18.769: Lecture 8 Notes

1.37. Quantum traces. Let C be a rigid monoidal category, V be an object in C, and a ∈ Hom(V, V ∗∗ ). Define the left quantum trace (1.37.1) TrLV (a) := evV ∗ ◦ (a ⊗ IdV ∗ ) ◦ coevV ∈ End(1). Similarly, if a ∈ Hom(V, trace (1.37.2) ∗∗ V ) then we can define the right quantum TrR V (a) := ev∗∗ V ◦ (Id∗ V ⊗ a) ◦ coev∗ V ∈ End(1). In a tensor category over k, TrL (a) and TrR (a) can be regarded as elements of k. When no confusion is possible, we will denote TrLV by TrV . The following proposition shows that usual linear algebra formulas hold for the quantum trace. Proposition 1.37.1. If a ∈ Hom(V, V ∗∗ ), b ∈ Hom(W, W ∗∗ ) then (1) T rVL (a) = T rVR∗ (a∗ ); (2) T rVL ⊕W (a ⊕ b) = TrLV (a) + TrLW (b) (in additive categories); (3) T rVL ⊗W (a ⊗ b) = TrLV (a)TrLW (b); R ∗∗ (4) If c ∈ Hom(V, V ) then TrLV (ac) = TrLV (c∗∗ a), TrR V (ac) = TrV ( ca). Similar equalities to (2),(3) also hold for right quantum traces. Exercise 1.37.2. Prove Proposition 1.37.1. If C is a multitensor category, it is useful to generalize Proposi­ tion 1.37.1(2) as follows. Proposition 1.37.3. If a ∈ Hom(V, V ∗∗ ) and W ⊂ V such that a(W ) ⊂ W ∗∗ then TrLV (a) = TrLW (a) + TrLV /W (a). That is, Tr is additive on exact sequences. The same statement holds for right quantum traces. Exercise 1.37.4. Prove Proposition 1.37.3. 1.38. Pivotal categories and dimensions. Definition 1.38.1. Let C be a rigid monoidal category. A pivotal ∼ →?∗∗ . structure on C is an isomorphism of monoidal functors a : Id − ∼ That is, a pivotal structure is a collection of morphisms aX : X − → X ∗∗ natural in X and satisfying aX⊗Y = aX ⊗ aY for all objects X, Y in C. Definition 1.38.2. A rigid monoidal category C equipped with a piv­ otal structure is said to be pivotal. Exercise 1.38.3. (1) If a is a pivotal structure then aV ∗ = (aV )∗ −1 . Hence, aV ∗∗ = a∗∗ V . (2) Let C = Rep(H), where H is a finite dimensional Hopf alge­ bra. Show that pivotal structures on C bijectively correspond to group-like elements of H such that gxg −1 = S 2 (x) for all x ∈ H. Let a be a pivotal structure on a rigid monoidal category C. Definition 1.38.4. The dimension of an object X with respect to a is dima (X) := Tr(aX ) ∈ End(1). Thus, in a tensor category over k, dimensions are elements of k. Also, it follows from Exercise 1.38.3 that dima (V ) = dima (V ∗∗ ). Proposition 1.38.5. If C is a tensor category, then the function X �→ dima (X) is a character of the Grothendieck ring Gr(C). Proof. Proposition 1.37.3 implies that dima is additive on exact se­ quences, which means that it gives rise to a well-defined linear map from Gr(C) to k. The fact that this map is a character follows from the obvious fact that dima (1) = 1 and Proposition 1.37.1(3). � Corollary 1.38.6. Dimensions of objects in a pivotal finite tensor cat­ egory are algebraic integers in k. 12 Proof. This follows from the fact that a character of any ring that is finitely generated as a Z-module takes values in algebraic integers. � 1.39. Spherical categories. Definition 1.39.1. A pivotal structure a on a tensor category C is spherical if dima (V ) = dima (V ∗ ) for any object V in C. A tensor category is spherical if it is equipped with a spherical structure. Since dima is additive on exact sequences, it suffices to require the property dima (V ) = dima (V ∗ ) only for simple objects V . Theorem 1.39.2. Let C be a spherical category and V be an object of −1 C. Then for any x ∈ Hom(V, V ) one has T rVL (aV x) = TrR V (xaV ). −1 ∗ Proof. We first note that TrR X (aX ) = dima (X ) for any object X by Proposition 1.37.1(1) and Exercise 1.38.3(1). Now let us prove the proposition in the special case when V is semisimple. Thus V = ⊕i Yi ⊗ Vi , where Vi are vector spaces and Yi are simple objects. Then x = 12If k has positive characteristic, by an algebraic integer in k we mean an element of a finite subfield of k. ⊕i xi ⊗IdVi with xi ∈ Endk (Yi ) and a = ⊕ IdYi ⊗aVi (by the functoriality of a). Hence � TrLV (ax) = Tr(xi ) dim(Vi ), � −1 TrR Tr(xi ) dim(Vi∗ ). V (xa ) = This implies the result for a semisimple V . Consider now the general case. Then V has the coradical filtration (1.39.1) 0 = V 0 ⊂ V1 ⊂ V2 ⊂ · · · ⊂ V n = V (such that Vi+1 /Vi is a maximal semisimple subobject in V /Vi ). This filtration is preserved by x and by a (i.e., a : Vi → Vi∗∗ ). Since traces are additive on exact sequences by Proposition 1.37.3, this implies that the general case of the required statement follows from the semisimple case. � Exercise 1.39.3. (i) Let Aut⊗ (IdC ) be the group of isomorphism classes of monoidal automorphisms of a monoidal category C. Show that the set of isomorphism classes of pivotal structures on C is a torsor over Aut⊗ (IdC ), and the set of isomorphism classes of spherical structures on C is a torsor over the subgroup Aut⊗ (IdC )2 in Aut⊗ (IdC ) of elements which act by ±1 on simple objects. 1.40. Semisimple multitensor categories. In this section we will more closely consider semisimple multitensor categories which have some important additional properties compared to the general case. 1.41. Isomorphism between V ∗∗ and V . Proposition 1.41.1. Let C be a semisimple multitensor category and let V be an object in C. Then ∗ V ∼ = V ∗ . Hence, V ∼ = V ∗∗ . Proof. We may assume that V is simple. We claim that the unique simple object X such that Hom(1, V ⊗X) �= 0 is V ∗ . Indeed, Hom(1, V ⊗ X) ∼ = Hom(∗ X, V ) which is non-zero if = V , i.e., X ∼ and only if ∗ X ∼ = V ∗ . Similarly, the unique simple object X such that Hom(V ⊗ X, 1) �= 0 is ∗ V . But since C is semisimple, dimk Hom(1, V ⊗ X) = dimk Hom(V ⊗ X, 1), which implies the result. � Remark 1.41.2. As noted in Remark 1.27.2, the result of Proposi­ tion 1.41.1 is false for non-semisimple categories. Remark 1.41.3. Proposition 1.41.1 gives rise to the following ques­ tion. Question 1.41.4. Does any semisimple tensor category admit a piv­ otal structure? A spherical structure? This is the case for all known examples. The general answer is un­ known to us at the moment of writing (even for ground fields of char­ acteristic zero). ∼ Proposition 1.41.5. If C is a semisimple tensor category and a : V − → V ∗∗ for a simple object V then Tr(a) = � 0. Proof. T r(a) is the composition morphism of the diagram 1 → V ⊗ V ∗ → 1 where both morphisms are non-zero. If the composition mor­ phism is zero then there is a non-zero morphism (V ⊗V ∗ )/1 → 1 which means that the [V ⊗ V ∗ : 1] ≥ 2. Since C is semisimple, this implies that dimk Hom(1, V ⊗ V ∗ ) is at least 2. Hence, dimk Hom(V, V ) ≥ 2 which contradicts the simplicity of V . � Remark 1.41.6. The above result is false for non-semisimple cate­ gories. For example, let C = Repk (GLp (Fp )), the representation cat­ egory of the group GLp (Fp ) over a field k of characteristic p. Let V be the p dimensional vector representation of GLp (Fp ) (which is clearly irreducible). Let a : V → V ∗∗ be the identity map. Then Tr(a) = dimk (V ) = p = 0 in k. 1.42. Grothendieck rings of semisimple tensor categories. Definition 1.42.1. (i) A Z+ −basis of an � algebra free as a module over Z is a basis B = {bi } such that bi bj = k ckij bk , ckij ∈ Z+ . (ii) A Z+ −ring is an algebra over Z with unit equipped with a fixed Z+ −basis. Definition 1.42.2. (1) A Z+ −ring A with basis {bi }i∈I is called a based ring if the following conditions hold � [a] There exists a subset I0 ⊂ I such that 1 = i∈I0 bi . [b] Let τ : A → Z be the group homomorphism defined by � 1 if i ∈ I0 (1.42.1) τ (bi ) = 0 if i �∈ I0 There i �→ i∗ of I such that induced map �exists an involution � a = i∈I ai bi �→ a∗ = i∈I ai bi∗ , ai ∈ Z is an anti-involution of ring A and such that � 1 if i = j ∗ (1.42.2) τ (bi bj ) = 0 if i = � j ∗. (2) A unital Z+ -ring is a Z+ -ring A such that 1 belongs to the basis. (3) A multifusion ring is a based ring of finite rank. A fusion ring is a unital based ring of finite rank. Remark 1.42.3. (1) It follows easily from definition that i, j ∈ I0 , i = � j implies that b2i = bi , bi bj = 0, i∗ = i. (2) It is easy to see that