On skein relations in class S theories

Loop operators of a class S theory arise from networks on the corresponding Riemann surface, and their operator product expansions are given in terms of the skein relations, that we describe in detail in the case of class S theories of type A. As two applications, we explicitly determine networks corresponding to dyonic loops of $N=4$ $SU(3)$ super Yang-Mills, and compute the superconformal index of a nontrivial network operator of the $T_3$ theory.


Introduction and summary
To understand the properties of 4d N =2 supersymmetric theories, it is extremely useful to realize them as twisted compactifications of the 6d N =(2, 0) superconformal field theories (SCFTs) on Riemann surfaces C with punctures [1,2]. These 4d theories are now called class S theories, and they are closely associated to certain 2d theories defined on C. For example, class S theories on S 4 are related to the Liouville/Toda CFT on C [3,4] and when considered instead on S 3 ×S 1 they give rise to 2d q-deformed Yang-Mills on C [5][6][7].
How do supersymmetric defects 1 appear under these 4d-2d duality relations? In this paper, we only consider one-dimensional objects, loops, on the 4d side. Since 6d N =(2, 0) SCFTs have codimension-two defects and codimension-four defects, a loop operator on the 4d side should come from a codimension-four defect on a codimension-one defect on the 2d side.
For class S theories of type SU(2), or equivalently A 1 , dyonic charges of loops on the 4d side is in one-to-one correspondence with topologies of non-intersecting loops on the 2d side [8]. When the 2d theory is the Liouville/Toda theory, these loops on the 2d side have an explicit realization in terms of the Verlinde operators [9] as shown in [10,11]. When the 2d theory is the q-deformed Yang-Mills, these loops are instead realized as 2d Wilson loops. 2 We only have to consider non-intersecting loops on the Riemann surface for the A 1 case, thanks to the existence of a skein relation resolving each crossing into a sum of two non-crossing ones : where q = e πib 2 in the Liouville theory and q = q 1/2 in the q-deformed Yang-Mills. However, in the higher rank cases, namely for SU(N) with N > 2 or equivalently for A k with k = N − 1 > 1, there is no such simple skein relation, since junctions of lines inevitably appear. This results in the networks 3 of lines on the Riemann surface, as already mentioned in [12,13]. Such networks were treated and discussed in [14,15], putting special emphasis on the SU(3) case. In [14] the analysis was mainly carried out using the approach of the higher Teichmüller theory classically (q = 1), and in [15] the study was done mainly in the framework of Toda CFTs for general q. These analyses gave rise to the skein relations that had been discovered in the context of mathematics before [16].
Our aim in this paper is to describe the skein relations of the networks in the general SU(N) case. We have, for example, the relation that reduces to the equation (1.1) when N = 2. These skein relations were first found in [17] in the context of knot invariants. The guiding principle for us is that the representation theory of the quantum group SU q (N) underlies these networks and their skein relations. The relation of the loop operators of 2d CFTs and the quantum group has been known for quite some time, mainly in the context when q is a root of unity, see e.g. [18]. In the case of 2d q-deformed Yang-Mills, the relation of their loops and the quantum group is very direct, because the q-deformed Yang-Mills is a gauge theory whose gauge group is the quantum group [19,20].
We have two applications of the skein relations. The first concerns the dyonic loop operators of N =4 SU(N) Yang-Mills theory. From the 4d gauge theory perspective, their classification was performed in [21]. When the electric charge and the magnetic charge of a dyonic loop are parallel in the weight system, there is an obvious realization of such dyonic loop in the class S language as a loop wrapping the torus. When they are not parallel, it was expected that they are represented by networks on the torus. We will give a complete description in the case of SU (3).
The second application is the study of a particular loop operator of the T 3 theory. The T N theory is a strongly-coupled class S theory that does not have a useful Lagrangian description, and there is no direct method to describe a loop operator using the path integral language. Still, using the class S language, it is easy to see that there is a loop operator described by a nontrivial network on a three-punctured sphere. We compute the superconformal index of the T 3 theory in the presence of this loop operator, and confirm that it indeed has an enhanced E 6 symmetry, as expected from the fact [1] that the T 3 theory is in fact the E 6 theory of Minahan and Nemeschansky [22].
The rest of the paper is organized as follows: in Sec. 2, we start by reviewing the relation of the operator product expansion of the loop operators on the 4d side and the skein relation on the 2d side. We emphasize the universality of the skein relation, independent of the choice of the 4d spacetime. Next, in Sec. 3, we describe the skein relations explicitly for general SU(N). We also explain how they can be understood in terms of the representation theory of quantum groups. In Sec. 4, we turn to an application on 4d N =4 gauge theory. We discuss the operator product expansion of a pure Wilson loop and a pure 't Hooft loop in the general SU(N) theory. We also describe all the dyonic loops in the SU(3) theory in terms of networks, and point out a relation to the brane tiling. In Sec. 5, we first explain how the 4d Wilson loop is mapped to the 2d Wilson loop in the context of the q-deformed Yang-Mills. We then use one of the skein relations to compute the superconformal index of the T 3 theory in the presence of a loop operator corresponding to the simplest nontrivial network. In the Appendix A, we have more examples of the realization of dyonic loops by networks in the general SU(N) theory.

Loop operators in 4d and skein relations in 2d
In the Liouville theory, a Verlinde loop operator is defined in terms of monodromy actions on the conformal block F along a loop γ. It is possible to insert more than one Verlinde operator and L γ A L γ B F = L γ B L γ A F in general when γ A and γ B intersect each other, as we see from the concrete calculations. See Fig. 1 for an illustration. Under the 2d-4d correspondence, the Verlinde operators map to loop operators of the 4d theory. Therefore, there should be a concept of ordering of loops on the 4d side, such that the product becomes non-commutative. In this section we review how this ordering arises, following [23][24][25][26]. See also the recent reviews [27,28].

