The entropic boundary law in BF theory

We compute the entropy of a closed bounded region of space for pure 3d Riemannian gravity formulated as a topological BF theory for the gauge group SU(2) and show its holographic behavior. More precisely, we consider a fixed graph embedded in space and study the flat connection spin network state without and with particle-like topological defects. We regularize and compute exactly the entanglement for a bipartite splitting of the graph and show it scales at leading order with the number of vertices on the boundary (or equivalently with the number of loops crossing the boundary). More generally these results apply to BF theory with any compact gauge group in any space-time dimension.


Introduction
Entropy is a key notion in the search for a quantum gravity, both as the thermodynamic quantity useful to probe the physics and potential phenomenology of the theory and as the measure of information useful to identify the physical degrees of freedom and their dynamics. Research has focused on the particular case of black holes and has lead to the holographic principle directly relating geometric quantities (the area) to the entropy in quantum gravity.
In the context of Loop Quantum Gravity (see [2] for reviews), most of the black hole entropy calculations have been performed in the framework of isolated horizons following the seminal work by Ashtekar, Baez and Krasnov [3]. Assuming precise boundary conditions for the gravitational fields on the horizon, they count the number of (kinematical) boundary states consistent with fixing the value of the area. Here we would like to get rid of the classical boundary: instead of considering a spin network state with a boundary specified classically, we look at a closed spin network and an arbitrary bipartite splitting into inside/outside regions (see e.g [4]). The aim is to compute the entanglement between these two parts of the spin network for a physical quantum geometry state solving the Hamiltonian constraint. The first step that we take here is to work this out for BF theory instead of gravity. Indeed BF theories are topological field theories lacking local degrees of freedom, thus allowing for an exact quantization (see e.g [1]). In this context, we know the physical states solving all the constraints and can compute the entanglement explicitly. This turns out to be similar to ground state entanglement calculations in some spin models developed for topological quantum computation [5,6].
The motivation to analyze BF theory is that it is very close to gravity. First, in three space-time dimensions, 3d gravity is actually a topological BF theory with the Lorentz group as gauge group. Second, in four space-time dimensions, gravity can be reformulated as an constrained BF theory and we can work on a quantization scheme with quantum BF theory as the starting point. Of course, the fact that gravity has local degrees of freedom should matter in the end. Nevertheless, studying BF theory should allow us to develop mathematical tools and procedures later useful for loop gravity.
In the present paper, we start with a quick overview of BF theory and the definition of the physical quantum states as spin network states for the flat connection with possibly particle-like topological defects. Then focusing on SU(2) BF theory, we explicitly compute the entanglement between the two parts of such a flat spin network states and we show its holographic behavior: it scales with the size of the boundary (more precisely, with the number of boundary vertices). We also study the influence of topological defects. We show that they do only affect the entropy when located on the boundary between the two regions and we compute the finite variation of entanglement that they create.

I. AN OVERVIEW OF BF THEORY
A. SU(2) BF theory: spin networks and physical states BF theory is a class of topological gauge field theories defined on a oriented smooth n-dimensional manifold M by the following action (see e.g. [1] for a review): (1) The gauge group is a (semi-simple compact) Lie group G whose Lie algebra g is equipped with an invariant (nondegenerate) bilinear form tr(·, ·). Picking a local trivialization of the principal G-bundle over M, the basic fields are a g-valued connection 1-form A with curvature 2-form F [A] and a g-valued (n-2)-form B. The action is invariant under the action of the gauge group, for h ∈ G: It is also invariant under shifts of the B field by an arbitrary (n-3)-form φ: This symmetry kills all local degree of freedom making BF theory a topological field theory. Its classical field equation impose a flat connection, F [A] = 0, and a vanishing "torsion", d A B = 0. In particular, in n = 3 space-time dimensions, BF theory for the gauge groups G = SU (2) and G = SU(1, 1) is equivalent to 3d Riemannian and Lorentzian gravity (in its first order formulation). For our present purpose, we are interested into BF theory from the point of view of its canonical quantizationà la loop quantum gravity. In the following, we will work with n = 3 and G = SU (2) although all the formalism and techniques apply to arbitrary space-time dimensions and arbitrary (compact semi-simple) Lie groups. We perform a 2+1 splitting of the 3d space-time viewing the 3d manifold M ∼ Σ × R as a two-dimensional space Σ evolving in time. Then the spatial parts of the connection A and the field B on Σ are canonically conjugate variables. The theory is completely constrained and the Hamiltonian vanishes on-shell. There are two sets of constraints, imposed by the time components of A and B, respectively, which turn out to be simply Lagrange multipliers. The first constraints impose the flatness of the connection on Σ. The second set of constraints imposes the vanishing of the torsion on Σ and is usually called the Gauss law. These are first class constraints, respectively generating the translational symmetry (3) and the gauge invariance (2) under the action of the group G. For more details on the canonical analysis and resulting structures, the interested reader can check [2].
