consider map There is one more symmetry of this map that we have not discussed so far, namely the gauge symmetry. A gauge transformation is a map f : P → G, which is G-equivariant, i.e. f (gp) = g −1 f (p)g. A gauge transformation induces an automorphism of P , which acts as p → f (p)p. Consequently, transformation f induces a pullback of connection forms. It can be shown that map S H is gauge invariant [56], i.e. depends only on the gauge equivalence class of connection H.
An important conclusion regarding flat bundles on spaces that do not have a differential structure comes from the second part of correspondence (4). This is the reconstruction of a flat principal bundle from a given homomorphism Hom(π 1 (B), G).
It turns out that any flat bundle over B can be realised as a particular quotient bundle of the trivial bundle over the universal cover of B. In order to formulate the correspondence, we first introduce the notion of a covering space and a universal cover 2 . The following theorem is also a definition of a flat principal bundle for spaces that are not differential manifolds.
Theorem 5 Any flat principal G-bundle P → B can be constructed as the following quotient bundle of the trivial bundle over the universal cover of B.
In the above formula, group π 1 (B) acts onB via deck transformations. Action on G is defined by picking a homomorphism ρ : π 1 (B) → G. Then the action reads ag := ρ(a)g for a ∈ π 1 (B), g ∈ G.
Summing up, in order to describe the moduli space of flat G-bundles, one has to classify conjugacy classes of homomorphisms π 1 (B) → G. All spaces that are considered in this paper have finitely generated fundamental group. This fact makes the classification procedure easier. Namely, one can fix a set of generators a 1 , . . . , ar of π 1 (B) and represent them as group elements g 1 , . . . , gr. Matrices g 1 , . . . , gr realise π 1 (B) in G in a homomorphic way iff they satisfy the relations between the generators of π 1 (B). This way, the moduli space of flat connections can be given the structure of an algebraic variety. In other words, we consider map which returns the values of words describing the relations between generators of π 1 (B). Then, We view Q −1 (e, . . . , e) as the zero locus of a set of multivariate polynomials. In general, such a zero locus has many path connected components. This reflects the topological structure of M(B, G). Namely, one can decompose the moduli space of flat connections into a number of disjoint components, that are enumerated by the isomorphism classes of bundles is the space of flat connections on principal bundles from the isomorphism class [ξ] modulo the gauge group. The following fact gives a necessary condition for two flat structures to be non-isomorphic. Example 4 -The moduli space of flat U (1) bundles over spaces with finitely generated fundamental group. As conjugation in U (1) is trivial, we have Moreover, Hom(π 1 (B), U (1)) is the same as the space of homomorphisms from the abelianization of π 1 (B) to U (1). A standard result from algebraic topology says that is the group commutator. H 1 (B, Z) as any finitely generated abelian group decomposes as the sum of a free component and a cyclic (torsion) part Therefore, we can generate H 1 (B, Z) as We represent a i as e ιφi , φ i ∈ [0, 2π[ and the cyclic generators as roots of unity e ι2kiπ/pi , where k i = 0, 1, 2, . . . , p i − 1. This way, we get q i=1 p i connected components in the space of homomorphisms Hom(H 1 (B, Z), U (1)) that are enumerated by different choices of numbers k i . Each connected component is homeomorphic to a p-torus, whose points correspond to phases φ i . In fact, the connected components are in a one-to-one correspondence with isomorphism classes of flat bundles. To see this, recall the fact that set of U (1)-bundles has the structure of a group, which is isomorphic to H 2 (B, Z). Recall that for particles in R 2 and R 3 , we had Hence, the moduli spaces read (see also Fig. 5) Characteristic classes of flat bundles From this point, we can move away from considering connections and use the wider definition of flat G-bundles, which makes sense for bundles over spaces that have a universal covering space. As stated in theorem 5, such flat bundles have the form where we implicitly use a group homomorphism ρ : π 1 (B) → G in the definition of the quotient. For such flat U (n)-bundles over connected CW -complexes we have the following general result about the triviality of rational Chern classes [42].
Theorem 6 Let G be a compact Lie group, B a connected CW -complex and ξ : P → B a flat G-bundle over B. Then, the characteristic homomorphism is trivial.
Remark 31 Theorem 6 in particular means that if B is a finite CW -complex, then by the universal coefficient theorem for cohomology (see e.g. [43]), the image of the characteristic map f * Specifying the above results for U (n)-bundles, we get that the lack of nontrivial torsion in H 2i (B, Z) has the following implications for the stable equivalence classes of flat vector bundles.
Proposition 7 Let B be a finite CW complex. If the integral homology groups of B are torsion-free, then every flat complex vector bundle over B is stably equivalent to a trivial bundle.
Proof If the integral cohomology of B is torsion-free, then by the Chern character we get, that the reduced Grothendieck group is isomorphic to the direct sum of even cohomology of B. Thus, if all Chern classes of a given bundle vanish, this means that this bundle represents the trivial element of the reduced Grothendieck group, i.e. is stably equivalent to a trivial bundle.
Interestingly, in the following standard examples of configuration spaces, there is torsion in cohomology.
1. Configuration space of n particles on a plane. Space Cn(R 2 ) is aspherical, i.e. is an Eilenberg-Maclane space of type K(π 1 , 1), where the fundamental group is the braid group on n strands Brn. Cohomology ring H * (Cn(R 2 ), Z) = H * (Brn, Z) is known [44,63]. Its key properties are i) finiteness -H i (Brn, Z) are cyclic groups, except Some computational techniques are presented in [64,46], but little explicit results are given. Ring [45] and H q (C 3 (R 3 )) = 0 for q > 4. However, it has been shown that there are no nontrivial flat SU (n) bundles over C 3 (R 3 ).