Sums and products of loops
We consider the 4d setups where some kind of localization computations is possible. Typically there is a supercharge preserved in the background, whose square involves a linear combination of two isometries k 1,2 . Supersymmetric loops wrap along the direction of k 1 and sit at the fixed point of k 2 .
In the neighborhood of the loop, we can approximate the geometry as S 1 × C × R, where k 1 shifts the coordinate along S 1 and k 2 is the phase rotation of C. The loop now wraps S 1 and sits at the origin of C, and the position x ∈ R is arbitrary. Therefore, we can place multiple loops L 1,2,... on x 1,2,... preserving the same supercharge.
This gives an intrinsic ordering of loops on the 4d side, and furthermore, the expectation values are unchanged under infinitesimal changes of the positions x i . The choice of the local supersymmetric background at a loop can be characterized by a single parameter which we denote by q. These statements can be explicitly checked in the case of the localizations on S 1 × S 3 [12], S 1 × R 3 [26] and S 4 b [29]. Given two types of loop defects L 1 and L 2 , we denote loops placed at x ∈ R by L i (x). We define a formal sum L 1 + L 2 of two loops by · · · (L 1 + L 2 )(x) := · · · L 1 (x) + · · · L 2 (x) (2.1) where the ellipses stand for other operator insertions. The product L 1 · L 2 of two loops are now defined by where we demand x 1 > x > x 2 so that L 1 and L 2 are the loops closest to x from the left and from the right. Since the expectation values depend only on the order but independent of the relative distance, this gives a consistent definition. At this stage, we can make sense of the operator product expansion for defects. Suppose that there is a set L 1,2,... of loops which cannot be decomposed into any sum of other simpler ones. We then have the following expansion of correlation functions that can be written succinctly as The OPE coefficients c k ij (q) are asymmetric under the exchange i ↔ j due to the intrinsic ordering along R. However, we can simultaneously flip the R direction and S 1 direction in the local S 1 × C × R geometry to obtain another supersymmetric background. The background is parameterized by q as remarked before, and we use the parameterization such that this flip is represented by q → q −1 . Then we should have the relation networks in 2d operators on H loops in 4d There is also an order for the defect networks on the two dimensional geometry side. In order to make the relation between two orders, let us consider the case when the 4d side is S 4 b and the 2d side is the Liouville/Toda theory [3,4]. In this case the Verlinde operators associated to the networks on the 2d side act on the space H conf of conformal blocks, and the loop operators on the 4d side act on the space H hemi of holomorphic Nekrasov partition functions defined on a hemisphere [10,11,15,30]. Since H conf and H hemi are naturally isomorphic, we can see the relations among three orderings as shown in Fig. 2.
Corresponding to (2.4), there should be the operator product expansions of defect networks which are indeed skein relations as resolutions of the crossings. In Sec. 4, we will see that we can calculate many OPE coefficients in terms of defect networks.

Non-commutativity and the angular momentum
Let us recall the origin of the non-commutativity when the geometry is globally S 1 ×C×R, following the discussions in [24,26]. Considering the S 1 as the time direction, the partition function of S 1 × C × R is given by where H is the Hilbert space of the system, J 3 = J 3 + I 3 is the sum of the spin along R and the Cartan of the SU(2) R-symmetry. Our parameter q is then given by q = e πiλ . Suppose now that we have a U(1) gauge theory, that the first loop L 1 is purely electrically charged with electric charge e and that the second loop L 2 is purely magnetically charged with magnetic charge m. Then there appears the Poynting vector carrying the angular momentum J P = ( /2)em along the R direction. The magnitude of J P is independent of the distance between two particles but the sign depends on the ordering, and therefore we have In the classical limit (q → 1), this product becomes commutative.

Networks and skein relations in 2d
In this section we discuss possible types of networks in class S theories of type SU(N) and their skein relations. Our guiding principle is that they are described by the structure of the quantum groups underlying both the q-deformed Yang-Mills and the Liouville/Toda CFT, and that the skein relations are universal under an appropriate parameter identification. The skein relations exhibited below already appear in the mathematical works [17,31,32], up to the overall factors and the changes in conventions. Skein relations introduce equivalence relations among all possible networks, and it would be extremely useful if we can pick a natural representative element out of a given equivalence class of networks allowing linear combinations of networks. In this paper we at least give a general method to simplify a given network: • In principle, edges in a network can carry arbitrary representations of SU(N). We will first rewrite them in terms of representations ∧ k for k = 1, . . . , N − 1. These are the fundamental representations in the mathematical terminology, 4 and within the diagrams we just denote them by 1, 2, . . . , N − 1.
• Then we rewrite all crossings in terms of linear combinations of junctions that are at most rectangular. The concrete formulas are given in (3.36). We call this process crossing resolutions.
In the A 1 case, these procedures eliminate all the crossings and no junctions remain, thus reproducing the classification in [8]. In the A 2 case, we will see that all digons and rectangles can be eliminated, and we will find a natural representative for a given equivalence class of networks. We will detail this process in Sec. 4.4.
Hereafter we use a version of the standard quantum number defined as and the factorial defined as As this section is rather long, let us pause here to explain the organization: in Sec. 3.1, we start by recalling that codimension-4 operators of the 6d N =(2, 0) theory are labeled by representations and they can have junctions corresponding to the invariant tensors. In Sec. 3.2, we describe how an arbitrary representation can be rewritten in terms of just the fundamental representations of the form ∧ k . In Sec. 3.3, we then describe the trivalent junctions where three edges labeled by fundamental representations meet. In Sec. 3.4, we show how a crossing of two edges can be rewritten in terms of junctions. We start from the crossing of two edges labeled by and then describe the general case. In Sec. 3.5 we summarize the Reidemeister moves that are fundamental equivalence relations guaranteeing the isotopy invariance. In Sec. 3.6 we note other useful skein relations that can be used to simplify networks. Finally in Sec. 3.7, we explicitly display the skein relations for A 2 and A 3 .
In general, it would also be important in the class S theory to study skein relations with full and other punctures in [1] or networks ending on other punctures. We do not, however, consider such objects in this paper.