The loop quantization scheme is based on a specific choice of wave functions. We choose cylindrical functionals of the connection A on Σ. More precisely, they depend on A through only a finite number of variables: they are functions of the holonomies of A along the edges of some arbitrary (finite) oriented graph in Σ. Considering a particular graph Γ with E and V vertices, we consider the holonomies g 1 [A], .., g E [A] of the connection along the edges e = 1, .., E of Γ and build wave functions of the following type: Further we require these functionals to be gauge-invariant. Since the action of the group on the connection translates into a group action at the end points of the holonomies, g e [A] → h −1 s(e) g e [A]h t(e) , where s(e) and t(e) are respectively the source and target vertices of the edge e, the gauge invariance of the wave functions ψ Γ involves a invariance under the group action at every vertex of the graph Γ: Then we need to impose the flatness condition F = 0 on these wave functions. To keep the discussion as simple as possible, we consider a graph Γ which forms a lattice faithfully representing the canonical surface Σ. More precisely, if we cut the surface Σ along the embedded graph Γ, then we are left with surfaces all homomorphic to the unit disk. These are the faces of the lattice. We now impose that the holonomy of the connection around each face is the identity. Thus the projection onto physical states satisfying the flatness condition is implemented by the multiplication by δ-function around each face: where L labels the loops around the faces of the lattice. This imposes trivial holonomies around every contractible loop in Σ while allowing for arbitrary holonomies around the non-contractible cycles of the surface. When Σ is the two-sphere, this gives a single flat quantum state. When Σ is an orientable compact surface of genus n, the physical space of flat quantum states is isomorphic to the space of L 2 functions of 2n group elements For instance, for the two-torus, physical states will be gauge invariant functions of two group elements A and B satisfying ABA −1 B −1 = ½.
Finally, in 2+1 spacetime dimensions particles create conical singularities, which leave the spacetime flat except along their worldline. They are represented as topological defects, which translate into non-trivial holonomies. For a spinless particle on a given face of the lattice, we replace the δ-function by a δ θ -function imposing that the holonomy around the face has a class angle θ. This deficit angle θ then defines the mass of the particle (e.g. see [7,8] for more details).
Our goal in the present work is to compute the entanglement on a physical state between a region of Σ and the rest of the surface and to check whether it satisfies to an "area-entropy" law. On a given graph Γ, we will consider a connected region of the graph and compute the von Neumann entropy of the reduced density matrix obtained after tracing out all the holonomies outside that region from a given physical state. In the next part, we will show how this works for a BF theory based on a discrete group, in which case the theory can be reformulated as a spin system.
To conclude this review section, we give a few mathematical details on the SU(2) group. We parameterize group elements g as where θ ∈ [0, 2π] is the class angle (or half of the rotation angle),û is a unit vector on the 2-sphere indicating the axis of the rotation and σ are the standard Pauli matrices. Let us point out g(θ,û) = g(−θ, −û). In these variables the normalized Haar measure dg on SU(2) reads as By the Peter-Weyl theorem, every function which is invariant under conjugation can be decomposed on the characters of the irreducible (spin) representations of SU (2). Such representations are labeled by a half-integer j ∈ N/2 and the corresponding characters are where U n is the n-th Tchebychev polynomial of the second kind. The δ-distribution decomposes as: Finally, we introduce the distributions δ θ that localize group elements on a specific equivalence class under conjugation by fixing their rotation angle: Its decomposition into characters reads: B. A word on Kitaev's spin system and Z2 BF theory In this subsection, we give a quick overview of Kitaev's spin system introduced in [5] and the corresponding ground state entanglement calculations presented in [6]. As in the previous section, we consider the (canonical) surface Σ provided with the lattice defined by the graph Γ. We attach a two-level system (which is called a qubit in the language of quantum information and represents a spin-1/2 particle) to each edge e of the graph. Label its basis states as |± e . Consider the following Hamiltonian which is a sum of local operators attached to the vertices and the faces of the lattice: with σ x |± = |∓ and σ z |± = ±|± . Calling the operators A v ≡ ⊗ e∋v σ (e) x and B f ≡ ⊗ e∈f σ (e) z , we first notice that all these operators commute with each other. We can thus diagonalize them simultaneously. Ground states |ψ 0 are then states which diagonalize all A v and B f with the highest eigenvalue: As shown in [5] (and reviewed in [6]), this spin system is equivalent to BF theory based on the discrete gauge group Z 2 . Generalizing this to systems with a higher number of levels attached to each edge allows to reformulate in a similar fashion BF theory for an arbitrary discrete group. Ground states correspond to physical states in BF theory. The holonomy can either be + or − along an edge. Then the A v |ψ 0 = |ψ 0 condition implements the Gauss law imposing gauge invariance at each vertex, while the face condition B f |ψ 0 = |ψ 0 imposes the flatness of the Z 2 holonomy. Thus the Hilbert space of ground states for an orientable surface of genus g has a 2 2g degeneracy, with the holonomy around all 2g cycles of the surfaces being free.