for a given Z+ −ring A, being a (unital) based ring is a property, not an additional structure. (3) Note that any Z+ -ring is assumed to have a unit, and is not necessarily a unital Z+ -ring. Proposition 1.42.4. If C is a semisimple multitensor category then Gr(C) is a based ring. If C is semisimple tensor category then Gr(C) is a unital based ring. If C is a (multi)fusion category, then Gr(C) is a (multi)fusion ring. Proof. The Z+ -basis in Gr(C) consists of isomorphism classes of simple objects of C. The set I0 consists of the classes of simple subobjects of 1. The involution ∗ is the duality map (by Proposition 1.41.1 it does not matter whether to use left or right duality). This implies the first two statements. The last statement is clear. � Example 1.42.5. Let C be the category of finite dimensional repre­ sentations of the Lie algebra sl2 (C). Then the simple objects of this category are irreducible representations Vm of dimension m + 1 for m = 0, 1, 2, . . . ; V0 = 1. The Grothendieck ring of C is determined by the well-known Clebsch-Gordon rule (1.42.3) i+j � Vi ⊗ V j = Vl . l=|i−j|,i+j−l∈2Z The duality map on this ring is the identity. The same is true if C = Rep(Uq (sl2 )) when q is not a root of unity, see [K]. Let C be a semisimple multitensor category with simple objects {Xi }i∈I . Let I0 be the subset of I such that 1 = ⊕i∈I0 Xi . Let Hijl := Hom(Xl , Xi ⊗ Xj ) (if Xp ∈ Cij with p ∈ I and i, j ∈ I0 , we p p will identify spaces Hpi and Hip with k using the left and right unit morphisms). � We have Xi ⊗ Xj = l Hijl ⊗ Xl . Hence, �� j = Hi1 i2 ⊗ Hjii43 ⊗ Xi4 (Xi1 ⊗ Xi2 ) ⊗ Xi3 ∼ i4 Xi1 ⊗ (Xi2 ⊗ Xi3 ) ∼ = j �� i4 l Hii14l ⊗ Hil2 i3 ⊗ Xi4 . Thus the associativity constraint reduces to a collection of linear iso­ morphisms � j � (1.42.4) Φii41 i2 i3 : Hi1 i2 ⊗ Hjii43 ∼ = Hii14l ⊗ Hil2 i3 . j l The matrix blocks of these isomorphisms, (1.42.5) (Φii41 i2 i3 )jl : Hij1 i2 ⊗ Hjii43 → Hii14l ⊗ Hil2 i3 are called 6j-symbols because they depend on six indices. Example 1.42.6. Let C be the category of finite dimensional repre­ sentations of the Lie algebra sl2 (C). Then the spaces Hijl are 0- or 1-dimensional. In fact, it is obvious from the Clebsch-Gordan rule that the map (Φii41 i2 i3 )jl is a map between nonzero (i.e., 1-dimensional) spaces if and only if the numbers i1 , i2 , i3 , i4 , j, l are edge lengths of a tetrahe­ dron with faces corresponding to the four H-spaces (i1 i2 j, ji3 i4 ,i1 li4 , i2 i3 l, such that the perimeter of every face is even (this tetrahedron is allowed to be in the Euclidean 3-space, Euclidean plane, or hyper­ bolic 3-space, so the only conditions are the triangle inequalities on the faces). In this case, the 6j-symbol can be regarded as a number, provided we choose a basis vector in every non-zero Hijl . Under an ap­ propriate normalization of basis vectors these numbers are the Racah coefficients or classical 6j-symbols. More generally, if C = Uq (sl2 ), where q is not a root of unity, then the numbers (Φii41 i2 i3 )jl are called q-Racah coefficients or quantum 6j-symbols [CFS]. Further, the evaluation and coevaluation maps define elements (1.42.6) αij ∈ (Hiij∗ )∗ and βij ∈ Hiij∗ , j ∈ I0 . Now the axioms of a rigid monoidal category, i.e., the triangle and pentagon identities and the rigidity axioms translate into non-linear al­ gebraic equations with respect to the 6j-symbols (Φii41 i2 i3 )jl and vectors αij , βij . Exercise 1.42.7. Write down explicitly the relation on 6j symbols coming from the pentagon identity. If C = Rep(sl2 (C)) this relation is called the Elliott-Biedenharn relation ([CFS]). Proposition 1.42.4 gives rise to the following general problem of cat­ egorification of based rings which is one of the main problems in the structure theory of tensor categories. Problem 1.42.8. Given a based ring R, describe (up to equivalence) all multitensor categories over k whose Grothendieck ring is isomorphic to R. It is clear from the above explanations that this problem is equiva­ lent to finding all solutions of the