3. Configuration space of n particles on a graph (a 1-dimensional CW -complex Γ ).
Spaces Cn(Γ ) are Eilenberg-Maclane spaces of type K(π 1 , 1). The calculation of their homology groups is a subject of this paper. Group H 1 (Cn(Γ ), Z) is known [22,26] for an arbitrary graph. We review the structure of H 1 (Cn(Γ )) in section 4.1. By the universal coefficient theorem, the torsion of H 2 (Cn(Γ )) is equal to the torsion of H 1 (Cn(Γ )), which is known to be equal to a number of copies of Z 2 , depending on the structure of Γ . We interpret this result as the existence of different bosonic or fermionic statistics in different parts of Γ . The existence of torsion in higher (co)homology groups of Cn(Γ ), which is different than Z 2 , is an open problem. In this paper, we compute homology groups for certain canonical families of graphs. However, the computed homology groups are either torsion-free, or have Z 2 -torsion.
As we have seen while studying the example of anyons, the parametrisation of different path-connected components of the moduli space of flat bundles corresponds physically to changing some fields. On the other hand, while studying the example of particles in

Configuration spaces of graphs
The general structure of configuration spaces of graphs has been introduced in section 1.2. For computational purposes, we use discrete models of graph configuration spaces. By a discrete model we understand a CW -complex, which is a deformation retract of Cn(Γ ). The existence of discrete models for graph configuration spaces enables us to use standard tools from algebraic topology to compute homology groups of graph configuration spaces. In particular, we use different kinds of homological exact sequences. There are two discrete models that we use.
1. Abram's discrete configuration space [32]. The Abram's deformation retract of Cn(Γ ) is denoted by Dn(Γ ). We use Abram's discrete model mainly in the first part of this paper, where we apply discrete Morse theory to the computation of homology groups of some small canonical graphs (section 5.2).
2. The discrete model by Świątkowski [20], that we denote by Sn(Γ ). We use this model in sections 5.3-5.6 to compute homology groups of configuration spaces of wheel graphs and some families of complete bipartite graphs.
Świątkowski model has an advantage over Abram's model in the sense that its dimension agrees with the homological dimension of Cn(Γ ), and as such, stabilises for sufficiently large n. The dimension of Abram's model is equal to n for sufficiently large n. Hence, the Świątkowski model is more suitable for rigorous calculations. However, sometimes it is more convenient to use Abram's model with the help of discrete Morse theory. The computational complexity of numerically calculating the homology groups of Cn(Γ ) for a generic graph is comparable in both approaches.
Abrams discrete model Let us next describe in detail the discrete configuration spaces Dn(Γ ) by Abrams. For the deformation retraction from Cn(Γ ) to Dn(Γ ) to be valid, the graph must be simple and sufficiently subdivided, which means that each path between distinct vertices of degree not equal to 2 passes through at least n − 1 edges, each nontrivial loop passes through at least n + 1 edges.
The discrete configuration space Dn(Γ ) is a cubic complex. The n-dimensional cells in Dn(Γ ) are of the following form.
We denote cells of Dn(Γ ) by the set notation using curly brackets. Lower dimensional cells are described by sets of edges and vertices from Γ , that are mutually disjoint. A d-dimensional cell consists of d edges and n − d vertices. In other words, cells from In particular, when there are not enough pairwise disjoint edges in the sufficiently subdivided Γ , the dimension of the discrete configuration space can be smaller than n.
In order to define the boundary map, we introduce a suitable order on vertices of Γ , following [28,26]. To this end, we choose a spanning tree T ⊂ Γ and fix its planar embedding. We also fix the root * of T by picking a vertex of degree 1 in T . For every v ∈ V (Γ ) there is the unique path in T that joins v and * , called the geodesic gv, * . For every vertex with d(v) ≥ 2 we enumerate the edges adjacent to v with numbers 0, 1, . . . , d(v) − 1. The edge contained in gv, * has label 0. The remaining edges are labelled increasingly, according to their clockwise order starting from edge 0. The enumeration procedure for vertices goes in an inductive manner. The root has number 1. If vertex v has label k and d(v) = 2, the vertex adjacent to v is given label k + 1. Otherwise, if d(v) ≥ 2, the vertex adjacent to v in the lowest direction with vertices that have not been yet labelled is given label kmax + 1, where kmax is the maximal label among all of the already labelled vertices. If d(v) = 1, we look for essential vertices in gv, * and go back to the closest essential vertex that contains a direction with unlabelled vertices. In other words, the vertices are labelled in the clockwise direction. This way every edge is given an initial and terminal vertex that we denote by ι(e) and τ (e) respectively. The terminal vertex is the vertex with the lower index, i.e. τ (e) < ι(e). We can unambiguously specify an edge by calling its initial and terminal vertices, hence we denote the edges by e ι τ . Given a cell from Dn(Γ ) σ = {e 1 , . . . , e d , v 1 , . . . , v n−d }, we order the edges from σ according to their terminal vertices, i.e. τ (e 1 ) < τ (e 2 ) < · · · < τ (e d ). The full boundary of σ is given by the following alternating sum of faces.
For examples, see section 4.1 and section 5. The boundary map reads The boundary map for elements of a higher degree is determined by the Eilenberg-Zilber theorem: There is a canonical basis for S(Γ ), whose elements of degree (d, n) are of the form d + k + n 1 + · · · + n l = n.
The basis elements form a cube complex. In calculations we use the notion of support of a given cell or a chain. The support of a chain b = i p i c i , p i ∈ Z is given by Supp(c i ).
In this paper we will also use a variation of S(Γ ), which we will call the reduced Świątkowski complex with respect to a subset of vertices U ⊂ V (Γ ) and denote bỹ S U (Γ ). In most cases, the reduced complexes lack a canonical basis, however they have a smaller number of generators than S(Γ ). The reduction is done by changing the generators at vertex v to differences of half edges h ij : Intuitively, this means, that effectively, the particles always slide from one half-edge to another without staying at the central vertex. Both reduced and the non-reduced Świątkowski complexes have the same homology groups [39]. From now on, the default complex we will work with is the complex, which is reduced with respect to all vertices of degree one. Intuitively, this means that we do not consider redundant cells, where particles move from an edge to some vertex of valency one. Such complexes have the canonical basis, which corresponds to cells of a cube complex of the form (6). By a slight abusion of notation, we will denote such a default reduced complex by S(Γ ). In other words, from now on

Sv.