Finally, using the stabilizer space methods, the entropy of one (bounded connected) region of the lattice was computed in any ground state [6]. It was shown that it does not depend on the particular ground state (i.e noncontractible cycles do not matter) and it is simply related to the perimeter of the region's boundary as where n L is the number of spins in the perimeter of a region (for a 2d square lattice).
In the present work, we generalize these entanglement calculation to the case of the continuous Lie group, making explicit calculations in the case of the group G = SU(2). We do not use the same methods as developed in the analysis of the spin systems but exploit simple loop quantum gravity tools.

II. ENTANGLEMENT: GENERIC STRUCTURES AND ENTROPY-BOUNDARY LAW
A. The setting: flat spin networks and choice of independent loops Let us consider a fixed connected oriented (abstract) graph Γ and a spin network wave functional ψ(g e ) living on it. We would like to study the completely flat wave functional, imposing that the holonomy along all loops of the graph. This corresponds to the flat physical state for BF theory for a trivial topology (a two-sphere) of the canonical surface Σ. It also gives the physical state for a non-trivial topology but with the additional constraint of trivial holonomies around every cycle of the canonical surface. To truly consider non-trivial topologies, we would need to alow for non-trivial holonomies around some of the loops of the graph (which would then represent the non-contractible cycles of the canonical surface).
In order to study the entanglement properties of this completely flat state, we write the flat wave functional ψ 0 (g e ) as a product of δ-functions on the group SU(2) for every loop of the graph. However, such a naive product would obviously lead to infinities due to redundant δ-functions and we need to consider the product over independent loops. We formalize this using by the gauge fixing procedure for spin networks that was developed in [11].
Considering a generic spin network functional ψ(g e ) on Γ, we can gauge fix the SU(2) gauge transformations (4) acting at the vertices by introducing a maximal tree T on the graph. T is a connected set of edges on Γ touching all vertices of Γ but never forming any loop. If we call E and V respectively the number of edges and vertices of Γ, then the number of edges in T is exactly V − 1. On the other hand, the number of edges not in T is exactly the number of independent loops on the graph L ≡ E − V + 1. The gauge fixing consists in fixing all the group elements g e living on edges e ∈ T belonging to the tree to the identity. More precisely, we choose a reference vertex v 0 . Then for all vertices v, there exists a unique path [v 0 → v] linking v 0 to v along the tree T . We perform a gauge transformation with parameters: For all edges on the tree e ∈ T , the resulting group elements h s(e) g e h −1 t(e) are set to the identity ½. For all edges not belonging to the tree e / ∈ T , this defines a loop variable G e ≡ h s(e) g e h −1 t(e) , which is the oriented product of the loop starting at v 0 and going to s(e) along the tree T , and then coming back to v 0 along the tree from the the vertex t(e). The procedure ensures that (see [11] for more details) the wave functions evaluated on the original g e 's is equal to its evaluation on the G e / ∈T while setting the other group elements to ½.
Finally, we define the completely flat spin network state as: It is straightforward to check that the resulting state actually does not depend on the choice of neither the reference vertex v 0 nor the maximal tree T . Moreover it truly imposes the condition that the holonomy around any loop of the graph Γ is constrained to be equal to the identity ½.
Such a distributional state is obviously not normalisable for the kinematic inner product defined with the Haar measure dG e . It is indeed L 1 but not L 2 . To deal with it, we need to regularize it. The method we will use is the standard one when working in loop quantum gravity and spin foam models [2]: expanding the state in SU (2) representations, we will introduce by hand a cut-off J in the representations and then study the behavior of the various quantities in the large spin limit J → +∞.