Generalities
Before proceeding, let us first recall the fact that a codimension-4 operator of the 6d N =(2, 0) theory of type SU(N) has a label given by a representation of SU(N), and how a multiple number of such operators can be joined.
A cylinder of the N =(2, 0) theory gives rise to a 4d N =2 vector multiplet with gauge group SU(N). On the 4d side, then, we can consider the Wilson loop operator in a representation R of SU(N). This should come from some codimension-4 operator of the N =(2, 0) theory wrapped around the cylinder. Then this codimension-4 operator also needs to be labeled by a representation R.
When we multiply two parallel Wilson loops with representations R 1 and R 2 , we get a Wilson loop with representation R 1 ⊗ R 2 , and the product is commutative. The same should be then true among codimension-4 operators of the N =(2, 0) theory.
On the 4d side, three Wilson loops in representations R 1,2,3 can be joined at a point consistently if R 1 ⊗ R 2 ⊗ R 3 contains an SU(N) invariant subspace, or equivalently when there is an invariant tensor in this triple tensor product. The number of independent ways to join them is given by the number of linearly independent invariant tensors. Then, three codimension-4 operators of the N =(2, 0) theory labeled by R 1,2,3 can be joined along a one-dimensional subspace when there are invariant tensors in R 1 ⊗R 2 ⊗R 3 . The number of distinct ways to connect is given by the number of linearly independent invariant tensors.
Since this should be an intrinsic property of codimension-4 operators of the 6d theory, we can join three codimension-4 operators along a one-dimensional loop on the 4d side. This gives a junction of three edges labeled by R 1 , R 2 , R 3 on the 2d side. Using many such junctions, we end up having networks on the 2d side, which are our main interest in this paper.

Restriction of labels
First, note that Wilson loop on the 4d side in a representation R in a direction can be thought of as a Wilson loop in the representationR in the opposite direction. This feature should also be carried over to the codimension-4 defects of the N =(2, 0) theory, and to the networks on the 2d side. This can be represented diagrammatically as As it is cumbersome to use arbitrary representations R as labels, we next rewrite them in terms of fundamental representations ∧ k , k = 1, . . . , N − 1. An irreducible representation R can be specified by a Young diagram (ℓ i ) where ℓ i is the number of boxes in the i-th row, so that ℓ i = N. For example, ∧ k is represented by (1 1 Note that any symmetric polynomial of N variables x 1 , x 2 , . . . , x N , under a constraint x 1 x 2 . . . x N = 1, can be written as a polynomial of the elementary symmetric polynomials. As a character χ R (diag(x 1 , . . . , x N )) of SU(N) in a representation R is such a symmetric polynomial, and χ (1 k ) (diag(x 1 , . . . , x N )) for k = 0, 1, 2, . . . , N − 1 are exactly elementary symmetric polynomials, this means that any representation R can be decomposed as the direct sum (allowing negative integral coefficients) of the tensor products of ∧ k .
For example, we have the equalities

4)
which we can diagrammatically depict, in the case of closed loops, as These relations are locally applicable on parallel edges. Therefore, we can insert some punctures or networks inside the circle, for example.
We can just use the integer k to label the edge, and an edge labeled by 0 can be removed.
Reversing the orientation now corresponds to replacing the label k by N − k. An edge labeled by k has charge k under the center of SU(N), and therefore we call these integer labels as the charge. 5 For each trivalent junction, the sum of three inflowing charges must equal to zero modulo N. Say we have three edges labeled by a, b and c = a + b. There is only a single invariant tensor in ∧ a ⊗ ∧ b ⊗ ∧ N −c , and this corresponds to the projection from ∧ a ⊗ ∧ b to ∧ c=a+b . Therefore, there is no need to place a label on a junction to distinguish the possible invariant tensors.
Sometimes these labels k are then taken to be defined modulo N as in [14,15], but it is useful to consider them just as integers between 0 and N − 1. This is because we can write down the invariant tensor rather explicitly using the quantum group representation theory when the net inflowing charge to a junction vanishes in Z. We call such a junction canonical. We call a junction non-canonical if the net inflowing charge vanishes only in Z N . See Fig. 3 for examples.
Let us now describe the invariant tensors associated to the canonical junctions. Let be spanned by the vectors e 1 , . . . , e N . We define the q-deformed wedge product by the rule where in particular e i ∧ e i = 0 where we do not sum over i. This defines the projection π from ⊗ to ∧ 2 by Furthermore, ∧ 2 can be naturally embedded within ⊗ by the rule More generally, we associate to any canonical junction that combines labels a, b to a + b the projection 3.11) and to any canonical junction that splits the label a + b to two labels a, b the map ι a+b→a,b where where we assume k 1 < · · · < k a+b , the sum is over the disjoint split of indices and n(i, j; k) is the minimal number of adjacent transpositions to bring the sequence i 1 , . . . , i a , j 1 , . . . , j b to k 1 , . . . , k a+b . These maps are described in more detail mathematically in [33]. Then these maps π a,b→a+b and ι a+b→a,b naturally combine according to the following diagrams together with the ones with reversed arrows:

Removal of digons
Now, we can check that any digons can be removed as Note that in the classical limit q → 1 the prefactor becomes due to the fact that the classical limit of ι a+b→a,b is (−1) ab times the standard map that follows from the classical epsilon symbol. This somewhat unusual sign is however necessary to match with the known skein relations in the Liouville/Toda theory, and it also simplifies the signs appearing in the general crossing resolutions (3.36). In the q-deformed Yang-Mills theory it would be more conventional to drop this sign and the canonical junctions would be defined to be π a,b→a+b and (−1) ab ι a+b→a,b . We will stick to the Liouville/Toda convention in this paper, except in Sec. 5.
When the sum of i l is N, we can use the rule to evaluate a network with two trivalent junctions, since edges labeled by N can be removed. For example, when i + j + k = N, we have More simply, we can evaluate a closed loop with label k by considering it as a digon with edges labeled by k and N − k: Again, this shows that our convention is different by a factor of (−1) k(N −k) from the convention in the q-deformed Yang-Mills. We also see at this point that, to compare with the skein relation of the Toda theory or the q-deformed Yang-Mills theory, we need to use the relation q = e πib 2 = q

Crossing resolutions 3.4.1 The R matrix
Let us first discuss the best-known case: the R-matrix for ⊗ of SU(N), which is given by Here, I V is the identity operator on a vector space V , Q is an operator where e ij is a matrix whose only non-zero entry is 1 at the i-th row and j-th column and A is the overall normalization which we will be fixed later. . (3.23) The resulting entries are the basis vectors of the second rank antisymmetric representation of SU q (N). Indeed, the operator Q is the composition of the projection π 1,1→2 : ⊗ → ∧ 2 q and the natural embedding ι 2→,1,1 : ∧ 2 q → ⊗ : Note also that Q satisfies This is a special case of the digon elimination.
We can now represent the R-matrix R diagrammatically as The inverse of the R-matrix R is that we represent as Below, we call the crossing (3.26) as positive and the crossing (3.28) as negative.
The A 1 -case: In this case, ∧ 2 is the trivial one-dimensional representation and there is the pseudo-reality condition ≃ which we can diagrammatically write as Then the general equation (3.26) reduces to where we have set A = q 1/2 . We then have These reproduce the standard skein relations of the Liouville theory found in [10,11] under the identification q = e iπb 2 . The relation Q 2 = 2 Q shows From this we see that q = q 1/2 SCI where q SCI is the parameter used commonly in the literature on the superconformal index. The minus sign here is a convention common in the Liouville/Toda literature, i.e. the definition of a loop in the representation differs by an overall minus sign between the Liouville theory and the q-deformed Yang-Mills.
The A 2 -case: Here we have = ∧ 2 , and therefore we have As we now only have one type of the label , we can drop it altogether. The general R-matrix (3.26) then becomes where we have set A = q 1/3 . This reproduces the fundamental skein relations of the SU (3) Toda theory found in [15], again under the identification q = e iπb 2 .
General case: The analysis so far suggests that we should take in general. As we will see soon, this is consistent with the general crossing resolutions (3.36).

General crossing resolutions
Now let us move on to the crossing resolutions in the general case. The expression was found in [17] up to an overall factor, which we quote here: where s = min(a, b, N − a, N − b) and a, b = 0, 1, 2, . . . , N − 1. With this choice of the overall factor, this equality is invariant with a reversal of an arrow and the rotation of the diagrams by 90 • . Note also that when a = b = 1, this equality reduces to (3.26) we already discussed. Let us introduce the names to the fundamental objects on the right hand side of (3.36): (3.37) Note that the number of possible choices of i matches with the number of irreducible summands of the decomposition of the tensor product ∧ a ⊗ ∧ b . We expect that all these Q ab (i) cannot be further decomposed as parts of networks.

The intersection number and the powers of q
Let us briefly discuss the significance of the prefactor q ab N in (3.36). In general, two loop operators in a class S theory of type SU(N) can be mutually nonlocal, and the nonlocality can be measured in terms of the Dirac pairing that takes values in Z N [34]. In terms of the 2d networks realizing the 4d loop operators, the Dirac pairing is given by their intersection number. We can define it by assigning a local intersection number to a crossing as follows: This is consistent with the reversal of arrows, since it sends the label a to N − a. The intersection number I(Γ 1 , Γ 2 ) of two networks Γ 1 and Γ 2 is then defined by summing the contributions from all the crossings: where sign(c) is the sign of the crossing c and a (i) c is the charge of Γ i at c. In the Liouville/Toda setup, we expect the expectation value of any network without crossings is a single-valued function of q = e πib 2 invariant under q → e 2πi q. Similarly, in the superconformal index, the expectation value of a loop operator on the 4d side is a single valued function of q = q 1/2 SCI , since the index is a trace Tr(−1) F q ∆−I 3 and the scaling dimensions ∆ of a class S theory are integral or half-integral.
Non-invariance of the expectation value under q → e 2πi q then captures the mutual non-locality, and we can think of the prefactor q ab N in (3.36) as encoding the difference in the local intersection number between the left hand side and the right hand side, to keep track of this non-locality. The difference in the powers of q among different summands in the resolutions of the crossings should be integral, and the relation (3.36) indeed satisfies this requirement.