For examples, see figure 6. As a direct consequence of the dimension of Sn(Γ ), we get the following fact.
Fact 41 Let Γ be a graph. Then, the following homology groups of Cn(Γ ) vanish.
Vertex blowup In the following, we will explore relations on homology groups that stem from blowing up a vertex of Γ : Γ → Γv (Fig. 7). We borrow this nomenclature and the methodology of this subsection from [39]. We start with the reduced complex with respect to vertex v,S v (Γ ). Any chain b ∈S v (Γ ) can be decomposed in a unique way by extracting the part that involves generators fromSv. In order to do it, we fix a half-edge h 0 ∈ H(v) and write b as Note that chains b 0 and b h belong to S(Γv). We associate two chain maps to the above decomposition. The first map φ is the embedding of any chain b 0 from S(Γv) toS v (Γ ).
Clearly, this map is injective and commutes with the boundary operator.
The other map ψ is the projection of b ∈S v (Γ ) to its h-components. It assigns a number of n − 1-particle d − 1-chains to a n-particle d-chain in the following way Map ψ is surjective, because any chain b ∈ S n−1 (Γv) can be obtained by ψ for exmaple from chain (h 0 − h)b ∈S v n (Γ ). In order to see that ψ is a chain map, consider a cycle c ∈S v n (Γ ). We have Grouping the summands that entirely belong to S n−1 (Γv), we get By the same argument, the second equation implies that ∂c h = 0 for all h ∈ H(v)\h 0 .
We can write down the two maps as a short exact sequence Short exact sequence (7) of chain maps implies the long exact sequence of homology groups . . .
where the connecting homomorphism reads Long exact sequence (8) implies a collection of short exact sequences   It has been shown in [22] that subject to certain relations, cycles c O and c (n) Y generate H 1 (Dn(Γ )) (see also [39] for the proof of an analogous fact for H 1 (Sn(Γ ))).
The fundamental relation between Y -cycles is shown on Fig. 9 and Fig. 10. Fig. 9 The fundamental relation between the two-particle cycle on a Y -graph and the ABcycle and a two-particle cycle c 2 in the lasso graph.
Cycle c (1) AB is the cycle, where one particle goes around the cycle in the lasso graph and the other particle occupies vertex 1.
Cycle c 2 is the cycle, where two particles go around the cycle in lasso.
It is straightforward to check that where S = {e 2 1 , e 4 3 }. Consider next a situation, where two disjoint Y -graphs share one cycle c O and their free ends are connected by a path pv 1 ,v2 , which is disjoint with c O (Fig. 10). In other words, consider an embedding of a graph, which is isomorphic to the Θ-graph 3 . Then, Subtracting both equations, we get But the existence of pv 1 ,v2 gives us that c will be called a Θ-relation. It turns out that considering all Θ-relations stemming from different Θ-subgraphs and relations (10) that express different distributions of particles in the O-cycles as differences of Y -cycles, one can compute the first homology group of Dn(Γ ). Let us next summarise the results concerning the structure of the first homology group of graph configuration spaces. We formulate the results assuming, that the considered graphs are simple. The general form of the first homology group reads where N and L are the numbers of copies of Z and Z 2 respectively. Numbers N and L depend on the planarity and some combinatorial properties of the given graph [22,26]. The Z 2 -components appear, when Γ is non-planar and have the interpretation of different fermionic/bosonic statistics that may appear locally in different parts of a given graph (see [22]).

Calculation of homology groups of graph configuration spaces
This section contains the techniques that we use for computing homology groups of graph configuration spaces. We tackle this problem from the 'numerical' and the 'analytical' perspective. The numerical approach means using a computer code for creating the boundary matrices and then employing the standard numerical libraries for computing the kernel and the elementary divisors of given matrices. The procedures for calculating the boundary matrices of Dn(Γ ), Sn(Γ ) and the Morse complex (see section 5.2) were written by the authors of this paper, based on papers [28,26]. The analytical approach means computing the homology groups for certain families of graphs by suitably decomposing a given graph into simpler components and using various homological exact sequences. Recently in the mathematical community, there has been a growing interest in computing the homology groups of graph configuration spaces. A significant part of the recent work has been devoted to explaining certain regularity properties of the homology groups of Cn(Γ ) [34,35,36,38,37,40].

Product cycles
Considering simultaneous exchanges of pairs of particles on disjoint Y -subgraphs of Γ and the O-type cycles with the remaining particles distributed on the free vertices of Γ , one can construct some generators of H * (Dn(Γ )) or H * (Sn(Γ )). Such cycles are products of 1-cycles, hence are isomorphic to tori embedded in the discrete configuration space. To construct a product d cycle in Dn(Γ ), we choose Y -subgraphs of Γ All the chosen subgraphs must be mutually disjoint.
In an analogous way, we form product cycles in Sn(Γ ).
We study such product cycles for configuration spaces of different graphs and describe relations between them. So far, it has been known, that product cycles generate the second homology of the two particle configuration space of a simple graph [33] and all homology groups for an arbitrary number of particles on tree graphs [18] (see also [30]). In this section, we find new families of graphs, for which product cycles generate some homology groups of their configuration spaces. These cases are all homology groups of the configuration spaces of wheel graphs (section 5.3), all homology groups of the configuration space of graph K 3,3 , except the third homology group (section 5.5), the second homology group of a simple graph, which has at most one vertex of degree greater than 3.
In sections 5.5 and 5.6 we also discuss examples of cycles that are different than tori. In particular, we compute all homology groups of configuration spaces of complete bipartite graphs K 2,p , that are often pointed out in the literature as an unsolved example, where the simple use of product cycles is not sufficient to generate the homology groups.
We show, that some of the generators of H * (Sn(K 2,p )) are cycles of a new type, that have the homotopy type of triple tori.

Discrete Morse theory for Abrams model
In this subsection, we apply a version of Forman's discrete Morse theory [27] for Abram's discrete model that was formulated in [28] (see also [31]). The results are listed in tables 1 and 2.