Our purpose is to consider a bounded region A of the graph Γ and compute the entanglement between A and the rest of the graph on the completely flat spin network state. This will give the entropy of A. More precisely, we choose a connected region A of Γ. We define it as a set of V int vertices and the E int edges that link them (i.e any edge whose both source and target vertices are in A also belongs to A). We further distinguish the E b boundary edges, that connect one vertex inside A to one vertex outside, and the E ext exterior edges who do not touch the considered region A. We call the interior graph Γ int the graph formed by the vertices and (interior) edges of A. The exterior graph Γ ext is defined as its complement: it consists in both exterior and boundary edges. We define the V b boundary vertices which belong to both interior and exterior graphs i.e vertices in A that touch some boundary edges. Then the exterior graph has To define the entanglement between the interior and the exterior regions, we consider the reduced density matrix on A obtained by tracing out all the holonomies outside A from the full density matrix: We are then interested in the standard measure of entanglement defined as E = − tr A ρ int log ρ int . We would like to compute it on the completely flat spin network state, which a physical state for BF theory. For this purpose, we need to adapt the choice of the tree T used in defining the flat state ψ 0 to the choice of the studied region A: we would like the definition of ψ 0 to respect the interior/boundary/exterior structure. We now choose two reference vertices v 0 in the interior (there is no problem if v 0 itself is a boundary vertex) and w 0 in the exterior (for the sake of simplicity, we choose w 0 so that it is not a boundary vertex) and two maximal trees T int , T ext respectively on the interior and outer graphs. We would like to form a maximal tree T on the whole graph by merging the two trees. The only issue is that the interior and outer graphs, and thus the two trees, both share the V b boundary vertices. Considering a boundary vertex v b , there exists a unique path along the tree T int from v 0 to v b and there also exists a unique path in the exterior along T ext from w 0 to v b . If we consider the straightforward gluing of the two trees T int ∪ T ext , then we obtain loops as soon as there are at least two boundary vertices of the In order to get a tree, we simply need to break these loops. For this purpose, we single out an arbitrary boundary vertex v and it is straightforward to check that T is a maximal tree on the whole graph Γ. Indeed there are no loops in T . Moreover there exists a path in T from v 0 to any vertex v ∈ Γ: if v is in the region A the path is simply the path with the first halves of the path in T int and the second in T ext .
We follow the previous gauge fixing procedure and define the flat state ψ 0 according to this chosen maximal T . We have one holonomy loop variable per edge e / ∈ T not belonging to the tree T . There are three types of such edges. First, we identify the edges e in the region A but not belonging to the interior tree T int . Second, we identify the edges e outside A but not belonging to the outer tree T ext . Finally, there are the boundary edges e (i) , i = 1, .., (V b − 1). For all these edges e / ∈ T , we associate the corresponding holonomies around the interior, exterior and boundary loops respectively which consist in edges f going from v 0 to s(e) along T then along the edge e then coming back from t(e) to v 0 along T . We call the loops L e ≡ [v 0 → s(e)] ∪ e ∪ [t(e) → v 0 ] and define the corresponding holonomies: The flat state is then the product of δ-functions over all these loops: We insist that the whole procedure with the choice of a tree T is simply to choose a set of independent loops in order to impose the flatness conditions with no redundant δ-distribution. In other words, the state ψ 0 as defined above still imposes that the holonomy around any loop of the whole graph Γ is trivial. Finally, a last detail is that we can cut the loops L e for exterior edges e ∈ Γ ext \ T ext . We can have them start at the exterior vertex w 0 instead of the reference vertex On the other hand, the loops L e have the advantage that there only involve edges in the outer graph Γ ext (i.e not belonging to the region A). We will therefore use this prescription for the following We would like to underline that the properties (entanglement,..) of the flat state ψ 0 do not depend on the specific choice of tree T that we made. Indeed the choice of tree is simply a choice to gauge fixing for the spin network functional. The particular tree that we defined using trees in the interior and exterior regions is a convenient choice allowing a clear description of the "boundary loops" which are the central objects for the entanglement calculation. We now compute the reduced density matrix ρ int obtained by tracing out all exterior holonomies from the flat state ψ 0 . The von Neumann entropy of this reduced density matrix defines the entanglement between the region A and the rest of the spin network. Our main result is that the entanglement scales with the size of the boundary: where (V b −1) is the number of boundary loops between the region A and the exterior and u(J) is a unit of entanglement which only depends on the regulator J (representation cut-off) and does not depend on either the graph Γ or the choice of the region A. In particular, if there is a single boundary vertex, V b = 1, the region A is totally disentangled from the rest of the spin network. Let us compute the reduced density matrix as defined by (14): First, the interior loops are unaffected by the integration over exterior edges since they do not involve any edges outside the region A. Second, the exterior loops involve only exterior edges, thus we have the identification H e = H e . This produces δ(½) infinities which we re-absorb in the normalisation of the reduced density matrix. Finally, the moot point is what happens to the boundary loops. More precisely, we have defined: With this decomposition the identification δ( g e g −1 e ) for all edges e ∈ Γ ext implies that the factors δ(B i )δ( B i ) can be re-written as δ(C −1 where the holonomy C −1 i C 0 is the (oriented) product of the group elements from the boundary vertex v along the interior tree T int . Therefore, the reduced density matrix reads up to a normalisation (later fixed by the requirement that tr A ρ int = 1): The next step is to compute the von Neumann entropy of this density matrix, S A = − tr A ρ int log ρ int . We have two types of terms: some δ(G e )δ( G e ) for edges not in the tree e / ∈ T int and some δ(g i g −1 i ) which only involve edges in the tree T int . Since these different terms do not involve the same group elements, they can be treated separately (more precisely, we can do a change of variable in the integrations replacing the g e 's by the G e 's for the edges not in the tree T int and the Jacobian of the transformation is trivial due to the left and right invariance of the Haar measure). The terms δ(G e )δ( G e ) corresponding to the interior loops are density matrices corresponding to pure states and thus have a zero entropy. The only non-vanishing contribution therefore comes from the boundary loops and gives: where S[δ(g g −1 )] is the entropy of the to-be-normalized density matrix σ(g, g) = δ(g g −1 ). It turns out that this density matrix σ is actually the identity matrix and its entropy is simply the log of the dimension of the Hilbert space of L 2 functions on SU(2). However, this Hilbert space has an infinite dimension and the result requires a regularization. The regularization consists in introducing a cut-off J in the representations of SU(2). We define a regularized δ J function in term of the SU(2) characters 1 : Corresponding we consider the Hilbert space of L 2 functions on SU(2) which decompose onto matrix elements of the group elements involving only representations j ≤ J. This Hilbert space H J is spanned by the (renormalized) 1 We could also use the q-deformation of SU(2) which provides a natural way to restrict to a finite number of representations [15]. For instance, given the deformation parameter q = exp(iπ/J), the highest representation is j = J − 1/2. Their q-dimensions are equal to d (q) j ≡ sin(2j + 1)π/J sin π/J = χ j (π/J), which is the same as the evaluation of the usual SU(2) character on an angle θ = π/J. When J → ∞ (and q → 1), we recover d (q) j → (2j + 1). Hence the q-deformed version of the reduced density matrix σ(g, e g) is j χ j (ge g −1 ).