system of algebraic equations coming from the axioms of the rigid monoidal category modulo the group of au­ tomorphisms of the spaces Hijk (“gauge transformations”). In general, this problem is very difficult because the system of equations involved is nonlinear, contains many unknowns and is usually over-determined. In particular, it is not clear a priori whether for a given R this sys­ tem has at least one solution, and if it does, whether the set of these solutions is finite. It is therefore amazing that the theory of tensor categories allows one to solve the categorification problem in a number of nontrivial cases. This will be done in later parts of these notes; now we will only mention the simplest result in this direction, which follows from the results of Subsection 1.7. Let Z[G] be the group ring of a group G, with basis {g ∈ G} and involution g ∗ = g −1 . Clearly, Z[G] is a unital based ring. Proposition 1.42.9. The categorifications of Z[G] are VecωG , and they are parametrized by H 3 (G, k × )/Out(G). Remark 1.42.10. It is tempting to say that any Z+ -ring R has a canonical categorification over any field k: one can take the skeletal semisimple category C = CR over k whose Grothendieck group is R, define the tensor product functor on C according to the multiplication in R, and then “define” the associativity isomorphism to be the identity (which appears to make sense because the category is skeletal, and therefore by the associativity of R one has (X ⊗Y )⊗Z = X ⊗(Y ⊗Z)). However, a more careful consideration shows that this approach does not actually work. Namely, such “associativity isomorphism” fails to be functorial with respect to morphisms; in other words, if g : Y → Y is a morphism, then (IdX ⊗g)⊗IdZ is not always equal to IdX ⊗(g ⊗IdZ ). To demonstrate this explicitly, denote the simple objects of the cat­ egory C by Xi , i = 1, ..., r, and let Xi ⊗ Xj = ⊕k Nijl Xl . Take X = Xi , Y = mXj , and Z = Xl ; then g is an m by m matrix over k. The alge­ bra End((X ⊗ Y ) ⊗ Z) = End(X ⊗ (Y ⊗ Z)) is equal to ⊕s Matmns (k), where � p � ns = Nij Npls = Niqs Njlq , p q and in this algebra we have (IdX ⊗ g) ⊗ IdZ = ⊕rp=1 IdNijp ⊗ g ⊗ IdNpls , IdX ⊗ (g ⊗ IdZ ) = ⊕rq=1 IdNiqs ⊗ g ⊗ IdNjlq , We see that these two matrices are, in general, different, which shows that the identity “associativity isomorphism” is not functorial. 1.43. Semisimplicity of multifusion rings. Definition 1.43.1. A ∗-algebra is an associative algebra B over C with an antilinear anti-involution ∗ : B → B and a linear functional τ : B → C such that τ (ab) = τ (ba), and the Hermitian form τ (ab∗ ) is positive definite. Obviously, any semisimple algebra B = ⊕ri=1 Mati (C) is a ∗-algebra. Namely, if pi > 0 are any positive numbers for i = 1, ..., r then one can define ∗ to�be the usual hermitian adjoint of matrices, and set τ (a1 , ..., ar ) = i pi Tr(ai ). Conversely, it is easy to see that any 8­ algebra structure on a finite dimensional semisimple algebra has this form up to an isomorphism (and the numbers pi are uniquely deter­ mined, as traces of central idempotents of B). It turns out that this is the most general example of a finite dimen­ sional ∗-algebra. Namely, we have Proposition 1.43.2. Any finite dimensional ∗-algebra B is semisim­ ple. Proof. If M ⊂ B is a subbimodule, and M ⊥ is the orthogonal comple­ ment of M under the form τ (ab∗ ), then M ⊥ is a subbimodule of B, and M ∩ M ⊥ = 0 because of the positivity of the form. So we have B = M ⊕ M ⊥ . Thus B is a semisimple B-bimodule, which implies the proposition. � Corollary 1.43.3. If e is a nonzero idempotent in a finite dimensional ∗-algebra B then τ (e) > 0. The following proposition is obvious. Proposition 1.43.4. Let A be a based ring. Then the algebra A ⊗Z C is canonically a ∗-algebra. Corollary 1.43.5. Let A be a multifusion ring. Then the algebra A ⊗Z C is semsimiple. Corollary 1.43.6. Let X be a basis element of a fusion ring A. Then there exists n > 0 such that τ (X n ) > 0. Proof. Since τ (X n (X ∗ )n ) > 0 for all n > 0, X is not nilpotent. Let r � q(x) := (x − ai )mi i=0 be the minimal polynomial of X (ai are distinct). Assume that a0 = � 0 (we can do so since X is not nilpotent). Let r � g(t) = (x − ai )mi xh(x), i=1 where h is a polynomial chosen in such a way that g(a0 ) = 1, g (j) (a0 ) = 0 for j = 1, ..., m0 − 1 (this is clearly possible). Then g(X) is an idempotent, so by Corollary 1.43.3, τ (g(X)) > 0. Hence there exists � n > 0 such that τ (X n ) �= 0, as desired. 