SL(2, Z) action on the torus
We can define the operation I which we call the inversion by reversing all the arrows simultaneously. This is an involution that we can identify with the charge conjugation on the 4d side.
When the network is on a torus, we can also consider the action of SL(2, Z) on the networks. Two basic actions are the T -action and the S-action. T corresponds to sending α → α, β → β + α and S corresponds to α → β, β → −α, where α, β are two bases of the 1-cycles on the torus.
The operation C = S 2 generates the center of SL(2, Z) and is the charge conjugation action in the 4d N =4 SYM theory. Then we need to have two operations, C and I, to be consistent on the torus. The relation (3.37) relating two forms of the networks representing the same object Q ab (i) is exactly the one required to have C = I, when these networks are put on the torus and open edges are connected to the opposite ones.

Reidemeister moves
In knot theory, a projective representation in two dimension of knots and links in a three dimensional space is not unique, and any different representations can be mapped to each other by a combination of three so-called Reidemeister moves, see e.g. [35]. The move I straightens a twist in an edge, the move II slides one edge over another edge to two parallel edges, and the move III changes the order of three crossings. In the presence of junctions, we need to add another move, where we move an edge over a junction. We call this as the move IV.
Since we expect that the charge of a loop in the 4d theory is determined by the isotropy class of networks, we would like to require that a network is invariant under these moves. This is indeed possible for the moves II, III and IV, but the move I results in a q-dependent factor. In the context of 3d Chern-Simons theory, this can be understood from the change in the framing of the link [36]. Let us describe these moves explicitly below.

R-Move II:
This relation says that the the negative crossing is given by the inverse R −1 of the R-matrix R corresponding to the positive crossing. (3.40) R-Move III This is the Yang-Baxter equation which the R-matrix R should satisfy. (3.41) R-Move IV This is the additional move for the networks with junctions. (3.42) R-Move I Finally, this move involves a nontrivial factor.
When the representation V is ∧ k , the coefficient C k (q) can be calculated using the general crossing resolution (3.36) and the relation (3.16) removing digons. The i-th network on the right hand side of (3.36) gives a coefficient (3.16). We then find (3.44) A direct understanding of this coefficient in 4d or 6d would be an interesting problem.

More simplifying relations
Let us list various other skein relations that can be used to simplify networks. All relations except (3.48) are known in [32] [31] and references therein.

Triangle contraction relations
We have rules to remove triangles. In order to express the rules, we first map all the junctions so that they are canonical. Then, there are four possibilities up to the mirror images: c . (3.46) Note that whether three vertices are totally ordered by arrows or not changes the look of the factors.

Rectangle decaying relations
The rectangles Q (i) ab that we had in (3.37) can not be further simplified, but there are many other rectangles that are equivalent to sums of simpler ones. Let us show one class: These relations assure that any network constructed from only rectangles around a tube can always be decomposed into a sum of closed loops around the tube. For example, using (3.47) recursively, we can see where the two horizontal thin parallel lines signify that they are to be identified so that the network is on a tube. There are various other relations. Here we just note one example: (3.49)

Examples: A 2 and A 3
The skein relations of the A 1 case and the A 2 case have already been described in the literature.

A 2
Let us record the A 2 case as a summary. We have two types of junctions : (3.50) The basic skein relation was (3.34), which we copy here [14][15][16]: The following two relations are useful to simplify the networks: (3.52)

A 3
Let us now discuss the next nontrivial case of A 3 . Note that the label 3 can be traded with 1 by reversing the arrow, and since 2 is a real representation we do not have to exhibit the direction for edges labeled by 2. In this case, there are also two types of junctions as we see below.
There are three types of crossing resolutions : (3.55) There are three decaying relations for one rectangle:

Networks for N =Yang-Mills
As the fundamental aspects of defect networks in the two dimensional theories have been discussed, we can move on to the discussion of the correspondence between networks on the 2d side and loops on the 4d side. In this section, we restrict ourselves to the most familiar N =4 case. On the gauge theory side, the charges of the loop operators were classified in [21]. It is not easy to construct the corresponding networks for the general A k case, but we will see that the skein relations allow us to describe and classify the networks for A 2 concretely. Before proceeding, let us quickly recall the possible charges of the loop operators of N =4 SU(N) Yang-Mills, following [21]. We denote the weight lattice by Λ. We use the notations ω i for the fundamental weights and h i for the weight vectors in the defining N-dimensional representation. They are explicitly given by Note that ω 1 = h 1 and ω N −1 = −h N . Let us consider a Wilson loop labeled by an irreducible representation R. We can also use its highest weight λ as the label, and possible highest weights are in one-to-one correspondence with Λ/W where W is the Weyl group. Similarly, a 't Hooft loop can be characterized by a charge vector in Λ, considered up to the action of the Weyl group.
For a dyonic loop operator, we need to specify a pair of electric and magnetic charges (µ, λ) ∈ Λ × Λ but the charges need to be identified under a simultaneous action of the Weyl group. Therefore a dyonic charge corresponds to an element in (Λ × Λ)/W and represent the element as [(µ, λ)]. We also call it [(µ ′ , λ ′ )] lower than [(µ, λ)] if µ ′ and λ ′ are lower than µ and λ respectively when mapped to Λ/W.

The product of Wilson loops and 't Hooft loops
It is well known how pure Wilson loops and pure 't Hooft loops are represented as loops on the torus : where we identify each pair of parallel opposite edges to make the parallelogram the torus.
Here W R is the Wilson loop in the representation R, and we identify an irreducible representation and its highest weight vector. We use a similar notation for the 't Hooft loop. We also fix the horizontal one cycle as α-cycle and the vertical one as β-cycle. Note that the S transformation on the torus is naturally identified with S duality transformation of N =4 gauge theory.
We can now decompose the product W ωa · T ω b using the crossing resolution (3.36). Here we express it in a form to make the data of electromagnetic charges manifest: Recalling the fact reviewed in Sec. 2.2 and that µ is the x-component of the classical angular momentum associated to the Poynting vector under electric charge λ  where we use the symbol D [(µ,ν)] to denote the dyonic loop operator with the charge [(µ, ν)] ∈ (Λ × Λ)/W. We also use a simple symbol . Below, we call these dyonic loops D b,a (i) and the corresponding networks elementary.