Its entropy (after properly normalization of the density matrix) is where the factor N (J, π/J) is introduced later when dealing with topological defects.
(Wigner) matrix elements √ 2j + 1 D j mn (g), where m and n run by integer step from −j to +j. Its dimension is thus: It is easy to check that the non-normalized density matrix δ J (g g −1 ) is the identity on H J by explicitly computing its matrix elements 2 and therefore its entropy is S(δ J ) = log ∆ J . Indeed, if we introduce the normalized kets through g|j, m, n = 2j + 1D (j) (g) m n , j, m, n|j ′ , m ′ , n ′ = δ jj ′ δ mm ′ δ nn ′ , then the normalized density matrix is obviously maximally mixed, hence is maximally entangled with the rest of the system [16]. This finally proves the results which we announced above (a more detailed proof is presented in Appendix A): Result 1. The entanglement between the region A and the rest of the spin network state for the completely flat state ψ 0 is: At the end of the day, it is the number of boundary vertices that is relevant for the entanglement. It could seem surprising since the SU(2) group elements live on the edges of the graph and therefore the degrees of freedom of the theory apparently live on the edges. We would then expect the entanglement to scale with the number boundary edges. However, the requirement of gauge invariance implies that the physical degrees of freedom actually live on the (independent) loops of the graph and not simply on its edges. Then we proved that the number of (independent) boundary loops is directly related to the number of boundary vertices (minus one). A last comment is that we would still have found the same leading order behavior for the entanglement in case we had first regularize and then calculate the reduced state instead of the contrary as we did above 3 .

C. Topological defects and renormalized entanglement
We can further generalize the procedure described above to take into account topological defects. In the context of 3d gravity, topological defects represent point particles. For instance, a (spinless) particle of given mass leads to a non-trivial holonomy around it: we impose δ θ (g) instead of δ(g) on loops around it with the angle θ related to the mass (see [7,8] for more details).
Moreover, we expect that the leading order of the entanglement does not change when including topological defects. Nevertheless, for any state ψ we introduce the renormalized (or relative) entanglement as a regularized entropy [14] of the reduced density operator of A, 2 We compute, using the orthogonality of SU(2) matrix elements: √ 2k + 1D k cd |δ J | p 2j + 1D j ab = Z dgde g p (2k + 1)(2j + 1)D j ab (g)D k cd (e g)δ J (ge g −1 ) = δ jk δacδ bd .