In general, the discrete Morse theory relies on constructing a discrete gradient flow F , which is a linear map mapping d-chains to d-chains. Moreover, map F has the property, that for any chain c, F r+1 (c) = F r (c) for some r. The Morse complex is the chain complex of chains invariant under F . Some chosen homology groups over integers for configuration spaces of exemplary graphs have been calculated using the discrete gradient vector field described in [28] and its computer implementation done by the authors of this paper. The results are collected in tables 1 and 2. Table 2 presents the results for the second and third homology groups for graphs from the Petersen family ( fig. 11). These graphs serve as examples, where torsion in higher homology groups appears. Interestingly, the torsion subgroups are always equal to a number of copies of Z 2 . This phenomenon can be explained by embedding a nonplanar graph in Γ and considering suitable product cycles. The question about the existence of torsion different than Z 2 in higher homologies remains open.

Wheel graphs
In this section, we deal with the class of wheel graphs. A wheel graph of order m is a simple graph that consists of a cycle on m − 1 vertices, whose every vertex is connected by an edge (called a spoke) to one central vertex (called the hub). We provide a complete description of the homology groups of configuration spaces for wheel graphs. In particular, we show, that all homology groups are free. Therefore, in addition to tree graphs, wheel graphs provide another family of configuration spaces with a simplified structure of the set of flat complex vector bundles. The general methodology of computing homology groups for configuration spaces of wheel graphs is to consider only the product cycles and describe the relations between them. We justify this approach in subsection 5.4. The simplest example of a wheel graph is graph K 4 , which is the wheel graph of order 4. Let us next calculate all homology groups of graph K 4 and then present the general method for any wheel graph.

Graph K 4
Graph K 4 is shown on figure 12. It is the 3-connected, complete graph on 4 vertices.
Second homology group There are three independent cycles in K 4 graph. These are the cycles that contain the hub and two neighbouring vertices from the perimeter. However, any two such cycles always share some vertices. Hence, there are no tori that come from the products of c O cycles. Hence, the product 2- There are four cycles of the first kind: . Therefore, the second Fig. 11 Graphs that form the Petersen family. homology of the three-particle configuration space is If n > 3, there are still three independent O × Y -cycles, as the differences between distributions of free particles in such cycles can always be expressed as combinations of Y × Y -cycles. To see this, consider the following example. For n = 4, consider the O × Y -cycles that involve cycle c O1 , subgraph Y 1 and one of three possible free vertices (Fig. 13). The cycles are Subtracting the above equations and multiplying the results by c Y1 , we get This means that the differences between distribution of particles in AB-cycles can be expressed as combinations of Y × Y cycles. This fact generalises to n > 4 in a straightforward way.
Consider next all possible ways of choosing two Y -subgraphs. There are six Y × Ycycles modulo the distribution of free particles. Hence, if there are no free particles, i.e. when n = 4, we have H 2 (D 4 (K 4 )) = Z 3 ⊕ Z 6 . If n > 4, we have to take into account the distribution of free particles in Γ − (Y ∪ Y ).
For a sufficiently subdivided graph one always ends up with two connected components (Fig. 14). A Y × Y -cycle involves 4 particles, hence one has to calculate the number of  Adding the contribution from O × Y -cycles, the rank of the second homology group is then given by β 2 (Cn(K 4 )) = 3 + 6(n − 3) = 6n − 15, n ≥ 3.
Higher homology groups The product generators of higher homologies are even simpler than in the case of the second homology. There are only basis cycles of Y × Y × · · · × Ytype. After removing three and four Y -graphs, K 4 graph always disintegrates into 4 and 6 parts respectively. Taking into account the distributions of free particles, we get the following formulae for the Betti numbers.
Because there are maximally four Y -graphs, group H 5 (Cn(K 4 ), Z) is zero.

General wheel graphs
In Table 3 we list Betti numbers of configuration spaces of wheel graphs of order 5, 6 and 7 that were calculated using the discrete Morse theory.  Table 3 Betti numbers of configuration spaces for chosen wheel graphs computed using the discrete Morse theory. In all cases the calculated groups were torsion-free.
Second homology Since there are no pairs of disjoint O-cycles in wheel graphs, we have β 2 (D 2 (Wm)) = 0. When n = 4, we have to count the Y × Y cycles in. Let us divide the Y × Y cycles into two groups: i) cycles, where one of the subgraphs is Y h and ii) cycles, where both subgraphs lie on the perimeter. There are no relations between the cycles within group i) and no relations between the cycles within group ii). However, there are some relations between the cycles of type i) and type ii). The relations occur between cycles Y h × Y and Y × Y , when subgraphs Y h and Y do not share any edges of the graph (like on Fig. 15b)). Then, as on Fig. 10, cycles c Y h and c Y are in the same homology class in D 2 (Wm − Y ), because they share the same O-cycle and they are connected by a path that is disjoint with Y . Therefore, by multiplying the relation by c Y we get that If m > 4, then for every pair Y × Y h that does not share an edge, one can find subgraph Y on the perimeter which gives rise to such a relation. There are ( m−1 2 ) tori coming from Y -subgraphs from the perimeter. For a fixed Y -subgraph, the contribution from Y × Y h -cycles turns out to be equal to the number of independent cycles in the fan graph, which is formed by removing subgraph Y from the wheel graph [22]. This number is equal to m − 3. Hence, For numbers of particles greater than 4, we have to take into account the distribution of free particles. Removing two Y -subgraphs from the perimeter may result with the decomposition of the wheel graph into at most two components. This happens iff two neighbouring Y -subgraphs have been removed. The number of nonequivalent ways of distributing the particles is n − 3. The number of ways one can choose two neighbouring Y -subgraphs from the perimeter is m − 1. This gives us the contribution of (n − 3)(m − 1). Furthermore, removing a Y -subgraph from the hub and a subgraph from the perimeter always yields two nonequivalent ways of distributing the