Analysis in the Liouville/Toda theory
Let us now connect our analysis so far with a computation on the Liouville/Toda theory side, using the localization in the gauge theory [30,37].
Hereafter we use symbols a = (a 1 , a 2 , . . . , a N ) under the constraint A 2 (a), . . . , A N (a)). We also denote the N =4 holomorphic partition function by Z(a).
As seen in [37] and [30], Wilson loop W R and 't Hooft loop T R are written in the form of matrix model : (4.8) where χ R is the character of R and Π(R) is the set of weights corresponding to the irreducible representation R. T (R) λ (a) are some functions of a related to the character χ R via a Fourier transformation in a [12,30] but the concrete expressions are unnecessary hereafter.
In general, any loop operator is expected to be represented as where ν runs over some finite set in the weight lattice Λ and X ν (a) are some functions the detail of which we do not need either. The additions of W R and T R in the ordering of loops seen in Sec. 2.1 are written as follows : In particular, let us choose R as one of the fundamental representations ∧ n and introduce W (k) := W ∧ k and T (ℓ) := T ∧ ℓ . Then consider insertions both of W (k) and of T (ℓ) . One way to insert is where the ellipsis represents further insertions of other loops. Recalling exp[2πib a, λ ], then define the following operators labeled by m = 1, 2, . . . , min(k, ℓ) : (4.14) where we decompose the set Π Note that the decomposition of T (ℓ) W (k) is independent of the ellipsis . . . and X assuring that this expansion is local and represent the product as T (ℓ) · W (k) or T (ℓ) × W (k) . On the other hand, the insertion in the opposite order is and we also have In summary, we have found the relations Comparing the product expansion (4.20) and the graphical expansion (4.4) we find the following identification : and the monodromy action under the change θ → θ + 2π is following : The Witten effect on the partition function can be re-expressed in the loop operators which acts on the partition functions. In particular, when λ is in Π(∧ ℓ ) = W(ω ℓ ), Z(a) * Z(a − bλ) is accompanied by q −ℓ+ ℓ 2 N e 2πib λ,a (4.23) as θ shifts by 2π. Summing it up over Π(∧ ℓ ), This θ → θ + 2π action is graphically represented as and matches with the T transformation on the torus.

Examples of products of loops in A 2
Let us focus on the A 2 case and perform some explicit computations. The examples in the general A k case will be given in Appendix A. We will see the geometric SL(2, Z) action on the torus is nicely mapped to the SL(2, Z) action on the electric and magnetic weight systems.
. This was originally found in [14,15] in the context of class S theory.
The dyonic loop D 1,1 (1) is obtained from the 't Hooft loop T by an application of the T operation. In particular this loop can be mapped to a Wilson loop in some duality frame. The object D 1,1 (0) cannot be mapped into a network localized on any one cycle by the torus modular transformations. In the language of charges, this means that the electric weight and the magnetic weight are not parallel. We can now work out how the SL(2, Z) transformations act on this particular network and the pair of weights, see Fig.4.
Example 2. The next example is W × T : In this example, the first term on the right hand side is a network that cannot be mapped by SL(2, Z) to any of the networks we already studied explicitly. It is natural to posit the following expansion + (loops with lower weights) (4.29) since we expect that the exponent of q multiplying D [(λm,λe)] equals − λ m , λ e to capture the angular momentum. Then we can identify up to the lower contribution D [(ω 1 ,h i +h j(>i) )] from lower weights. Hereafter, we try to map networks and charges of the dyonic loops up to the contributions from lower weights. 7 Example 3. The third example is W Adj × T : The skein relation gives us while from gauge theory we expect For a graphical representation of weights involved, see Fig. 5.
The first term and the third term can be obtained by SL(2, Z) transformations on D 1,1 (0) . The second term is a new type: while the gauge theory computation yields + (loops with the lower weights) (4.35) where λ Adj = ω 1 + ω 2 is the highest weight of the adjoint representation and see Fig. 6 for λ 1,2± . It would be interesting to reproduce the terms with lower weights from a purely 4d gauge theoretic computations.

