The Dauns-Hofmann theorem and tensor products of C*-algebras Dave McConnell Functional Analysis Seminar, Oxford, 22.10.13 C∗ -bundles Definition Let X be a locally compact Hausdorff space. A C*-bundle over X is a triple (X , A, πx : A → Ax ), where A is a C∗ -algebra, πx : A → Ax surjective ∗-homomorphisms, such that T 1. x∈X ker πx = {0} (equivalently, kak = supx∈X kπx (a)k for all a ∈ A), 2. for f ∈ C0 (X ) and a ∈ A, there is f · a ∈ A with πx (f · a) = f (x)πx (a) for all x ∈ X . ` I may regard a ∈ A as a section â : X → I If for all a ∈ A the function x 7→ kπx (a)k ∈ C0 (X ), we say that (X , A, πx : A → Ax ) is a continuous C∗ -bundle. Upper- and lower-semicontinuos bundles are defined analogously. Ax with â(x) = πx (a), Examples of C∗ -bundles 1. By the Gelfand-Naimark theorem, any commutative C∗ -algebra is a continuous C∗ -bundle; X = Max(A), Ax = C for all x ∈ X and πx the evaluation maps. 2. Trivial bundles: for any C∗ -algebra B and locally compact Hausdorff space X , A = C0 (X , B) (with pointwise operations and supremum norm) is a continuous C∗ -bundle with constant fibre Ax = B. 3. Subtrivial bundles: with A trivial and A0 ⊆ A a C∗ -subalgebra of A closed under the action of C0 (X ), we get a continuous C∗ -bundle (X , A0 , σx : A0 → Bx ) with Bx ⊆ B for all x ∈ X . Generalised Gelfand-Naimark Problem: given a non-commutative C∗ -algebra A, can we choose a space X and ∗-homomorphisms πx : A → Ax such that (X , A, πx : A → Ax ) is a C∗ -bundle. Representation over Prim(A) Let A be a C∗ -algebra and Prim(A) the set of kernels of irreducible representations π : A → B(H). For P = ker π ∈ Prim(A) we identify the quotient C∗ -algebra A/P ≡ π(A). I kak = sup{ka + Pk : P ∈ Prim(A)}, I Hull-kernel topology has open sets of the form U(I ) = {P ∈ Prim(A) : P 6⊇ I }, where I is an ideal of A, I If Prim(A) is Hausdorff, then (Prim(A), A, πp : A → A/P) is a continuous C∗ -bundle (Fell), I In particular if A ≡ C0 (X ) is commutative then Prim(A) = X , I Prim(A) is non-Hausdorff in general (it is a locally compact T0 -space), I Dauns-Hofmann Theorem: represents any C∗ -algebra A as the section algebra of a bundle over the complete regularisation of Prim(A). Glimm(A) I Define an equivalence relation ≈ on Prim(A) via P ≈ Q iff f (P) = f (Q) for all f ∈ C b (Prim(A)), I As a set, let Glimm(A) = Prim(A)/ ≈, and let ρA : Prim(A) → Glimm(A) be the quotient map, I For each f ∈ C b (Prim(A)) define f ρ : Glimm(A) → C via f ρ ([P]) = f (P), where [P] is the ≈-equivalence class of P ∈ Prim(A), Topology τcr on Glimm(A) induced by {f ρ : f ∈ C b (Prim(A))}, I I For each p T = [P] in Glimm(A), denote by Gp the closed two sided ideal Gp = [P], the Glimm ideal of A corresponding to p, I If A is unital then Glimm(A) = Max(Z (A)), and the Glimm ideals are the ideals of A generated by the maximal ideals in Z (A). The Dauns-Hofmann Theorem In what follows we will assume that Glimm(A) locally compact. Theorem (J. Dauns, K. Hofmann, 1968) Let A be a C∗ -algebra. Then the triple (Glimm(A), A, πp : A → A/Gp ), where (i) Gp is the Glimm ideal of A corresponding to p ∈ Glimm(A), and (ii) πp : A → A/Gp is the quotient ∗-homomorphism, is an upper semicontinuous C∗ -bundle. The centre ZM(A) of the multiplier algebra of A is ∗-isomorphic to C b (Glimm(A)), where f ∈ C b (Glimm(A)) acts by pointwise multiplication, πp (f · a) = f (p)πp (a) for all a ∈ A, p ∈ Glimm(A). We will construct the Dauns-Hofmann representation of the tensor product A ⊗ B of C∗ -algebras in terms of the corresponding representations of A and B. Need to determine Glimm(A ⊗ B) in terms of Glimm(A) and Glimm(B). Continuous bundles over Glimm(A) If Prim(A) is Hausdorff, then Prim(A) = Glimm(A) (as sets of ideals and topologically), and the C*-bundle (Glimm(A), A, πp : A → A/Gp ) = (Prim(A), A, πP : A → A/P) is continuous. More generally, Lee’s Theorem shows that (Glimm(A), A, πp : A → A/Gp ) is a continuous C*-bundle if and only if the canonical map ρA : Prim(A) → Glimm(A) is open (it is continuous by construction). Quasi-standard C*-algebras Definition An ideal I / A is said to be primal if, given ideals J1 , . . . , Jn of A with J1 J2 . . . Jn = {0}, there is 1 ≤ i ≤ n with Ji ⊆ I . I Primitive ideals are primal. I A primal ideal is the kernel of the GNS representation associated with a state φ of A which is a weak*-limit of factorial states (Archbold & Batty). Definition A C*-algebra A is called quasi-standard if (i) (Glimm(A), A, πp : A → A/Gp ) is a continuous C*-bundle, and (ii) every Gp ∈ Glimm(A) is a primal ideal of A. For separable A satisfying (i), (ii) is equivalent to the condition (ii’) there is a dense subset D ⊆ Glimm(A) with Gp primitive for all p ∈ D. Examples 1. Let A be the C∗ -algebra of sequences x = (xn ) ⊂ M2 (C) such that xn → x∞ ∈ M2 (C). Then Prim(A) is homeomorphic to N̂ (one-point compactification), where n ∈ N̂ is identified with the ideal Pn = {x ∈ A : xn = 0}, 1 ≤ n ≤ ∞. Thus the Dauns-Hofmann representation of A gives the trivial bundle C (N̂, M2 (C)). 2. Let B = {x ∈ A : x∞ = diag(λ(x), µ(x))} ⊂ A. Then Prim(B) = {Pn : n ∈ N} ∪ {ker(λ)} ∪ {ker(µ)}, where the Pn are isolated and ker(λ) ≈ ker(µ). Thus Glimm(B) = {Pn : n ∈ N} ∪ {ker(λ ⊕ µ)} ≡ N̂, and the Dauns-Hofmann bundle (N̂, B, πn : B → Bn ) has fibres M2 (C) if n ∈ N Bn = C ⊕ C if n = ∞, and B is quasi-standard. A discontinuous example Take the C∗ -algebra C of sequences x = (xn ) ⊂ M2 (C) such that x2n → diag(λ1 (x), λ2 (x)) x2n+1 → diag(λ2 (x), λ3 (x)). Then Prim(C ) = {Pn : n ∈ N} ∪ {ker(λi ) : i = 1, 2, 3}. Glimm(C ) = {Pn : n ∈ N} ∪ {ker(⊕3i=1 λi )} ≡ N̂ The bundle (N̂, C , πn : C → Cn ) has fibres Cn = M2 (C) for n ∈ N, and C∞ = C3 , where π∞ (x) = (λ1 (x), λ2 (x), λ3 (x)). Upper-semicontinuous, but not continuous: take x = (xn ) ∈ C with x2n = diag(1, 0) and x2n+1 = 0. Classes of Dauns-Hofmann representations We have the following relations: { C∗ -algebras A with Prim(A) Hausdorff } = { continuous C*-bundles over Prim(A)} ( { quasi-standard C*-algebras A} ( { continuous C*-bundles over Glimm(A)} ( { u.s.c. C*-bundles over Glimm(A)} = { all C*-algebras }. Question: are these classes closed under tensor products? The minimal tensor product of C∗ -algebras Let A and B be C∗ -algebras, A B their ∗-algebraic