3 It can be understood by noting that Z dgδ(g)δ(g) = δ(½) = X which is the difference of the regularized measure of entanglement of the region A with the rest computed in the spin network state ψ and the entanglement computed for the completely flat state ψ 0 . When ψ only includes topological defects, we actually expect the renormalized entanglement to converge to a finite value as the regulator J is taken to infinity. Let us first consider inserting a topological defect in an interior loop or an exterior loop 4 , i.e replacing δ(G e ) or δ(H e ) by δ θ (G e ) or δ θ (H e ). In the case of an exterior loop, we will get in the reduced density matrix after integration over the group elements living on the exterior edges a term dHδ θ (H) 2 instead of dHδ(H) 2 . Such a term only enters the normalization of the reduced density matrix and thus does not affect the entanglement. In the case of an interior loop, the modified loop constraint gives a term δ θ (G e )δ θ ( G e ) in the reduced density matrix instead of the original δ(G e )δ( G e ). This still represents a pure state, thus has zero entropy. Once again, it does not affect the entanglement. This proves the following result: Result 2. Starting off from the completely flat state ψ 0 and then adding topological defects along some loops of the spin network state, if no topological defect is inserted on boundary loops then the entanglement between the region A and the rest of the graph does not change: On the other hand, if we insert a topological defect on the boundary between A and the outside, or more precisely replace the constraint δ(B i ) by δ θ (B i ) on one given boundary loop, this will modify our previous entanglement calculation. For the completely flat state ψ 0 , the term in ρ int corresponding to that loop was dgd g δ(g g −1 )δ(g i g)δ( g i g) = δ(g i g −1 i ) which gives the totally mixed state (identity density matrix on the Hilbert space of L 2 functions over SU (2)). This now replaced by: where we used the decomposition of the distribution δ θ into SU(2) representations as given by eqn. (10). Cutting off the sum over representations to J, we compute the matrix elements of this reduced density matrix: where the normalisation factor N (J, θ) ensures that the reduced density matrix σ (θ) J has a unit trace: One can check that doing a Taylor expansion around θ ∼ 0, we recover N (J, θ → 0) = ∆ J . Finally, computing the von Neuman entropy of this density matrix gives the renormalized entanglement between the region A and the outside in the state ψ θ with one topological insertion along a boundary loop: Result 3. Inserting a single topological defect δ θ along a boundary loop between the region A and the outside leads to a non-zero renormalized entanglement: We prove in Appendix B that this expression has a finite limit for all values of θ when the regulator J is sent to ∞. This renormalized entanglement has a universal value when θ is not a rational fraction of π: We can also compute this value analytically 5 for some specific values of the angle θ: We see that the renormalized entropy (and entanglement) is discontinuous. This is actually a generic feature of infinite-dimensional systems [9,13]. It is also known that if a reasonable constraint is put on the accessible set of states, such as bounded mean energy, tr Hρ < ∞, then the entropy will be a continuous function on this set. In our case the entropy is continuous if the cut-off J is large but finite. This can be motivated by appealing to the geometric interpretation of the representation labels. Indeed, in the context of 2+1 (loop) gravity, the spin network states are the eigenstates of the length operator and its eigenvalues are exactly the representation labels (up to ordering ambiguities). Thus requiring the finite extent of the system in space would naturally impose an upper cut-off on the edge representation labels. The negativity of the above result can be easily understood if we recall the the flat holonomy state is maximally entangled. Therefor, the states with non-trivial holonomies are less than maximally entangled, hence the negativity of the renormalized entropy.
This result extends to the more general case of several topological defects inserted along different boundary loops, each boundary loop contributing independently to the (renormalized) entanglement between the region A and the rest of the spin network state.
Finally, we are left with the possibility of a non-trivial (2d) topology, i.e non-contractible cycles in the canonical surface. It is easy to see that the entanglement calculations are not affected by non-contractible cycles as long as the region A has a trivial topology (isomorphic to the unit disk). Otherwise, if the region A contains some non-contractible cycles, the holonomies around them will couple to the holonomies around the cycles outside A. As an example, taking the case of a 2-torus, let us consider that one cycle with holonomy G is contained inside A while the dual cycle with holonomy H is outside. Now, flatness of the connection does not require δ(G)δ(H) but the weaker condition that they commute GHG −1 H −1 . Then the reduced density matrix ρ int for the region A will be given in term of integrals of the type: for some central function f (invariant under conjugation). The representation decomposition of such density matrix involves the {6j} symbol and are more complicated to analyze. We postpone this study to future investigation.