free particles. The first one being the edge e joining the hub and the central vertex of Y , the second one being the remaining part of the graph, i.e. Wm − (Y Y h e). The contribution is (n − 3)(m − 1)(m − 3). Adding the contribution from O × Y -cycles and from non-neighbouring Yp × Yp-cycles, we get that the final formula for the second Betti number reads Higher homologies In computing the higher homology groups, we proceed in a similar fashion as in the previous section. However, the combinatorics becomes more complicated and in most cases it is difficult to write a single formula that works for all wheel graphs. Let us start with an example of H 3 (Dn(W 5 )). The possible types of product Cycles of the first type arise in W 5 only when graphs Y and Y are neighbouring subgraphs from the perimeter. There are four possibilities for such a choice of Y -subgraphs, hence When n > 5, the free particles can be placed either on the edge joining the Y -subgraphs or on the connected part of W 5 that is created by removing subgraphs Y and Y . By arguments analogous to the ones presented in section 5.3.1, the distribution of free particles on the connected component containing cycle O does not play a role. Hence, the contribution to β 3 is equal to the number of different distributions of free particles on the edge connecting Y and Y and on the connected component. In other words, there are two bins and n − 5 free particles. Hence, the total contribution from O × Y × Y -cycles is 4(n − 4). We split the contribution from Y × Y × Y -cycles into two groups. The first group consists of cycles only from perimeter (Yp × Y p × Y p ), for whom the combinatorial description is straightforward. The number of possible choices of Y -subgraphs is ( 4 3 ) and it always results with the decomposition of W 5 into 3 components. Hence, with n − 6 free particles the number of independent Yp × Y p × Y pcycles is 4( n−4 2 ). In order to determine the number of independent cycles Yp × Y p × Y h (two subgraphs from the perimeter and one from the hub), one has to consider different graphs that arise after removing two Y -subgraphs from the perimeter of W 5 . The number of independent Y h -cycles for a fixed choice of Yp and Y p is the same as in a certain fan graph, which is determined by the choice of the Yp-subgraphs. Choosing Yp and Y p to lie on the opposite sides of the diagonal of W 5 , the resulting fan graph is the star graph S 4 . The free particles outside Yp and Y p can always be moved to the S 4 -subgraph. Hence, the contribution from such cycles is given by the number of independent Y -cycles in S 4 for n − 4 particles. We denote this number by β (n−4) 1 (S 4 ).
The last group of cycles that we have to take into account are Yp × Y p × Y h , where Yp and Y p are neighbouring subgraphs. The resulting fan graph is shown on Fig. 16.
The n − 4 particles that do not exchange on the perimeter subgraphs are distributed between the fan graph and the edge joining Yp and Y p . There have to be at least 2 particles exchanging on a Y h -subgraph of the fan graph. The number of independent Y h -cycles for k + 2 particles on the fan graph is given in the caption under Fig. 16.
After summing all the above contributions, the final formula for the third Betti number reads The fourth Betti number is easier to compute, because removing three Yp-subgraphs Fig. 16 The fan graph that is created after removing two neighbouring Y -subgraphs from the perimeter of W 5 . It has µ = 3 leaves. There are two types of Y -cycles at the hub: a) cycles, where the Y -graph is spanned in three different leaves -the number of such cycles for k + 2 particles is β where the Y -graph is spanned in two different leaves -the number of such cycles for k + 2 particles is k+3 k+1 − 1, see [22]. Let us next generalise the above procedure to an arbitrary wheel graph Wm. The dth Betti number is zero whenever the number of particles is less than 2(d − 1) + 1 = 2d − 1. If n = 2d − 1 the only possible tori come from the products of d − 1 Y -cycles and one O-cycle. The graph also cannot be too small, i.e. the condition m − 3 ≥ d − 1 must be satisfied. Otherwise, there is no cycle that is disjoint with d − 1 Y -subgraphs. Hence, Otherwise, for n = 2d − 1, if the graph is large enough, one has to look at all the possibilities of removing Y -subgraphs from the perimeter and what fan graphs are created. We are interested in the number of leaves (µ) of the resulting fan graph. The number of cycles in such a fan graph with µ leaves is m − 1 − µ. It is a difficult task to list all possible fan graphs for any Wm in a single formula. The results for graphs up to W 7 are shown in Table 4. Using the notation from Table 4, the general formula for β d reads (1) 6 2 (1,1) 9 4 (2) 6 3 (1,1,1) 2 6 (2,1) 12 5 (3) 6 4 (2,2) 3 6 (3,1) 6 6 (4) 6 5 (5) 6 6 (6) 1 6 Table 4 The possibilities of choosing a number of Y -subgraphs from the perimeter of a wheel graph. The groups of Y -subgraphs are denoted by sequences (n 1 , n 2 , . . . , n l ), where l + l i=1 n i ≤ m − 1. A group n i means that n i neighbouring Y -subgraphs were chosen. The groups have to be separated by at least one spoke. For a fixed set of groups there are many possibilities for distributing the remaining Y -subgraphs. The number of possibilities is written in the third column. The number of leaves of the resulting fan graph is written in the fourth column. It is independent on the distribution of the remaining Y -subgraphs and is given by For higher numbers of particles, one has to take into account the Y × Y × · · · × Y cycles and distribution of free particles. If n = 2d, the free particles are only in O × Y × Y × · · · × Y -cycles, where they are distributed between the edges that come from removing a group of Y -subgraphs. Group n i gives n i − 1 edges. Hence, groups (n 1 , . . . , n l ) give |n| − l edges. The final formula reads where #n is the number of groups in n (the length of vector n). The contribution β (2) 1 (Sµ n ) + (µn − 1)(m − 1 − µn) comes from the number of independent Y h -cycles in the relevant fan graph. The general formula when n > 2d reads as follows.
The first sum describes the O × Y × Y × · · · × Y -cycles and the distribution of the free n − 2d + 1 particles. Second sum is the number of Y × Y × · · · × Y -cycles, where all Ysubgraphs lie on the perimeter -there are n − 2d free particles. The last sum describes the number of independent Y h × Yp × · · · × Yp-cycles. Here we used the formula for the number of Y h -cycles for n particles on a fan graph with µ leaves and m − 1 spokes [22] n (n) Sometimes, in formula (13), we get to evaluate ( 0 0 ) = 1, ( 0 −1 ) = 0, ( −1 −1 ) = 1. (S m−1 ), n ≥ 2m.