tensor product. I Let π : A → B(Hπ ), σ : B → B(Hσ ) be faithful representations, then there is a ∗-homomorphism π σ : A B → B(Hπ ) B(Hσ ) ⊆ B(Hπ ⊗ Hσ ) such that (π σ)(a ⊗ b) = π(a) ⊗ σ(b), I For t ∈ A B we define ktkα = k(π σ)(t)kB(Hπ ⊗Hσ ) , I k · kα is independent of faithful representations π, σ, I Define A ⊗ B as the completion of A B with respect to k · kα , I For ∗-homomorphisms π : A → A0 , σ : B → B 0 we get a ∗-homomorphism π ⊗ σ : A ⊗ B → A0 ⊗ B 0 extending π σ, I If π and σ are injective or surjective then π ⊗ σ has the same property. If k·kγ is another C∗ -norm on A B, A ⊗γ B will denote the completion of A B in this norm. Ideals of A ⊗ B If I / A and J / B, let qI : A → A/I and qJ : B → B/J be the quotient maps. Then qI qJ : A B → (A/I ) (B/J) has ker(qI qJ ) = I B + A J, which, by injectivity, has closure I ⊗ B + A ⊗ J / A ⊗ B. Extending qI qJ to qI ⊗ qJ : A ⊗ B → (A/I ) ⊗ (B/J) gives a closed two-sided ideal ker(qI ⊗ qJ ) / A ⊗ B. Clearly ker(qI ⊗ qJ ) ⊇ I ⊗ B + A ⊗ J. but this inclusion may be strict, If ker(qI ⊗ qJ ) = I ⊗ B + A ⊗ J for all I / A, J / B, we say that A ⊗ B satisfies Tomiyama’s property (F). Tensor Products of C∗ -bundles Let A = (X , A, πx : A → Ax ) and B = (Y , B, σy : B → By ) be C∗ -bundles. We can form the ’fibrewise tensor product’ of A and B, to get a C∗ -bundle (X × Y , A ⊗ B, πx ⊗ σy : A ⊗ B → Ax ⊗ By ), (1) (note that A ⊗ B/(ker(πx ⊗ σy )) ≡ Ax ⊗ By ). Or we may take as fibres the quotient C∗ -algebras A⊗B ≡ Ax ⊗x,y By =: (A ⊗ B)(x,y ) , ker(πx ) ⊗ B + A ⊗ ker(σy ) and get a C∗ -bundle (X × Y , A ⊗ B, θx,y : A ⊗ B → (A ⊗ B)x,y ). (2) If A and B are continuous bundles, then (1) is lower-semicontinuous and (2) is upper-semicontinuous (Kirchberg and Wassermann). In general continuity may fail. Characterisation of Glimm(A ⊗ B) Theorem Let A and B be C∗ -algebras. Then the map Glimm(A) × Glimm(B) → Glimm(A ⊗ B) (Gp , Gq ) 7→ Gp ⊗ B + A ⊗ Gq is an open bijection , which is a homeomorphism if (i) A is σ-unital and Glimm(A) locally compact (in particular if A unital), (ii) The complete regularisation map ρA : Prim(A) → Glimm(A) is an open map (iff Dauns-Hofmann representation of A continuous). Remark: in general, the topology on Glimm(A ⊗ B) depends only on Prim(A) × Prim(B). We will use Glimm(A) × Glimm(B) to denote this space with this possibly stronger topology, so that the above map is a homeomorphism. Sectional representation of A ⊗ B Theorem The Dauns-Hofmann representation (Glimm(A ⊗ B), A ⊗ B, θr : A ⊗ B → (A ⊗ B)/Gr ) of A ⊗ B may be canonically identified with the upper-semicontinuous C∗ -bundle Glimm(A) × Glimm(B), A ⊗ B, θp,q : A ⊗ B → (A ⊗ B)(p,q) , where the fibre algebras are given by (A ⊗ B)(p,q) = A⊗B = (A/Gp ) ⊗(p,q) (B/Gq ), Gp ⊗ B + A ⊗ Gq identifying r ∈ Glimm(A ⊗ B) with (p, q) ∈ Glimm(A) × Glimm(B) such that Gp ⊗ B + A ⊗ Gq = Gr . Corollary ZM(A ⊗ B) is canonically ∗-isomorphic to C b (Glimm(A) × Glimm(B)). The fibrewise tensor product I The norm k · k(p,q) is the minimal C∗ -norm on (A/Gp ) (B/Gq ) if and only if Gp ⊗ B + A ⊗ Gq = ker(πp ⊗ σq ), (∗) where πp : A → A/Gp and σq : B → B/Gq are the quotient