III. SOME SIMPLE EXAMPLES OF ENTANGLEMENT
A. The Θ-graph The simplest spin network that allows a non-trivial entropy calculation is the Θ-graph. It is too simple to be decomposed into two non-trivial inside/outside regions, but it allows to illustrate how to compute the entanglement between two sets of holonomies. The totally flat state is (the choice of maximal tree is a single edge T = {e 3 }): where we chose the two independent loops [e 1 , e 3 ] and [e 2 , e 3 ]. Notice that the δ-functions also impose that the loop [e 1 , e 2 ] carries a trivial holonomy, g 1 g −1 2 = ½. Constructing the (formal) density operator ((up to regularisation/renormalisation) for the edge e 3 one obtains The expectation values are calculated according to the trace formula O = tr(ρO)/ tr ρ. Hence we regularize the density operator by dropping the infinite multiplicative constant and truncating the remaining δ-functions in its expansion in term of SU(2) representations: where the sum on j 3 is over all half-integers and the D j mn (g) are the matrix elements of the (Wigner) matrix representing the group element g in the representation of spin j (with m and n running by integer step from −j to +j). The normalization constant is determined by the trace condition tr ρ = 1, so Introducing the following ket notation (as in [10]), g|j, m, n = 2j + 1 D j mn (g), the state induced on the edge e 3 can be written as For a pure state |Ψ the von Neumann entropy of ρ J (3) is the measure of entanglement between the edge e 3 and the rest (see e.g [10]). Hence The result is obviously divergent when J → ∞. We can check that computing the entanglement between the edge e 1 and the rest (edges e 2 , e 3 ) gives the same result. Moreover, as a consistency check, we calculate the complementary density matrix ρ(g 1 , g 1 ; g 2 , g 2 ), The loop G 12 = g 2 g −1 1 is the analog of an interior loops while g 1 g −1 3 plays the role of the boundary holonomy. Finally, as the reduced state ρ (12) is the direct product of the pure interior state δ(G 12 )δ( G 12 ) and of the mixed state δ(g 1 g −1 1 ), its entropy is determined only by the latter and we see that S(ρ (12) We can further introduce a topological defect along one of the loops of the Θ-graph. Let us consider the state which satisfies the following constraints, for all j ∈ N/2, Expanding the δ θ -distribution in SU(2) characters, we compute the reduced density matrix for the edge e 1 : After truncation of the sum over representations and proper normalisation, this gives in the ket notation: For θ = 0, we recover the previous reduced density matrix computed for the flat state. This leads to a regularized entanglement as we obtain in the generic case: On the the hand, we can also compute the reduced density operator ρ (2) θ for the edge e 2 : Up to a normalisation, it is actually equal to the density matrix ρ (2) computed above for the flat state, ρ θ (g 2 , g 2 ) ∝ δ(g 2 g −1 2 ) = ρ (2) (g 2 , g 2 ). Similarly, we obtain that ρ A simple graph that is represented on Fig. 4a has an equal number of all three types of edges, E int = E b = E ext = 2 and allows a simple inside/outside distinction. The holonomies along the internal, boundary and external edges are denoted by g, q, and h, respectively. It has two internal and two external vertices, so the three independent loops are produced by one internal, one external, and one boundary loop. The loops are marked on Fig.4a. The flat wave functional is given by (taking into account the orientation of the edges) Computing the reduced density matrix for the two interior edges gives: where G = g 1 g 2 is the holonomy along the interior loop. This interior density matrix decomposes into a direct product of a pure state on the interior loop and the mixed state δ( g 1 g −1 1 ). According to the previous calculations, the regularized entropy is simply As expected, it results in the same reduced density matrix for the interior, ρ int Ψ = ρ int Ψ0 , thus also S(ρ int Ψ ) = S(ρ int Ψ0 ). Coming back to the original graph in Fig.4a, we impose a non-trivial holonomy around the boundary loop: The interior density operator is and it leads to a renormalized entropy equals toĒ(θ) as we computed in the previous section (see result.3).

IV. CONCLUSIONS
We computed the entropy for a bounded region on a physical state of BF theory. Indeed, being solvable and lacking local degrees of freedom, BF theory allows for explicit calculations and a precise analysis of the relationship between boundaries and degrees of freedom. Looking at physical spin network states, we showed that the entanglement between the two regions for an arbitrary bipartite splitting of the spin network only depends on the structure of the boundary: the entropy simply scales with the size of the boundary. More precisely, we proved that the entanglement grows with the number of boundary vertices, E ∝ (V b − 1). We also showed that the introduction of topological defects does not affect this result at leading order and we computed exactly the finite entropy difference due to such particle-like defects on the boundary.
Technically, we developed the necessary mathematical tools required to analyze the graph structure of a flat distributional spin network state and to regularize the resulting wave functional and entropy calculations. These tools are also relevant to entanglement calculations for the spin systems used for topological quantum computation such as the Kitaev model [5,6].
We hope to apply these techniques to study the precise relation between gauge breaking and entropy on one hand, and to compute the entanglement for a more general class of spin network states relevant for loop quantum gravity [17].

APPENDIX A: DETAILED PROOF OF THE RESULTS
In this Appendix we discuss the structure of independent loops of Γ and prove our Results 1-3 in a more general setting. The connected graphs Γ int and Γ out (that consists of of the outer vertices, V out = V − V int and E ext exterior edges) contribute E int − V int + 1 and E ext − V out + 1 independent interior and exterior loops, respectively. It is convenient to introduce a further classification of the remaining E b − 1 loops. A true boundary loop in addition to the boundary (and, possibly) exterior edges includes at least one interior edge. Loops that contain at least one boundary edge, but no interior ones, we (for the lack of a better term) name frontier loops. Exterior loops that were defined above, together with the frontier loops form all the independent loops of Γ ext (as defined in Sec. IIA it consists of V ext = V out + V b exterior and boundary vertices, and E ext + E b exterior and boundary edges).