Wheel graphs via Świątkowski discrete model
In this section we show, that the homology of configuration spaces of wheel graphs is generated by product cycles. The strategy is to consider two consecutive vertex cuts that bring any wheel graph to the form of a linear tree.  Throughout, we use the knowledge of generators of the homology groups for tree graphs to construct a set of generators for net graphs and wheel graphs. Translating the results of paper [18] to the Świątkowski complex, we have, that the generators of H d (S(Tm)) are of the form subject to relations This means that computing the rank od H d (S(Tm)) boils down to considering all possible distributions of n − 2d free particles among the connected components of where n i is the number of particles on ith connected component of In the first step, we connect two endpoints of Tm to obtain net graph Nm (Fig. 18).
The Betti numbers read Proof Long exact sequence corresponding to vertex blow-up from figure 18 reads Let us next show that the connecting homomorphism δ is in this case injective. Map δ n,d acts on generators (15) as Hence, the rank of H d (Sn(Nm)) is equal to rk(coker n,d ) = β d (Sn(Tm))−β d (S n−1 (Tm)).
Let us next consider the homology sequence associated with the vertex blow-up from W m+1 to Nm ( fig. 17).
We next describe the kernel of map δ. Our aim is to show that it is free abelian, which in turn gives us that the short exact sequences for H d (Sn(W m+1 )) split and yield H d (Sn(W m+1 )) ∼ = coker(δ n,d ) ⊕ ker(δ n−1,d ). Map δ n,d assigns to generators (16) of H d (Sn(W m+1 )) the differences of generators derived from a given generator by adding one particle to a connected component of In order to write down the action of map δ, let us first establish some notation. The connected components of Nm − (v h (Y 1 ) ∪ · · · ∪ v h (Y d )) are either isomorphic to edges or to linear tree graphs. The number of connected components that are edges, which have one vertex of degree one in Nm is equal to d. The number of the remaining connected components is always equal to d, but their type depends on the distribution of subgraphs Y 1 , . . . , Y d in Nm. The situations, that are relevant for the description of ker δ are those, where a particle is added by map δ to two connected components, which contain an edge, which before the blow-up was adjacent to the hub of W m+1 . There are at most 2d such components, as removing the hub-vertices of two neighbouring Ysubgraphs of Nm yields a connected component of the edge type, which is not adjacent to the hub of W m+1 . We label these components by numbers 1, . . . , l (we always have d ≤ l ≤ 2d) and the occupation numbers of these components are n 1 , . . . , n l . We . The last type of elements of ker δ n,d are combinations of generators, that when acted upon by δ n,d , compose to the boundary of a Y -cycle centred at the hub of W m+1 . Such kernel elements correspond to cycles The precise form of such kernel elements is the following.
where i < j. In order to manage the relations between the above kernel elements, we use the already mentioned fact, that they are in a one-to-one correspondence with 1-cycles (O-cycles and Y -cycles) in a configuration space of the disconnected graph . Fan graphs are planar, hence by equation (12) there is no torsion in ker δ n,d . Hence, H d (Sn(W m+1 )) is torsion-free and short exact sequence for H d (Sn(W m+1 )) gives in this case The computation of ranks of kernels of maps δ n,d is a combinatorial task, which has been accomplished using the correspondence with cycles in configuration spaces of fan graphs in subsection 5.3.2.
Graph K 3,3 has the property, that all its vertices are of degree three. High homology groups of graphs with such a property have been studied in [39]. In particular, we have the following result.
Theorem 8 Let Γ be a simple graph, whose all vertices have degree 3. Denote by N the number of vertices of graph Γ and label the vertices by labels 1, . . . , N . Moreover, denote by Y = {Y 1 , . . . Y N } the set of Y -subgraphs of Γ such, that the hub of Y k is vertex k. Group H N (Sn(Γ )) is freely generated by product cycles Group H N −1 (Sn(Γ )) is generated by product cycles of the form The above generators are subject to relations As we show in section 5.7, the second homology group of configuration spaces of such graphs is also generated by product cycles. Later in this section, by comparing the ranks of homology groups computed via the discrete Morse theory, we argue, that H 4 (Cn(K 3,3 )) is also generated by product cycles. Interestingly, in H 3 (Cn(K 3,3 )) there is a new non-product generator. Using this knowledge, we explain the relations between the product and non-product cycles that give the correct rank of H 3 (Cn(K 3,3 )).
Second homology group There are no pairs of disjoint cycles in K 3,3 , hence the product part for n = 2 is empty. When n = 3, there are 12 O × Y -cycles. This can be seen by choosing the Y -graph centered at vertex 1 on Fig. 19b) -there are 2 cycles disjoint with such a Y -subgraph. There are 6 Y -subgraphs in K 3,3 , hence we get the number of O×Ycycles. One checks by a straightforward calculation, that 8 of them are independent. Hence, When n = 4, there are new product cycles of the Y × Y -type. There are ( 6 2 ) = 15 cycles of this type, however there are relations between them. Such relations between the Y × Y -cycles arise, when one of the cycles is in relation with a different Y -cycle. This happens only when we have a situation as on Fig. 10. Therefore, cycles of the Y × Y -type, where the hubs of the Y -subgraphs, are connected by an edge, are all independent (Fig. 20a)). The number of such cycles is 9. The relations occur between Y ×Y -cycles, where the hubs of the subgraphs are not connected by an edge (Fig. 20b)).
There are 6 such cycles. The number of relations is 4. To see this, consider Y -subgraph, whose hub is vertex 1 (Fig. 19). Denote this subgraph by Y 1 . It is straightforward to Analogous relations for Y -subgrphs that lie on the same side of the K 3,3 graph as Y 1 (see Fig. 19a)) read From the above equations only two are independent. Similar situation happens for relations between pairs of graphs from the other side. The complete set of relations reads Therefore, β 2 (D 4 (K 3,3 )) = 8 + 9 + 2 = 19.