Classification of networks on T 2 for A 2
We have seen some basic examples of products of loops and the identification of the charge and the network. Here we establish the general mapping between the networks and the charges in the case of the A 2 theory on the torus, or equivalently the N =4 SU(3) Yang-Mills. This is a minimal extension of the dictionary of Drukker, Morrison and Okuda [8].
Let us first classify the possible A 2 networks on the torus purely in terms of the skein relation. First, recall that all networks with crossings are resolved into those with junctions only. In particular, for the A 2 case, there are only two types of junctions, namely the one where the heads of three arrows meet and another one where the tails of three arrows meet. Therefore the networks are bipartite [14] and there appear only polygons with degree-even vertices.
We now use the skein relations we discussed so far. Recall the basic conventions we discussed in Sec. 3.7. All digons can be contracted, and all rectangles are resolved to two pairs of curves, as we discussed in (3.52).
At this point, the network might contain several disconnected components. If there are no vertices at all, the network consists of parallel loops wrapping the same one-cycle on T 2 . Assume now there is at least one vertex. Pick a connected component. It has the topology of either a disk, an annulus or a torus. Now, let us denote the number of edges, or equivalently the number of vertices, of the i-th polygon in this connected component by p i (= 2, 4, 6, . . .). Denote the number of polygons by f . The total number of vertices, edges and faces of the network is then given by since the connected component is either a disk (χ = 1, B > 0), an annulus (χ = 0, B > 0), or a torus (χ = 0, B = 0). Since we removed all digons and rectangles, p i ≥ 6, and therefore we have From this we see that the connected component has the topology of the torus, and every polygon is a hexagon. Therefore, the possible A 2 networks on T 2 are mapped into the bipartite hexagon tilings with three corner condition at every vertex. It is interesting to note at this point that bipartite hexagon tilings of the torus appeared in the string theory literature in the context of brane tilings [41][42][43]. In this case the bipartite hexagon tilings corresponded to Abelian orbifolds of C 3 . Now let us make the dictionary between the bipartite hexagon tilings and the dyonic charges. Instead of thinking of filling a torus by hexagons, we can take the quotient of the bipartite hexagon tiling filling the entire plane, and then we define the vectors e 1 and e 2 there, see Fig.7. To specify a bipartite hexagon tiling, we choose the α and the β cycles of the torus from Ze 1 ⊕ Ze 2 so that they are linearly independent. In Fig. 8 we show the hexagonal tilings and the dyonic charges that already appeared in our analysis so far. Figure 8: Some examples of the hexagonal tilings and the dyonic charges that we have already identified From these examples, we can find the general map. We first naturally identify the A 2 weight lattice Λ and the dual lattice of the hexagonal tiling. Then, the rule is (λ e , λ m ) → (α, β) = (λ m , −λ e ). (4.39) under the condition q 1 p 2 − q 2 p 1 > 0 where λ e = q 1 ω 1 + q 2 ω 2 and λ m = p 1 ω 1 + p 2 ω 2 . It is clear that the action of the Weyl group is consistent. Because the cycles α, β define the basis of charges, the action of SL(2, Z) on the dyonic charges (λ e , λ m ) and that on the cycles (α, β) should be transpose of each other, and indeed the mapping (4.39) satisfies this condition. Let us end this section by exhibiting some more examples of the mapping between the dyonic charges and the hexagonal tilings, see Fig. 9. In this section we only discussed the A 2 case. It would be interesting to find a general map from the charge of the dyonic loops to the networks for or general A k>2 cases.

Some examples in the superconformal index
So far we have been studying the properties of the networks of the class S theories. In this section, we would like to study a few concrete manifestations of our analysis when the four-dimensional side is S 1 × S 3 , or equivalently, in the setting of the superconformal index.
As has been well established, the superconformal index of a class S theory of type SU(N), defined by a Riemann surface C of genus g with n full punctures, is given by the (zero-area limit of the) (p, q, t)-deformed Yang-Mills theory on the same Riemann surface [5][6][7]. Here we consider the case q = t where the p dependence automatically drops out, and the superconformal index becomes the standard q-deformed Yang-Mills theory, defined in [19,20,44]. In this case, the supercharge defining the superconformal index is compatible with the presence of a half-BPS line wrapping the S 1 direction [25,45].
Therefore, in this setup, a defect of a class S theory, specified by a network with labels on the Riemann surface C should be realized as a concrete object in the q-deformed Yang-Mills theory, defined by the same network with labels on C. Below, we show that they are just given by the Wilson lines 8 and the Wilson junctions in the q-deformed Yang-Mills theory. In particular, we compute the superconformal index of the nontrivial line operator of the T 3 theory, namely the 6d N =(2, 0) theory of type SU(3) compactified on a sphere with three full punctures. Reassuringly, we will find that the result shows an enhancement of the symmetry from SU(3) 3 to E 6 , as it should be, since the T 3 theory is the E 6 theory of Minahan and Nemeshcansky.
Below, we will first very briefly review the relation between the superconformal index and the q-deformed Yang-Mills in Sec. 5.1. Then we check that the superconformal index with a Wilson line in the 4d gauge theory is given by the Wilson line in the 2d gauge theory in Sec. 5.2. Finally, we will compute the superconformal index of a network in the T 3 theory in Sec. 5.3. Note that in this section we use the 2d Yang-Mills normalization of the junctions as discussed in Sec. 3.3.2.

The superconformal indices without lines
Consider the class S theory S g,n for a genus g surface with n full punctures. Its superconformal index is defined as I g (a 1 , . . . , a n ) = Tr(−1) F q ∆−I 3 a 1 · · · a n (5.1) where the trace is over the Hilbert space of the system on S 3 or equivalently on the space of the operators, ∆ is the scaling dimension, I 3 is the Cartan generator of SU(2) R symmetry, and a i ∈ SU(N) i is a group element of the flavor symmetry SU(N) i of the theory associated to the i-th full puncture. This is known to be given by the following explicit expression: where the sum is over all the irreducible representations λ of SU(N), χ λ (a) is the character of a in the representation λ, q ρ is the element and K(a), K 0 are given by Here we took a to be diagonal, a = diag(a 1 , . . . , a N ) ∈ SU(N), (5.5) and PE is the plethystic exponential, defined by PE n≥1 a n t n t = n≥1 (1 − t n ) −an . (5.6) Take two class S theories S g,n and S g ′ ,n ′ , pick one full puncture from each, say the last one from S g,n and the first one from S g ′ ,n ′ . Let us then couple SU(N) gauge multiplets to the diagonal combination of the SU(N) symmetries carried by the two full punctures thus chosen. The combined system has the superconformal index I g,g ′ (a 1 , . . . , a n−1 ; b 2 , . . . , b n ′ ) = [dz] Haar K(z) −2 I g (a 1 , . . . , a n−1 , z)I g ′ (z −1 , b 1 , . . . , b n ′ ) (5 .7) where [dz] Haar is the natural measure on the Cartan of SU(N) given by Plugging in the expression (5.2) to (5.7) and using the orthogonality of characters we see that the equation (5.7) reduces to (5.2) for the genus g + g ′ surface and n + n ′ − 2 full punctures. This is as it should be. When g = 0 and n = 3, what we have is the T N theory. For N = 3, the symmetry SU(N) 3 enhances to E 6 . The decomposition is where α = (a, b, c) is the fugacity for E 6 flavor symmetry, and χ λ = χ λ (α) is the character of E 6 in the irreducible representation with Dynkin label λ.