From the definition of frontier loops it follows that their number in the set of independent loops is given by where the total number of the exterior loops is Finally, the number of boundary loops in the set of independent loops of Γ is Fig. 1 illustrates the loop construction that is used to define the new variables. First we select L int independent interior loops. Next, L b = V b − 1 boundary loops are constructed as follows. We select V b different paths (v 0 v i w 0 ) that link the two representative vertices, which are distinguished by the boundary vertices v 1 , v 2 , . . . , v V b they are passing through. The numbering of boundary vertices induces a natural numbering of the boundary loops that are formed by combining the sequential paths (v 0 v i w 0 ) and (v 0 v i+1 w 0 ). Finally, the set of L independent loops is completed by a necessary number of the external loops.
The i-th boundary loop carries a holonomy B i ≡ g i Q i g −1 i+1 , where g i is the (internal) holonomy acquired on the way from v 0 to v i , and the outer holonomy Q i is acquired along the path (v i w 0 v i+1 ). The holonomies Q i are the first L b new exterior variables. Holonomies H i around i = 1, . . . , L ext exterior loops (calculated with respect to w 0 ) provide another L ext independent variables. Since we complete the set of outer variables by adjoining to the list the holonomy Q V b which is acquired by going from v V b to w 0 along the edge v 0 v V b w 0 , and picking additional variables r.
To express an arbitrary holonomy R as a product of L independent holonomies and their inverses, all of them should be calculated with respect to some fixed reference point, such as v 0 . Hence, we pick up an arbitrary edge (say, v V b ), so with respect to v 0 the outer flower holonomies become Holonomies G i , i = 1, . . . , L int , around the independent interior loops are part of the set of new internal variables.
When V b = V int , the representative interior point is on the boundary, and can be taken to be v V b . Then g V b ≡ ½, and the set of interior variables is completed by the holonomies g i , i = 1, V b − 1. Indeed, they resulting set is independent and On the other hand, when V int > V b , we add a non-trivial g V b to the set of interior variables, and possibly more additional variables g.
In these new variables the flat state ψ 0 becomes where the last product contains the delta-functions of additional holonomies. The index i can cover all possible remaining loops of Γ. Those holonomies (calculated with respect to v 0 ) can be expressed as Pi·{r,s,t} where number and ordering of the factors, as well as their powers π a = 0, ±1 are determined by the decomposition of R i . The variableR i is defined by the same equation with g →g, G →G. As a result, the interior density operator of Result 1 is The integration results in ρ int (g, g, G;g,g,G) = Lint r=1 δ(G r )δ(G r ))δ(½) Lout where the conditions g s Q s g −1 s+1 = ½ and the factors δ(g −1 s+1 g sgsgs+1 ) and δ(G r ), were used to transform R i to the identity. Dropping the infinite constants and making the final change of variables g ′ s = g −1 s+1 g s brings the interior density operator to the form ρ int (g, g, G;g,g,G) = σ int (G,G) which completes the proof of Result 1. Result 2 deals with the situation where L b independent boundary loops that carry trivial holonomies can be chosen. In this case the state is given by for some assignment {θ} to the external and internal independent loops, and any compatible assignment to the dependent loops. After the integration over Q it reduces to where R ′ i = Pi·{r,t} G has the same form as in Eq. (A9). 2 Finally, in the setting of Result 3 we pick the first boundary loop to have a non-trivial holonomy. Then where θ i is either 0,or θ, depending on the loop. Integration over all variables but Q 1 results in the reduced state being given as where n is a number of "other" loops that carry holonomies of the class θ. Since δ 2 θ (g) = δ θ (g)c θ , where the infinite constant can be expressed in a regularized fashion using the formulas of Ec. I.A, the reduced state finally becomes ρ θ int (g, g, G;g,g,G) = σ int (G,G)δ θ (g ′ 1g ′−1 1 ) which trivially exhibits the desired entropy.

APPENDIX B: COMPUTING THE RENORMALIZED ENTANGLEMENT
Here we prove that the renormalized entanglementĒ(θ), as introduced in the Sec. II.C, converges for any value of the angle θ. Because of the symmetry θ ↔ π − θ, we can restrict ourselves to 0 < θ ≤ π/2. Since it follows that the limit exists if asymptotically for some function f (θ). We have to consider two different cases, depending on whether the class angle is a rational or irrational fraction of π. In the first case the class angle is θ = p/qπ, with relatively prime p, q ∈ N, and there are only finitely many different values that sin 2 nθ takes. It is periodic with period K θ = q, and q l=0 sin 2 (lpπ/q) = q/2. (B3) To simplify the notation we further take J = kq, k ∈ N, and consider the limit k → ∞.
which establishes the claim for θ = p/qπ, p, q ∈ N. For a generic value of θ we establish the limit by using the Euler-Maclauren integration formula for sums. The asymptotic behavior of the entropy is Using the known integral π 0 dx sin 2 x log sin 2 x = π( 1 2 − log 2), the second term becomes 1 J 2J 0 dn sin 2 (nθ) log sin 2 (nθ) ∼ 1 − 2 log 2.