maps. I A ⊗ B may be constructed as the minimal fibrewise tensor product of the Dauns-Hofmann bundles of A and B, i.e. (Glimm(A)×Glimm(B), A⊗B, πp ⊗σq : A⊗B → (A/Gp )⊗(B/Gq )) if and only if (∗) holds for all (p, q). Theorem Let A and B be C∗ -algebras such that (∗) holds for all pairs of Glimm ideals of A and B. Then A ⊗ B is a continuous C∗ -bundle over Glimm(A ⊗ B) if and only if A and B are continuous bundles over Glimm(A) and Glimm(B) respectively. The fibrewise tensor product ii. I What can we say about the condition Gp ⊗ B + A ⊗ Gq = ker(πp ⊗ σq ) (∗) for all (p, q) ∈ Glimm(A) × Glimm(B)? I Kaniuth showed that if A ⊗ B has property (F), then (∗) holds. I This condition is not necessary: B(H) ⊗ B(H) does not satisfy property (F), but Glimm(B(H)) = {0}, so clearly (∗) holds. I If A and B are continuous C∗ -bundles over their Glimm spaces then, for (p0 , q0 ) ∈ Glimm(A) × Glimm(B), we have (p, q) 7→ k(πp ⊗ σq )(c)k is continuous at (p0 , q0 ) if and only if (∗) holds at this point (Archbold). I No counterexamples given. I Using a result of Kirchberg, we can construct C∗ -algebras A and B for which (p, q) 7→ kc + (Gp ⊗ B + A ⊗ Gq )k (1) is continuous on Glimm(A) × Glimm(B) for all c ∈ A ⊗ B, but there is a c 0 ∈ A ⊗ B with (p, q) 7→ k(πp ⊗ πq )(c 0 )k I discontinuous at some (p0 , q0 ) ∈ Glimm(A) × Glimm(B): Q We take A = B(H) and B = n≥1 Mn (C), then Glimm(A) = {0} and Glimm(B) = βN, hence Glimm(A ⊗ B) = {B(H) ⊗ Gq : q ∈ βN}. I (2) We get q0 ∈ βN\N with B(H) ⊗ Gq0 ( ker(id ⊗ σq0 ). Then (2) is discontinuous at q0 for some c 0 ∈ A ⊗ B, but we can show that (1) is continuous (in fact A ⊗ B is quasi-standard). Thus the condition (∗) is not equivalent to continuity of the Dauns-Hofmann representation. Exact C∗ -algebras I Let B be a C∗ -algebra and J an ideal of B. Then the sequence 0 /J ι /B q / (B/J) /0 (†) is called a short exact sequence of C∗ -algebras. I We say that a C∗ -algebra A is exact if the sequence 0 / A⊗J id⊗ι / A⊗B / A ⊗ (B/J) id⊗q /0 is exact for every such sequence (†), that is, A ⊗ J = ker(id ⊗ q). I Examples include all commutative C*-algebras, liminal C*-algebras, nuclear C*-algebras, subalgebras and quotients of nuclear C*-algebras. Exact C*-algebras For H separable, then B(H) is inexact; in fact the short sequence 0 / B(H) ⊗ K (H) / B(H) ⊗ B(H) B(H) ⊗ (B(H)/K (H)) /0 is not exact; ker(id ⊗ qK ) ) B(H) ⊗ K (H) (Wassermann). Theorem (Archbold, Batty, Kirchberg, Wassermann) A is exact ⇔ A ⊗ B has property (F) for all C∗ -algebras B. Note that (∗) is equivalent to exactness of 0 / Gp ⊗ B + A ⊗ Gq / A⊗B (A/Gp ) ⊗ (B/Gq ) /0 Characterisation of exactness Theorem For a C∗ -algebra A, the following are equivalent: (i) A is exact, (ii) For every C∗ -algebra B and q ∈ Glimm(B), the sequence 0 / A ⊗ Gq id⊗ι / A⊗B id⊗σq / A ⊗ (B/Gq ) /0 is exact, where σq : B → B/Gq is the quotient map, (i.e. A ⊗ Gq = ker(id ⊗ σq )). (iii) For every C∗ -algebra B and (p, q) ∈ Glimm(A) × Glimm(B), we have A ⊗ Gq + Gp ⊗ B = ker(πp ⊗ σq ), I Thus for a C*-algebra A: A ⊗ B has (∗) for all B ⇔ A exact ⇔ A ⊗ B has (F) for all B. I But given A and B, A ⊗ B has (∗) 6⇒ A ⊗ B has (F). Exact C*-bundles Corollary Let A and B be C∗ -algebras, one of which is exact. Then (i) The Dauns-Hofmann representation of A ⊗ B over Glimm(A ⊗ B) is a continuous C∗ -bundle if and only if the same is true for both A and B, (ii) A ⊗ B is quasi-standard if and only if A and B are, (iii) Prim(A ⊗ B) is Hausdorff if and only if Prim(A) and Prim(B) are. We will show that this ’stability property’ characterises exactness. Theorem (Kirchberg & Wassermann) Let (X , A, πx : A → Ax ) be a continuous C∗ -bundle. TFAE: (i) A is an exact C∗ -algebra, (ii) For any continuous C∗ -bundle (Y , B, σy : B → By ), the C∗ -bundle (Y , A ⊗ B, id ⊗ σy : A ⊗ B → A ⊗ By ) is continuous, (iii) For any continuous C∗ -bundle (Y , B, σy : B → By ), the C∗ -bundle (X × Y , A ⊗ B, πx ⊗ σy : A ⊗ B → Ax ⊗ By ) is continuous. Remarks: I Ax exact ∀ x ∈ X 6⇒ A exact, I For (ii) or (iii) ⇒ (i), enough to take B separable, unital, and Y = N̂. With A inexact, the Kirchberg & Wassermann result gives a separable, unital, continuous C∗ -bundle (N̂, B, σn : B → Bn ) such that (N̂, A ⊗ B, id ⊗ πn : A ⊗ B → A ⊗ Bn ) is discontinuous at ∞. Using this we can construct a continuous bundle (N̂, C , γn : C → Cn ), with the property that (N̂, A ⊗ C , θn : A ⊗ C → (A ⊗ C )/(A ⊗ ker(γn ))) is discontinuous at ∞. By a result of Blanchard, we may assume that (N̂, C , γn : C → Cn ) has simple fibres. It follows that Prim(C ) = Glimm(C ), canonically homeomorphic to N̂. Thus, for each inexact A there is a continuous bundle (N̂, C , γn : C → Cn ), with Prim(C ) = N̂, such that (N̂, A ⊗ C , θn : A ⊗ C → (A ⊗ C )/(A ⊗ ker(γn ))) is discontinuous at ∞. Suppose now that the bundle (Glimm(A), A, πp : A → A/Gp ) is continuous. As a consequence of the above, we can show that the Dauns-Hofmann representation of A ⊗ C , i.e. the bundle A⊗C Glimm(A ⊗ C ), A ⊗ C , θp,n : A ⊗ C → , A ⊗ Gn + Gp ⊗ C is necessarily discontinuous. Characterisation of exactness Theorem Let A be a C∗ -algebra. (i) If (Glimm(A), A, πp : A → A/Gp ) is a continuous C∗ -bundle, then A is exact ⇔ for all C∗ -algebras B with (Glimm(B), B, σq : B → B/Gq ) continuous, the bundle (Glimm(A ⊗ B), A ⊗ B, θr : (A ⊗ B)/Gr ) is continuous, (ii) If A is quasi-standard, then A is exact ⇔ A ⊗ B is quasi standard for all quasi-standard C∗ -algebras B, (iii) If Prim(A) is Hausdorff, then A is exact ⇔ Prim(A ⊗ B) is Hausdorff for all C∗ -algebras B with Prim(B) Hausdorff. Tensor products of C*-bundles Thus none of the classes { C∗ -algebras A with Prim(A) Hausdorff } = { continuous C*-bundles over Prim(A)} ( { quasi-standard C*-algebras A} ( { continuous C*-bundles over Glimm(A)} are closed under tensor products. In each case, their intersection with the class of exact C*-algebras is the largest ⊗-closed subclass. The maximal tensor product For C*-algebras A and B, the maximal C*-norm on A B is defined as ktkmax = sup{ktkγ : k·kγ a C*-norm on A B} for t ∈ A B. A C*-algebra A is said to be nuclear if k·kmax = k·kα on A B for all C*-algebras B. Theorem Let A be a unital quasi-standard C*-algebra. Then A is nuclear ⇔ A ⊗max B is quasi-standard for all quasi-standard C*-algebras B.