For n > 4, we have to take into account the distribution of free particles. Whenever two non-neighbouring Y -subgraphs are considered, all distributions of free particles are equivalent (Fig. 20b)). When the subgraphs are adjacent, there are two different parts of K 3,3 , where the particles can be distributed, see Fig. 20a). This gives the formula β 2 (Dn(K 3,3 )) = 8 + 2 + 9(n − 3) = 9n − 17, n ≥ 4. Higher homology groups Let us first look at the third homology group. The are no product cycles for n = 4 however, from the Morse theory for the subdivided graph from Fig. 21 we have β 3 (D 4 (K 3,3 )) = 1. The number of all such cycles is equal to the number of pairs of adjacent Y -subgraphs, which is 9. Adding the properly embedded generator of H 3 (D 4 (K 3,3 )), we get β 3 (D 5 (K 3,3 )) = 10.
For n ≥ 6, all Y × Y × Y -cycles are independent. Consider two ways of choosing three Y -subgraphs. The first way is to remove two Y -graphs from the same side and one from the opposite side. This results with the partition of K 3,3 into three components (Fig. 22a)). Removing three Y -graphs from the same side splits K 3,3 into three parts (Fig. 22b)). Therefore, β 3 (Dn(K 3,3 )) = 1 + 9(n − 4) + 6 3 n − 4 2 , n ≥ 6. The product contribution to higher homology groups requires considering different choices of Y -subraphs. There are no Y × Y × · · · × Y × O-cycles in Hp(Dn(K 3,3 )) for p ≥ 4. As direct computations using discrete Morse theory show, there are also no non-product generators (see table 1). Therefore, only Y × Y × · · · × Y -cycles contribute to Hp(Dn(K 3,3 )) for p ≥ 4. Removing four Y -graphs from K 3,3 always results with the splitting into 5 parts, removing five Y -graphs gives 7 parts and removing all six Y -graphs gives 9 parts. Summing up, All homology groups higher than H 6 are zero for any number of particles.

Triple tori in Cn(K 2,p )
In this section we study a family of graphs, where some cycles generating the homology groups of the n-particle configuration space are not product. This is the family of complete bipartite graphs K 2,p (see figure 23a). The first interesting graph from this family is K 2,4 . As we show below, its 3-particle configuration space gives rise to a 2-cycle, which is a triple torus. It turns out, that such triple tori together with products of Y cycles generate the homology groups of Cn(K 2,p ). The most convenient discrete model for studying Cn(K 2,p ) is the Świątkowski model. In fact, we study the Świątkowski configuration space of graph Θp (see 23b), which is topologically equivalent to K 2,p , but it has the advantage, that its discrete configuration space is of the optimal dimension.
This in turn means, that H 2 (Cn(K 2,p )) as the top homology group is a free group.
The first homology group can be computed using the methods of papers [22,26].

Lemma 2
The first homology group of Cn(K 2,p ) is equal to Z p(p−1) for n ≥ 2 and p − 1 for n = 1.

Lemma 3
The Euler characteristic of Sn(K 2,p ) for n ≥ 3 and p ≥ 3 is On the other hand, χ(Sn(K 2,p )) = 1 − β 1 (Sn(K 2,p )) + β 2 (Sn(K 2,p )). Therefore, we compute the second Betti number of Cn(K 2,p ) as In the remaining part of this section we describe the generators of H 2 (Cn(K 2,p )) and the relations that lead to the above formula. We represent them in terms of 2-cycles in Sn(Θp).
This can be seen by comparing the number of cycles of the above form with β 2 (Sn(Θ 3 )) from formula 17. In both cases the answer is the number of distributions of n − 2 particles among edges e 1 , e 2 , e 3 (the problem of distributing n − 2 indistinguishable balls into 3 distinguishable bins), which is ( n−2 2 ) = 1 2 (n − 2)(n − 3).
From now on, we denote the Y -cycles as Cycle c ijk is the Y -cycle of the Y -subgraph, whose hub vertex is v and which is spanned on edges e i , e j , e k . Cycle c ijk corresponds to an analogous Y -subgraph, whose hub is v . Hence, all Θ-cycles generate a subgroup of H 2 (Sn(Θp)), which is isomorphic to Z ( p−1 3 ) .
The last type of relations we have to account for 5 are the new relations between products of Y -cycles. Again, many relations of type 11 can be written by picking different Θ 5 subgraphs.
Similarly as in the case of relations 20, the linearly independent ones are chosen by fixing e 1 to be the common edge of the Θ 5 subgraphs. Hence, the number of linearly independent relations is ( p−1 4 ). In particular, we have where (β 1 (C 2 (Sp))) 2 is the number of independent product cycles after taking into the account the relations within the two opposite star subgraphs. All the above relations are inherited by the cycles in Sn(Θp) after multiplying them by a suitable polynomial in the edges of Θp. In this way, they yield equation (17).
In this section we prove the following theorem.
In the proof we use the Świątkowski discrete model. The strategy of the proof is to first consider the blowup of the vertex of degree greater than 3 and prove theorem 12 for graphs, whose all vertices have degree at most 3. For such a graph, we choose a spanning tree T ⊂ Γ . Next, we subdivide once each edge from E(Γ ) − E(T ). We prove the theorem inductively by showing in lemma 4, that the blowup at an extra vertex of degree 2 does not create any non-product generators. The base case of induction is obtained by doing the blowup at every vertex of degree 2 in Γ − T . This way, we obtain graph, which is isomorphic to tree T and we use the fact, that for tree graphs the homology groups of Sn(T ) are generated by products of Y -cycles.
Lemma 4 Let Γ be a simple graph, whose all vertices have degree at most 3. Let T be a spanning tree of Γ . Let v ∈ V (Γ ) be a vertex of degree 2 and Γv the graph obtained from Γ by the vertex blowup at v. If H 2 (Sn(Γv)) is generated by product cycles, then H 2 (Sn(Γ )) is also generated by product cycles.