1.44. The Frobenius-Perron theorem. The following classical the­ orem from linear algebra [Ga, XIII.2] plays a crucial role in the theory of tensor categories. Theorem 1.44.1. Let B be a square matrix with nonnegative entries. (1) B has a nonnegative real eigenvalue. The largest nonnegative real eigenvalue λ(B) of B dominates the absolute values of all other eigenvalues µ of B: |µ| ≤ λ(B) (in other words, the spec­ tral radius of B is an eigenvalue). Moreover, there is an eigen­ vector of B with nonnegative entries and eigenvalue λ(B). (2) If B has strictly positive entries then λ(B) is a simple positive eigenvalue, and the corresponding eigenvector can be normalized to have strictly positive entries. Moreover, |µ| < λ(B) for any other eigenvalue µ of B. (3) If B has an eigenvector v with strictly positive entries, then the corresponding eigenvalue is λ(B). Proof. Let B be an n by n matrix with nonnegative entries. Let us first show that B has a nonnegative eigenvalue. If B has an eigenvector v with nonnegative entries and eigenvalue 0, then there is nothing to n prove. Otherwise, let Σ be the set of column � vectors x ∈ R with with nonnegative entries xi and s(x) := xi equal to 1 (this is a simplex). Define a continuous map fB : Σ → Σ by fB (x) = s(BBxx) . By the Brouwer fixed point theorem, this map has a fixed point f . Then Bf = λf , where λ > 0. Thus the eigenvalue λ(B) is well defined, and B always has a nonnegative eigenvector f with eigenvalue λ = λ(B). Now assume that B has strictly positive entries. Then f must have strictly positive entries fi . If d is another real eigenvector of B with eigenvalue λ, let z be the smallest of the numbers of di /fi . Then the vector v = d − zf satisfies Bv = λv, has nonnegative entries and has one entry equal to zero. Hence v = 0 and λ is a simple eigenvalue. Now�let y = (y1 , ..., yn ) ∈ Cn be a row vector. Define the norm |y| := |yj |fj . Then |yB| = � � � | yi bij |fj ≤ |yi |bij fj = λ|y|, j i i,j � and the equality holds if and only if all the complex numbers yi bij which are nonzero have the same argument. So if yB = µy, then |µ| ≤ λ, and if |µ| = λ then all yi which are nonzero have the same argument, so we can renormalize y to have nonnegative entries. This implies that µ = λ. Thus, part (2) is proved. Now consider the general case (B has nonnegative entries). Assume that B has a row eigenevector y with strictly positive entries and eigen­ value µ. Then µyf = yBf = λyf , which implies µ = λ, as yf = � 0. This implies (3). It remains to finish the proof of part (1) (i.e. to prove that λ(B) dominates all other eigenvalues of B). Let ΓB be the oriented graph whose vertices are labeled by 1, ..., n, and there is an edge from j to i if and only if bij > 0. Let us say that i is accessible from j if there is a path in ΓB leading from j to i. Let us call B irreducible if any vertex is accessible from any other. By conjugating B by a permutation matrix if necessary, we can get to a situation when i ≥ j implies that i is accessible from j. This means that B is a block upper triangular matrix, whose diagonal blocks are irreducible. So it suffices to prove the statement in question for irreducible B. But if B is irreducible, then for some N the matrix 1 + B + · · · + B N has strictly positive entries. So the nonnegative eigenvector f of B with eigenvalue λ(B) is actually strictly positive, and one can run the argument in the proof of part (2) with a norm bound (all that is used in this argument is the positivity of f ). Hence, the result follows from (2). �