4d Wilson Lines
Let us move on to the superconformal index in the presence of a loop operator wrapping S 1 [45,46]. In this subsection, for simplicity, we only study a loop operator that is just a Wilson line with respect to a 4d gauge group. For class S theories, this covers every operator that is a genuine loop L labeled by a representation R on the Riemann surface (i.e. all networks without junctions), since we can always cut the Riemann surface along L to go to a duality frame where that particular loop wraps a tube once.
To compute the superconformal index with a Wilson loop, let us first consider a more general situation. Take a theory X with G flavor symmetry whose superconformal index is I X (a). Suppose we can couple it with the G vector multiplet, such that the gauge coupling is exactly marginal. Let us insert a BPS Wilson line in the representation R of G. The resulting index is simply I R (q) = [dz] Haar K(z) −2 χ R (z)I X (z). This reduces to the formula (5.7) when R is a trivial representation, as it should be. Now consider the case when X consists of two copies of the T N theory, such that we gauge a diagonal combination of G symmetries. We have where n λ ′ λR = [da] Haar χ R (a)χ λ (a)χ λ ′ (a −1 ) (5.14) counts how many times the irreducible representation λ ′ appear in the tensor product λ ⊗ R. The result (5.13) is, up to the prefactor involving K, the unnormalized correlator of the q-deformed Yang-Mills theory with the Wilson line with the representation R, around the tube associated to the gauge group, see Fig. 10. There, we displayed full punctures as boundaries, as would be more common in the 2d Yang-Mills viewpoint.
For non-deformed 2d Yang-Mills theory, this is an immediate consequence from the fact that the Hilbert space of the system on S 1 is the space of class functions on G, that is spanned by χ λ (z), and the Wilson loop in the representation R acts by a multiplication by χ R (z) almost by definition. For q-deformed 2d Yang-Mills theory, we need to use the fact that the structure of the tensor product decomposition of the representations of the quantum group G q is not deformed as long as q is generic.

A network in the T 3 theory
As an example of the network that is not just a loop, let us consider the T 3 theory and the 2d network shown in Fig. 11. The superconformal index is given by where c R 1 ,R 2 ,R 3 is the amplitude of the q-deformed Yang-Mills, coupling three states χ R i (a i ) (i = 1, 2, 3) on the three boundaries.
In the undeformed 2d Yang-Mills theory, c R 1 ,R 2 ,R 3 is given by the integral × ǫ ijk ǫ¯ijk(U 1 ) ī i (U 2 ) j j (U 3 ) k k , (5.16) where U 1,2,3 are holonomies of the 2d gauge field from one junction point to the other junction point, along three different segments.
In the q-deformed Yang-Mills theory, we need to perform the integral above in the sense of the quantum group [19,20]. The U 3 integral gives a nonzero result only when R 1 is contained in R 2 ⊗ V , where V is the standard three-dimensional representation of SU(3) q . Since the rules of irreducible decompositions of tensor products are unchanged under the q deformation, we see that c R 1 ,R 2 ,R 3 is nonzero only when the highest weight of R 1 is given by adding to the highest weight of R 2 one of the three weights of V . This means that there is an arrow connecting R 1 → R 2 in the diagram of irreducible representations as shown in Fig. 12. We immediately see that there we should have arrows similarly for R 2 → R 3 and R 3 → R 1 . Therefore c R 1 ,R 2 ,R 3 is nonzero only when R 1 → R 2 → R 3 → R 1 forms a triangle in Fig. 12.
Next, consider what happens when the holonomy at a puncture a 3 is set to the special value a 3 = q ρ = (q, 1, q −1 ). This is equivalent to the absence of the puncture, and we have the situation in Fig. 13. Using the skein relation, we see that the left hand side and the right hand side should be proportional by a factor of [2]. The right hand side is just R 1 →R 2 χ R 1 (a 1 )χ R 2 (a 2 ). Therefore, c R 1 ,R 2 ,R 3 is determined by an overconstrained set of equations R 2 →R 3 →R 1 →R 2 c R 1 ,R 2 ,R 3 χ R 3 (q ρ ) = [2] q δ R 1 →R 2 (5.17) ω 2 ω 1 Figure 12: An arrow R → R ′ connects two highest weights R, R ′ of SU (3) if R ′ ⊗V contains R. c R 1 ,R 2 ,R 3 is nonzero only when the arrows form a triangle R 1 → R 2 → R 3 → R 1 . where the sum on the right hand side is over R 3 that fit in the triangle. This equation can be recursively solved starting from the triangle closest to the origin. We can check that the following gives the general solution 9 : dim q (nω 1 + mω 2 ) = χ nω 1 +mω 2 (q ρ ) = . This is essentially the same as the discussion in [8].
A.2 W × T Let us compute the skein relation of W and T . Comparing with what we expect from the gauge theory, we can then identify various networks with complex dyonic loops. From the skein relation, we have The second term is new, and the first and the third ones are obtained by some duality actions of some elementary one. Our identifications are: The first one does not appear when N = 3 but reduces to a sum of simpler ones.
Then the charge/network dictionary for the new ones is: where λ Adj = h 1 − h N is the highest weight of the adjoint representation.