Proof Long exact sequence corresponding to the vertex blow-up reads . . . We aim to show, that the corresponding long exact sequence 0 → coker (δ n,2 ) − → H 2 S v n (Γ ) − → ker (δ n,1 ) → 0 splits. To this end, we construct a homomorphism f : ker (δ n,1 ) → H 2 S v n (Γ ) such, that Ψ n,2 • f = id ker(δn,1) . In the construction we use the explicit knowledge of elements of ker (δ n,1 ). Such a knowledge is accessible, as we know the generating set of H 1 (S n−1 (Γv)) -because all vertices of Γv have degree at most 3, it consists of Y -cycles and O-cycles, subject to the Θ-relations (equations (11) and (10) This way, we obtained, that H 2 S v n (Γ ) ∼ = ker (δ n,1 ) ⊕ coker (δ n,2 ) and that elements of ker (δ n,1 ) are represented by product c O ⊗ c Y cycles. By the inductive hypopaper, elements of coker (δ n,2 ) are the product cycles that generate H 2 (S n−1 (Γv)) subject to relations ce ∼ ce .
The last step needed for the proof of theorem 12 is showing, that the blowup of Γ at the unique vertex of degree greater than 3 does not create any non-product cycles.
Here we only sketch the proof of this fact, which is analogous to the proof of lemma 4. Namely, using the knowledge of relations between the generators of H 1 (S n−1 (Γv)), one can show, that the elements of ker (δ n,1 ) are of two types: i) the ones, that are of the form ∂(c ⊗ b p(c) ), where [c] ∈ H 1 (S n−1 (Γv)) and b p(c) is the 1-cycle corresponding to path p(c) ⊂ Γv, which is disjoint with Supp(c) and whose boundary are edges incident to v, ii) pairs of cycles of the form (c(e j − e 0 ), c(e 0 − e j )), where e 0 , e i , e j are edges incident to v and [c] ∈ H 1 (S n−2 (Γv)). Such pairs are mapped by δ n,1 to c ⊗ ((e j − e 0 )(e 0 − e i ) + (e 0 − e i )(e 0 − e j )), which is equal to ∂(c ⊗ c 0ij ), where c 0ij is the Y -cycle corresponding to the Y -graph in Γ centred at v and spanned by edges e 0 , e i , e j . Next, in order to show splitting of the homological short exact sequences, we consider a homomorphism f : ker (δ n,1 ) → H 2 S v n (Γ ) , for which Ψ n,2 • f = id ker(δn,1) . Such a homomorphism maps [c] to [c ⊗ c O p(c) ], where O p(c) is the cycle, which contains path p(c) and vertex v. Pairs ([c(e j − e 0 )], [c(e 0 − e j )]) are mapped by f to cycles c ⊗ c 0ij . We obtain, that H 2 S v n (Γ ) ∼ = ker (δ n,1 ) ⊕ coker (δ n,2 ), where the generators of ker (δ n,1 ) are in a one-to-one correspondence with the product homology classes of H 2 S v n (Γ ) described above. Elements of coker (δ n,2 ) are also represented by product cycles. These cycles are the generators of H 2 (Sn(Γv)) subject to relations ce 0 ∼ ce i , i = 1, . . . , d(v), where e 0 , e 1 , . . . , e d(v) are edges incident to v.
The task of characterising all graphs, for which H 2 (Sn(Γ )) is generated by product cycles requires taking into account the existence of non-product generators from section 5.6. As we show in section 5.6 the existence of pairs of vertices of degree greater than 3 in the graph implies that there may appear some multiple tori in the generating set of H 2 (Cn(Γ )) stemming from subgraphs isomorphic to graph K 2,4 . Furthermore, the class of graphs, for which higher homologies of Cn(Γ ) are generated by product cycles is even smaller. Recall graph K 3,3 whose all vertices have degree 3, but H 3 (Cn(K 3,3 )) has one generator, which is not a product of 1-cycles (see section 5.5).

Summary
In the first part of this paper we explained, that the quantum statistics on a topological space X are classified by the conjugacy classes of unitary representations of the fundamental group of the configuration space Cn(X). On the other hand, every such a unitary representation gives rise to a flat complex vector bundle over space Cn(X). We interpret different isomorphism classes of flat complex vector bundles over Cn(X) as fundamentally different families of particles. Among these families we find for example bosons and anyons (corresponding to the trivial flat bundle) and fermions, that correspond to a non-trivial flat bundle. We argue, that, the existence of more than only these two isomorphism classes is possible. However, the construction of non trivial flat bundles for X = R 2 or X = R 3 is difficult, hence some simplified mathematical models are needed. This motivates the study of configuration spaces of particles on graphs, which are computationally more tractable. Topological invariants, that give some information about the structure of the set of complex vector bundles over Cn(X) are the homology groups of configuration spaces. In particular, Chern characteristic classes map the flat vector bundles to torsion components of the homology groups with coefficients in Z.
In the second part of this paper, we compute homology groups of configuration spaces of certain families of graphs. We summarise the computational results as follows.
-Configuration spaces of tree graphs, wheel graphs and complete bipartite graphs K 2,p have no torsion in their homology. This means, that the set of flat bundles over configuration spaces of such graphs has a simplified structure, namely every flat vector bundle is stably equivalent to a trivial vector bundle. Hence, these families of graphs are good first candidates for a class of simplified models for studying the properties of non-abelian statistics. -Computation of the homology groups of configuration spaces of some small canonical graphs via the discrete Morse theory shows, that in some cases there is a Z 2 -torsion in the homology. However, we were not able to provide an example of a graph, which has a torsion component different than Z 2 in the homology of its configuration space. -It is a difficult task to accomplish a full description of the homology groups of graph configuration spaces using methods presented in this work. One fundamental obstacle is that such a task requires the knowledge of possible embeddings of d-dimensional surfaces in Cn(Γ ), which generate the homology. However, cycles generating the homology in dimension 2 of graph configuration spaces have the homotopy type of tori or multiple tori. This fact allowed us to find all generators of the second homology group of configuration spaces of a large family of graphs in section 5.7.